<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Media Distillery on Medium]]></title>
        <description><![CDATA[Stories by Media Distillery on Medium]]></description>
        <link>https://medium.com/@mediadistillery?source=rss-35ebb815930c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*0P_-fKH8tDyn1jNVMiiGlQ.png</url>
            <title>Stories by Media Distillery on Medium</title>
            <link>https://medium.com/@mediadistillery?source=rss-35ebb815930c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:45:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mediadistillery/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Zero to Production in less than a day]]></title>
            <link>https://blog.devgenius.io/from-zero-to-production-in-less-than-a-day-f4e16f628f15?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4e16f628f15</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[github]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Thu, 23 Jun 2022 07:42:40 GMT</pubDate>
            <atom:updated>2022-07-01T17:43:47.043Z</atom:updated>
            <content:encoded><![CDATA[<p>At Media Distillery we make ample use of the micro-services paradigm. Each of our products is powered by a multitude of microservices that provide functionalities like (semantic) search, data storage, dashboards, EPG correction, face recognition, face clustering, text recognition, thumbnail selection, topic detection, topic segmentation, and many more.</p><p>The microservices in our platform can be divided into two groups: Java services and Python services. The latter takes care of the majority of our ML needs; and is the focus of this article. We’ll be going over how we make a new Python service, what such a service looks like and which tools and technologies we use to create and deploy it.</p><h3>An average Python service</h3><p>In the simplest case, the pipelines behind our products are separated into two services: a Java service and a Python service. The former takes care of things such as clustering, load balancing, data storage, and many more. The latter hosts the relevant machine learning model(s), taking care of both passing the input, such as an image or a piece of text, through the model and the pre- and post-processing of the input and output. The Java and Python services communicate through a simple HTTP REST API.</p><p>Thanks to this setup we can take advantage of the maturity and stability of the Java landscape while giving the ML side the flexibility and freedom to keep up with developments in the quickly changing, and slowly maturing, Python environment.</p><h3>Tools, Tools, Tools! (and tech)</h3><p>To create and maintain a Python service we make use of a host of open-source tools. For simplicity, we’ve divided the tools that we use into two groups: on one hand, we have the ones used for creating the service and, on the other hand, those used for getting that service onto our platform.</p><figure><img alt="Overview of the tools used for creating the service and for getting that service onto our platform." src="https://cdn-images-1.medium.com/max/800/1*SN2hRUlAgB37vYzvI3295Q.png" /></figure><h3>Creating the service</h3><h4>FastAPI</h4><p>For our HTTP REST framework, we use FastAPI. This framework has been gaining a lot of popularity in recent years and, as we were on the lookout for an alternative, FastAPI seemed like an excellent replacement. What caught our attention was primarily the high performance of the framework, but the overall feature set and design philosophy also scored high in our list.</p><p>FastAPI is an optimal candidate, however we’re yet to take full advantage of the asynchronous side of the framework.</p><h4>Marshmallow / Pydantic</h4><p>For data serialisation and validation we’ve mainly been using Marshmallow over the years, without any complaints. It has not only proven to improve the overall quality of our services, but also the quality of our prototyping, evaluations and other investigations. Even though it has been a great library, we’ve recently been looking at switching over to Pydantic. Simply because FastAPI supports Pydantic models out of the box (and not Marshmallow).</p><h4>YAML</h4><p>For configuration purposes, we make use of YAML files. Not too exciting, but it ties into another configuration tool that comes into play in the deployment phase.</p><h4>Poetry</h4><p>Similar to FastAPI, this setup.py replacement has been gaining a lot of popularity in recent years. Although setup.py and setuptools were mostly working fine for us, they were slowly becoming outdated and it was time for us to start looking for a new dependency and package management tool. In the end, we settled on Poetry, which we found very easy to work with and in constant improvement thanks to an active community.</p><h4>Pytest</h4><p>Naturally, we also need unit tests for our code. For this, we make use of Pytest, which appears to be the de facto standard when it comes to unit testing in Python.</p><h4>Nexus</h4><p>As we need to build a lot of internal Python libraries, it makes perfect sense to store all of them in a shared location. For this purpose, we use Nexus, which provides PyPI repositories out of the box. It is also used for Maven repositories, aligning nicely with our Java and Python-focused services.</p><h4>Black</h4><p>Black is a very recent entry to the list of our tools. In the past we haven’t made use of code formatters, but the need for them has been growing. Black, which recently released its first stable version, is slowly becoming the standard for Python code formatting. Our first impressions so far have been good and we’ll likely be using it more often in the future.</p><h3>Deploying the service</h3><h4>Kubernetes</h4><p>All our services are deployed as containers to an on-premise installation of Kubernetes. It is mainly managed by our Backend department.</p><h4>Docker</h4><p>For creating the deployable containers we use Docker. For each new version of an ML service, we create a new docker image that will expose the API.</p><h4>Helm</h4><p>Helm is essentially a layer on top of Kubernetes to ease the deployment and management of services. For each service, in addition to a Docker image to encapsulate the service, a new Helm “chart” needs to be created, which defines how the Docker image should be deployed on Kubernetes.</p><h4>Jenkins</h4><p>Jenkins provides a way to automate arbitrary development tasks. For the ML services we have defined various automation pipelines to build, for instance, service snapshots and make new releases. Those pipelines have not only sped up our development time but also eased collaboration between different teams and products.</p><h4>Spring Cloud Config</h4><p>As previously mentioned, we use YAML files for configuration. In order to make services easily configurable, it is best to not package their configuration files as part of the Docker image but instead have them hosted at a separate location. To achieve this, we use Spring Cloud Config, a service that connects to a Git repository hosting all our configuration files, making them easily accessible through an HTTP REST API.</p><h4>Prometheus</h4><p>To create visibility on the state of our running services we need to expose metrics on that service and monitor them. To achieve this goal we use Prometheus. The metrics are created and exposed using a Prometheus client in each service. This is done together with a separate Prometheus server that pulls these metrics from all client services. Alerts are also configured in the Prometheus server based on those metrics.</p><h4>Grafana</h4><p>Grafana is a layer on top of Prometheus, allowing us to create much nicer visualisation dashboards of Prometheus metrics.</p><h3>Tying it all together</h3><p>All these tools are great and all, but in the end, what matters most is how they all form a fully-fledged service. To help with this goal, we created a service template.</p><h4>Our approach</h4><p>For quite a while, whenever we needed to build some new services, we either started from scratch or copied an existing service. Although the latter can be a time-saving option, it often happens that a given copied service is lagging behind on current standards, transferring over to the new service and creating more technical debt. Furthermore, as we’re working in multiple teams, each group is responsible for different services adopting different standards. This eventually causes more drift between the teams and reduces the maintainability of our services as a whole.</p><p>To combat these issues we created a service template, adapted from an existing <a href="https://github.com/4OH4/kubernetes-fastapi">FastAPI template</a>. Essentially nothing more than a folder containing the file structure that forms a proper service; using the correct tooling. If we want to introduce some new tools, they can be added to the service template, to ensure that all the new services are always up to date with the current standards.</p><h4>Giving back</h4><p>The service template has helped us a lot, enabling faster developments and improving the overall quality and consistency of our codebase.</p><p>Since we make use of a lot of open-source tools, we want to give something back to the community by open-sourcing our service template. Whether you’re interested in using the template directly or simply want to learn more about what a production-quality Python service might look like, we hope this template will be as useful to you as it keeps being for us.</p><p>The service template is available on <a href="https://github.com/mediadistillery/ExampleTemplateService">Github</a>. You can simply copy the template, adjust the naming and implement the new logic.</p><h4>Closing thoughts</h4><p>In this article, we went over the tooling we use for Python services and the service template that ties them together. Although it’s been working very well for us, we’re always looking at how we can improve further. For instance, the service template doesn’t work retroactively, meaning that applying new standards to existing services still needs to be done manually.</p><p>We’ve also recently had our eye on improving our testing standards, such as the inclusion of test automation tools like Tox or Nox. On the ML side, we’ve recently benefited greatly from the use of inference servers like NVIDIA Triton, which may end up significantly impacting what our service template looks like.</p><p>All-in-all, we see that there are several aspects to further investigate and improve, so keep an eye out on our publications regarding refinements and additions to our tech stack.</p><p>Here are the links to the tool GitHub pages:</p><ul><li><a href="https://github.com/tiangolo/fastapi">FastAPI</a></li><li><a href="https://github.com/marshmallow-code/marshmallow">Marshmallow</a></li><li><a href="https://github.com/samuelcolvin/pydantic">Pydantic</a></li><li><a href="https://github.com/yaml/pyyaml">Yaml</a></li><li><a href="https://github.com/python-poetry/poetry">Poetry</a></li><li><a href="https://github.com/pytest-dev/pytest">Pytest</a></li><li><a href="https://github.com/sonatype/nexus-public">Nexus</a></li><li><a href="https://github.com/psf/black">Black</a></li><li><a href="https://github.com/kubernetes/kubernetes">Kubernetes</a></li><li><a href="https://github.com/docker/cli">Docker</a></li><li><a href="https://github.com/helm/helm">Helm</a></li><li><a href="https://github.com/jenkinsci/jenkins">Jenkins</a></li><li><a href="https://github.com/spring-cloud/spring-cloud-config">Springcloud</a></li><li><a href="https://github.com/prometheus/prometheus">Prometheus</a></li><li><a href="https://github.com/grafana/grafana">Grafana</a></li><li><a href="https://github.com/mediadistillery/ExampleTemplateService">Example_Template_Service</a></li><li><a href="https://github.com/4OH4/kubernetes-fastapi">Original Service Template</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4e16f628f15" width="1" height="1" alt=""><hr><p><a href="https://blog.devgenius.io/from-zero-to-production-in-less-than-a-day-f4e16f628f15">From Zero to Production in less than a day</a> was originally published in <a href="https://blog.devgenius.io/">Dev Genius</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Query Expansion for Snackable Content]]></title>
            <link>https://mediadistillery.medium.com/query-expansion-for-snackable-content-62bc17853720?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/62bc17853720</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[snackable-content]]></category>
            <category><![CDATA[query-expansion]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Thu, 03 Dec 2020 14:30:28 GMT</pubDate>
            <atom:updated>2021-01-29T09:49:45.962Z</atom:updated>
            <content:encoded><![CDATA[<p>At Media Distillery we have been working on a solution we’ve dubbed Snackable Content<em>™</em>. Using technologies like speech recognition, face recognition, and topic recognition, we can automatically create short clips based on a consumer’s favourite person, topic, or interest from TV content. With this solution, TV operators can present bite-sized content extracted from long-form video content. In this article, we discuss in-depth one of these technologies and how it contributes to our innovative Snackable Content<em>™</em> solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A3UAmyPVyrzhlvRJnd3rNA.jpeg" /></figure><p>For the main use-case of Snackable Content<em>™</em> we look for interesting topic clips in TV content. In this process, we decide which topics to search for and how to present them. We can also allow the user themselves to provide a user query (or “topic”). This way the task becomes more closely related to a classic information retrieval problem which we refer to as “topic search.”</p><p>While we have been investigating the topic search use-case, we have experimented with a number of technologies to improve its performance. Let’s now have a look at what we learned from our experiment with query expansion.</p><h3>Query Expansion</h3><p>In a nutshell, query expansion is a common technology used in information retrieval to increase the number of relevant search results (“recall”) by adding related terms to a user query. There are roughly two steps in this process: how to find new terms and how to add these new terms to the existing query.</p><p><strong>Step 1: How to find new terms</strong></p><p>Finding new terms can be done in a bunch of different ways. One simple way is by finding synonyms of each term in the original query. You can do this in an automated way for the English language through the readily available WordNet lexical database which has modeled these types of word relations. The benefit of this method is that the terms are guaranteed to be relevant to the input term, which depending on the method is not always guaranteed. One of the downsides is that lexical databases in languages other than English aren’t as readily available, which in our case could make it more difficult for us to scale our solution to other countries.</p><p>Another more scalable method is using word embeddings. Each word embedding represents a point in high-dimensional space where the embeddings are trained in such a way that word embeddings that are closer together represent words that are semantically related. By converting each term in the user query to a word embedding we can find semantically related terms. A common method to accomplish this is called Word2Vec. Many more methods have been developed since its original creation, but Word2Vec has remained very popular. Since it’s easy to start and scale with, this is the method we’ve mostly been focusing on.</p><p><strong>Step 2: How to add these new terms to the existing query</strong></p><p>Once you have the expanded terms the next part is to create the expanded query. This can be as trivial as appending the expanded terms to the original terms.</p><p>A possible side-effect of this — and query expansion in general — is that the expanded terms may cause the user query to drift too far from what the user originally intended. A simple way to mitigate this is to introduce a form of weighting on the expanded query. The weight of the expanded terms, w.r.t. finding related clips, is reduced as compared to the original terms. If and how this would work depends greatly on how the search is done.</p><p>Now we would like to demonstrate through some code examples our experiences and main findings using Word2Vec approaches for Query Expansion.</p><h3>Demo</h3><p>We start by setting up a Python 3.6 virtual environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/677/1*EB6o7Jm8Ya82M4IesmjGMA.jpeg" /></figure><p>We’ll be working with the Word2Vec models provided by the library Gensim. These are not the models we used internally, but they are adequate for this demonstration. Start by installing the library itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*guYp5IdN_XTsFRl6uBhRPg.jpeg" /></figure><p>The model will automatically be downloaded and stored locally the first time you try to load the model in a script.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*TimrLmoFVKQqWFyhmRiSeQ.jpeg" /></figure><p>Now that we have a model let’s see what we get with a simple topic like “cooking”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*CKLfnoYsgZhwbHMhhF5W8g.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*VFdfuLRGAf6DLF00pmwIGg.jpeg" /></figure><p>We get back a list of the top 10 (by default) most similar terms to the input term and an associated confidence score for each. We could directly add these terms to the query, but we may end up adding terms that are completely unrelated to the original query. We can mitigate this by filtering only terms with a high confidence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*EyWhncwgd3ESB5O7J8De5w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*x-KJFvcx7-az7Rd2Cp14Sg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*v2EzlXIsxJ9am4RiHsuk1w.jpeg" /></figure><p>Now what if we want to expand user queries that contain multiple words?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*-PuwBI_tfeW2XC2QmHfX1A.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*Zepm7VzfBE6hEgTktfg6DQ.jpeg" /></figure><p>As expected, most Word2Vec models are trained on single words. It doesn’t recognize “clean energy” as a single term, but it surely will recognize “clean” and “energy” separately. This means we need some extra preprocessing. This step can get much more complex, but for now, we’ll keep it simple and use an existing utility from Gensim.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*kJRqCd5oQYsHl7NVj9VSww.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*xU5jIH44CQOec-N86c1nGQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*Qs-dsGzdZBSg3tZFEA9UEQ.jpeg" /></figure><p>As we see, we can still expand the query by expanding each term separately. Lastly, we create a utility function to combine the preprocessing, thresholding, and query (re-)building.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*XQ9btIunEb_R-FwtvbhioQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*QUT0MvdlN4DNKmlMHDV1OA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*EQ6xZ1vBw6I1-ow18kRoZA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*2oysXEdoa-JNXZOM32hwlQ.jpeg" /></figure><p>This approach appears to work well enough for the query “cooking”. The model adds useful contextual information to the query. Unfortunately for the “clean energy” example it mostly adds noise. Words like “stick” and “value” have very little to do with “clean energy”. Putting it differently, through preprocessing we have lost the meaning, or “sense”, of the query. We can address this issue using Sense2Vec.</p><p>Sense2Vec is a word embedding model that was trained to preserve the sense of its words. For example, “clean” and “energy,” separately, have different meanings than “clean energy” together. For our use-case, this means the model can work directly with multi-word phrases and preserve the sense within them. We can easily access the model without having to download it through a handy RESTful API. Using that API let’s try the same query as before. To re-use the previous “expand_query” function we’ll need a small wrapper class and an extra dependency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/681/1*OrubbdN0MVJGE2atcDGHhg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*ikManZskldxOWTKw7BXwSQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*gOYHgUe3x9DlAZz6n8wJzg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*d4XELvXUVzWrH4my0xtGjQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*G6I0kzEOWpeF7BdwSivGIg.jpeg" /></figure><p>We lowered the threshold for the “cooking” query for demonstrative purposes. We see it is able to expand the query with words that are intuitively more closely related to the original query. With the “clean energy” example we see by the expanded terms it is able to preserve the original sense of the terms in the query.</p><h3>Concluding Thoughts</h3><p>During the experimentation on our platform, while we found the implementation of query expansion did not significantly improve our topic search use-case to find more related content, it certainly didn’t hurt our overall performance either. For some topics, query expansion introduced relevant results that we otherwise wouldn’t have found and for others, it introduced irrelevant results, resulting in a small overall improvement that we concluded wasn’t statistically significant. We only had a limited amount of annotated data available which was most likely a big factor in this result. Regardless, we gained some useful knowledge from our experimentation:</p><ul><li>We only looked at very simple user queries. While this was already sufficient to get decent results, there are much more advanced tokenization and preprocessing steps that can be applied to deal with more complex queries.</li><li>With both approaches, we needed to set some kind of threshold. Setting it too low meant potentially too many irrelevant terms were added to the query. Conversely, setting it too high meant nothing might be added to the query. We didn’t go in-depth to figure out a good setting for these thresholds, but they are still worth exploring further.</li><li>You can’t expand out-of-vocabulary words. Concrete examples are “corona” or “COVID”. These terms either didn’t exist yet when the models were trained or didn’t carry the same weight they do now. This means any model we use would need to be frequently updated to keep up with current affairs.</li><li>Query expansion is more focused on recall, whereas for our use-case <em>precision </em>is much more important. Often, not finding any results is better than finding irrelevant results.</li><li>We did not explore the usefulness of lexical databases like WordNet for query expansion. (One of our issues was introducing irrelevant results.) This approach could help to mitigate that.</li></ul><p>While the output from our experiments with query expansion for topic search wasn’t a smashing success, we did get multiple interesting findings, and there are plenty more avenues to explore. Based on these findings and experiences, we continue to be excited to explore the possibilities of query expansion in future projects. We hope you gained some useful insights, as well, about whether your use-case could benefit from this approach to query expansion, how you could quickly and easily get started, and how you could deal with some of the issues you may encounter in the process.</p><p><em>Written by </em>Ryan Amirkhan<em>, </em>Machine Learning Expert at Media Distillery.</p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/query-expansion-for-snackable-content/"><em>https://www.mediadistillery.com</em></a><em> on December 3, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62bc17853720" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What’s wrong with the EPG?!]]></title>
            <link>https://mediadistillery.medium.com/whats-wrong-with-the-epg-d778fd75ef4?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/d778fd75ef4</guid>
            <category><![CDATA[epg]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[videos]]></category>
            <category><![CDATA[epg-correction]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Tue, 20 Oct 2020 08:29:43 GMT</pubDate>
            <atom:updated>2020-10-20T08:44:34.880Z</atom:updated>
            <content:encoded><![CDATA[<p>The year was 1941 when the age of commercial television began. It was in that same year that one New York area TV station — WNBT — had the foresight to print the first TV guide, known as a “Program Card,” to distribute to local TV owners. Though it would be another two decades or so until TV consolidated into the mass media we know of today, and not until 1981 that the first Electronic Program Guide (EPG) service would be available to consumers, this small paper piece of TV history already contained a warning about an issue that has remained until present day: “All programs subject to change without notice.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*w1j9WXj6ULsyP3rscnNXTw.jpeg" /></figure><p>We’ve come to accept that events broadcasted “live” will vary their beginning and end times — or maybe never make it to air. We can understand when paper guides printed days in advance are unreliable. But what prevents online TV guides from being accurate?</p><p><strong>Hedging with Live Events</strong></p><p>The prime example here is sporting events. We know when events start but we cannot be sure about when some of them will end due to the nature of the competition (e.g. tennis, cycling). In some cases, TV schedulers take this potential delay into consideration by adding padding programming after the event that can absorb the delay generated by the live event. If the padding is missing, the delay will have a trickle down effect on the programming for the rest of the day.</p><p><strong>The Ratings Chess Match</strong></p><p>TV scheduling is a strategy game that depends not only on the programming goals of each channel but also on the programming of competitor channels. It’s during the primetime of major channels that this game reaches its climax and can turn into a chess match, where channels need to protect their king (highlight content) or attack another channel’s king. Normally one of the first pawns to be sacrificed is sticking to scheduled times. In some cases, schedulers will avoid overlapping the beginning of a show with the end of an established and successful show or event in a competing channel. If that hit show runs over its time slot, programs in other channels will be delayed to avoid a collusion that will impact its ratings. Channels can also use their most successful content as a weapon against powerful premieres in other stations aiming for that clash in content. Once the competitor’s program has started before the end of your content, viewers will be less likely to switch to something that has already started.</p><p><strong>Outdated Data = Delays</strong></p><p>It usually goes unnoticed, but having an EPG in place with all its data normally requires the collaboration of several companies. Each broadcaster elaborates their own EPG with their own formats. There are then big metadata aggregators in charge of compiling and standardizing this information, who send it to TV operators (among other players.) This transfer of information that takes place between a broadcaster updating its EPG until it’s visible for the user is <em>not </em>instant, as it requires time to ingest, process and deliver the information among the different actors. Things can go wrong, delays can happen, information can be sent too late to be processed… all have the same outcome: the EPG available to the users becomes outdated and programs start to deviate from their scheduled time.</p><p><strong>Arbitrary Time Slots</strong></p><p>Take a look at the EPG of any channel and you’ll see that most shows are scheduled at round times, where in general the smallest unit is five minutes. We are used to having meetings, appointments and events scheduled at round times for one simple reason: they are easier to remember. However, the length of programs are far from round, making the task of fitting programming into nice and round timeslots challenging. Ad breaks, and particularly content promotions, are used to adjust this but it’s very difficult to maintain for the full schedule, especially if we consider some of the factors mentioned before. In the best cases, the resulting EPG can be a close enough schedule, with up to a couple minutes deviation, but in others it can look more like rough guide.</p><p>In the era of video on-demand we have been conditioned to hit play and watch our program begin immediately. As viewers are increasingly shifting away from live TV to catch-up TV, a new problem arises for many TV operators that need to base their catch-up services on EPG times. Hitting the play button in this scenario results in videos starting with a ton of ads, the end of the previous program, a totally different show, or in some cases the start of the program we wanted to watch. This uncertain output can drive viewers away from this functionality and ultimately look for content to watch elsewhere.</p><p><strong>The Solution</strong></p><p>While broadcasters can use various technical systems to signal more accurate start times, they are often inconsistently available across channels. Using Artificial Intelligence to understand what is inside video, Media Distillery can detect in real-time visual cues that mark the beginning and the end of a programme, thanks to our EPG Correction™ solution. It means that the scheduled programme start and end time can be automatically updated with the actual programme start and end time based on observed features in the channel as broadcast. This approach has been already successfully applied by global video service providers to deliver a smarter programme guide.</p><p>Therefore, the smarter start and end times enabled by Media Distillery enhance the viewing experience and offer a solution to a source of viewer frustration that has been a problem ever since the early days of home video recording. While there may be alternative technical solutions to improving the metadata in the programmes as transmitted, this requires the technical co-operation of multiple broadcast channels. By using Artificial Intelligence to analyse the channels as they are actually transmitted, television and video service providers are enabled to improve the experience so that it is consistent across multiple channels.</p><p><em>Written by </em>Adrian Arroyo Mirantes<em>, Product Support Coordinator at Media Distillery.</em></p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/whats-wrong-with-the-epg/"><em>https://www.mediadistillery.com</em></a><em> on October 19, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d778fd75ef4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Survive the Streaming Wars]]></title>
            <link>https://mediadistillery.medium.com/how-to-survive-the-streaming-wars-6d35a990f26?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/6d35a990f26</guid>
            <category><![CDATA[streaming]]></category>
            <category><![CDATA[vod]]></category>
            <category><![CDATA[television]]></category>
            <category><![CDATA[subscription]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Wed, 20 May 2020 12:33:31 GMT</pubDate>
            <atom:updated>2020-05-20T14:02:04.019Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://variety.com/2019/tv/news/phoebe-waller-bridge-amazon-deal-1203347221/">“Phoebe Waller-Bridge Inks Overall Deal at Amazon”</a></p><p><a href="https://techcrunch.com/2020/03/06/quibi-launch-lineup/?guccounter=1">“Quibi will launch with 50 shows on April 6”</a></p><p><a href="https://www.cnbc.com/video/2020/01/16/nbc-announces-details-of-new-streaming-service-peacock.html">“NBCUniversal announces details of new streaming service Peacock”</a></p><p>It seems like every week we read headlines like these. Time and time again we are reminded that VOD platforms are gaining strength as key players in the entertainment industry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aAkkcYPN2ADGki2NXcx9Wg.jpeg" /></figure><p>To stay competitive and secure their position in the trenches of the streaming wars, online services (and all those competing with them) are intensifying their efforts with content production. Over the last few years, subscribers/viewers have been flooded with programming from all over the world on numerous platforms. And yet, even with this enormous amount of content to choose from, how many times have you caught yourself navigating from one streaming platform to the other, and (rather paradoxically), breathily murmuring to yourself: “I can’t find anything to watch”?</p><p><strong>A New Era, A New Model</strong></p><p>According to data from Nielsen, viewers in the United States can spend up to 10 minutes looking for something to watch in search menus before they give up. Most of them end up tuning in to their trusted, favourite TV channels, while the viewers that do know what they want to watch face another problem: Is the show I want to binge tonight available on one of the subscriptions that I have, and on which platform?</p><p>Many think that one day an entity will come and devour the competition, and put all the content under the same roof. Although that’s highly unlikely (and wishful thinking), TV providers have the opportunity to fulfill this role in a different way, but with a similar outcome. Similar to the cable television boom of the 1980s and 1990s when TV providers were competing for our subscriptions with bundled channels on their platforms, they <em>now </em>have the opportunity to put all or most streaming services under their umbrella.</p><p>Making content available from different streaming platforms accessible through their UI is something that we already see in Europe. Prior to its broad launch in March 2020, Disney Plus signed exclusive deals with Movistar in Spain and with Canal+ Group in France. These deals were similar to the ones achieved with Netflix back in the day. Despite any potential discounts, promotions, or small savings in their subscription, the key value that these TV providers need to ensure for their client is<em> content integration.</em></p><p><strong>Solving for (U)X</strong></p><p>Adding a button on your platform to access a particular streaming service adds little extra value to viewers. In order to keep offering a service that their clients will continue paying for month after month, TV providers should focus their efforts on integrating the content from all these streaming services to make it feel like a natural extension of their platform, and a seamless, natural extension of the viewer’s user experience. Additionally, the exponential increase of content available on their platforms should be paired with an investment in the current search and recommendation capabilities to guarantee viewers can spend more time watching content and less time surfing through irrelevant recommendations, or worse: Coming across the dreaded “Zero Search Results” prompt.</p><p><strong>Integrate, Then Maneuver</strong></p><p>Discounts, promotions, and whatever micro-savings TV providers have previously used to lure in customers is simply not enough to compete in the battle for relevance in the age of streaming wars. The key value that TV providers need to ensure for their customers is content integration <em>that enables</em> a seamless user experience. Whether you’re searching for an old, nostalgic movie from the provider’s own catalog, or you want to binge the latest docuseries from Netflix, or peruse the news highlights from yesterday, the experience you get when you search, find, and play this content should be the same, regardless of its origin.</p><p>This combination of content integration paired with search recommendation improvements will be vital in ensuring the prolonged relevance of TV providers. It is also a large victory in a time when TV as we know it is facing the largest battle of its life.</p><p><em>Written by </em>Adrian Arroyo Mirantes<em>, </em>Product Support Coordinator at Media Distillery.</p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/topic-detection-for-video-the-important-points/"><em>https://www.mediadistillery.com</em></a><em> on May 20, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6d35a990f26" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Topic Detection for Video: the Important Points]]></title>
            <link>https://mediadistillery.medium.com/topic-detection-for-video-the-important-points-aed156853112?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/aed156853112</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[customer-behavior]]></category>
            <category><![CDATA[medium]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Tue, 28 Apr 2020 15:20:25 GMT</pubDate>
            <atom:updated>2020-04-28T15:21:05.416Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Video topic detection is a technology with the potential to revolutionise how we understand customer behaviour. But there are serious challenges. How can we identify relevant topics from moving images? How can you deal with multiple topics in one video? How do we choose the best from the many available techniques?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/533/1*Fn9gEQj6DZ9hMzNjd5o_4g.jpeg" /></figure><p>In this blog, I reflect on our first trials using speech recognition as an intermediate step to tackle some of the challenges listed above, and share what we have learned so far.</p><p>Topic detection is most notably a subfield of Natural Language Processing (NLP), but there have also been research efforts towards video topic detection¹, and topic detection from speech². Our focus here is on the latter, using video and speech recognition to improve the user experience for Video on Demand (VoD) or Over-The-Top (OTT) Television services with three different use cases:</p><ol><li>Allow the user to search for programs by topic — something that is currently not possible on many platforms.</li><li>Give the user the option to search for topics within videos and start playback at the exact moment when the relevant item begins within the program.</li><li>Recommend video content related to a topic that the user has expressed interest in, or that is particularly trending at the moment.</li></ol><p>Features like these could revolutionise the way we approach viewing behaviour, and increase usability for viewers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mcJDAho-FZKMjhn_3Q-i8A.png" /><figcaption>An example: by applying topic detection to a news video, we can detect what each segment is about, when it starts and ends, or recommend it to users based on the topics they are interested in. But what is the best way to approach video topic detection?</figcaption></figure><h3>Supervised or unsupervised?</h3><p>There are two distinct flavours of topic detection, and we need to choose upfront which to use. Unsupervised detection (for example the popular LDA) involves clustering similar words and discovering topics from the emerging clusters. The results are a little like a word cloud and cannot be predicted in advance. Supervised detection involves pre-labelling topics — deciding in advance what is of interest.</p><p>Our own mission includes that we present users with a menu of topic choices — a predetermined taxonomy that naturally pushes us towards a supervised topic detection approach.</p><figure><a href="https://in.mathworks.com/help/textanalytics/ug/analyze-text-data-using-multi-word-phrases.html"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4s4E9I4C-_TuhS1AiMG5Nw.png" /></a></figure><h3>From video to text</h3><p>Extracting meaningful features from a raw video signal is difficult due to the high dimensionality. Every frame is represented by a tensor of (width x height x 3), where the last dimension represents the three RGB channels. For video content this adds up quickly: if we use common image recognition models like ResNet or VGG19 with an input size of 224 x 224, this already gives us 226 million features for a one minute video segment at 25 fps.</p><p>The best way to tackle this problem is to use text transcripts instead of video streams: conversations or voice-overs offer more direct and accessible ways to the topic. You can then apply NLP techniques to further process the data. The easiest and cleanest option is to use subtitle files as a text transcript, but these are clearly not available for all videos. We therefore used a speech recognition engine to extract a text transcript.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F411363754%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F411363754&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F883367399_295x166.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="426" height="240" frameborder="0" scrolling="no"><a href="https://medium.com/media/772611bfd2dfa3faf7b11cd62d9dac60/href">https://medium.com/media/772611bfd2dfa3faf7b11cd62d9dac60/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VTIgRANEaY59wEUcoKHqlw.png" /></figure><p>Frequently, speech recognition output contains quite a lot of errors, but typically there are enough keywords correctly identified (<em>democrats, impeachment, president, …)</em> to detect sensible topic labels.</p><p>We want to detect not only which topics are covered in a video, but the beginning, duration, and end of each topic segment. To enable this, we first split our text transcript into overlapping one-minute chunks and apply topic detection for each chunk separately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xPFJY8JmFXxFI8vrocydSQ.png" /></figure><h3>Text feature extraction</h3><p>The next step is to translate the words into <em>features</em> that can be used as input to a topic classifier. There are two alternative approaches here. You can use a Bag-of-Words approach, which results in a count of how many times each word appears in your text, or a Word Embedding model that converts every word into a vector, or <em>embedding</em> (numeric values) representing a point in a semantic space, pictured below. The idea behind these vectors is that words that are closely related semantically should have vectors that are similar.</p><figure><a href="https://migsena.com/build-and-visualize-word2vec-model-on-amazon-reviews/"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2xn5bCdnBjn2CtxAFLbYJg.png" /></a></figure><p>The advantage of using a Bag-of-Words representation is that it is very easy to use (<a href="https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.CountVectorizer.html">scikit-learn has it built in</a>), since you don’t need an additional model. The main disadvantage is that the relationship between words is lost entirely. Word Embedding models do encode these relations, but the downside is that you cannot represent words that are not present in the model. For domain-specific texts (where the vocabulary is relatively narrow) a Bag-of-Words approach might save time, but for general language data a Word Embedding model is a better choice for detecting specific content. Since our data is general language from television content, we chose to use a Word2Vec model pre-trained on <a href="https://fasttext.cc/docs/en/english-vectors.html">Wikipedia data</a>. <a href="https://radimrehurek.com/gensim/">Gensim</a> is a useful library which makes loading or training Word2Vec models quite simple.</p><p>Text transcripts are first preprocessed by removing stopwords, very short words, punctuation and numbers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JaM6ZU_BDtGWUxJ6pY6bJg.png" /></figure><p>Then, we calculate the word vector of every word using the Word2Vec model. We use the average over the word vectors within the one-minute chunks as features for that chunk. Word2Vec is a relatively simple feature extraction pipeline, and you could try other Word Embedding models, such as CoVe³, BERT⁴ or ELMo⁵ (for a quick overview see <a href="https://towardsdatascience.com/from-pre-trained-word-embeddings-to-pre-trained-language-models-focus-on-bert-343815627598">here</a>). There is no shortage of word representations with cool names, but for our use case the simple approach proved to be surprisingly accurate.</p><h3>Choosing a Classification Model</h3><p>The task of our classifier is to predict which topics are related to a given piece of text, represented as a single vector containing the document embedding.</p><p>We can use a simple feedforward neural network, but we must choose the right function for our final layer. For multiple topic labels, a sigmoid output function is the way to go; not softmax. Why? Well, a softmax function will result in a probability distribution between classes — answering relative questions like ‘what is the probability that this document is about topic A compared to the likelihood of topic B?’ This probability inevitably sums to 1.</p><p>A sigmoid function, on the other hand, models every class probability as an independent distribution. This means that the model can be interpreted in terms of independent questions like “What is the probability that this document is about topic A? What is the probability that this document is about topic B?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oRx82lK6pyleGTFByLHemQ.png" /></figure><h3>Results</h3><p>The below picture shows the full pipeline, going all the way from the source video to the output topic labels, using the steps described in detail in the previous sections.</p><p>As a final output, we take the top 5 topics as classified by our model. These are the detected topics for the example video:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cMMB-7ZSouI1jeAAu1JkoQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XaFFL_chH7WMYWD__CPdUA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EfG7u-AEk0xHAGxoLlKfgw.png" /></figure><p>Remembering that this model uses noisy speech-to-text transcripts: even with a fairly simple preprocessing pipeline the output is pretty decent! On our internal tests, we found that with this method we reach an average precision of 0.73, an average recall of 0.81, and 88% of the video snippets have at least one correct topic prediction.</p><h3>Must Knows:</h3><ol><li>Use supervised topic modeling if you have annotated training data and want to have control over which topics you want to detect. Use unsupervised topic modeling if you do not have access to annotated training data and you are okay with having less clear topic labels.</li><li>Use subtitles or speech recognition since conversations and voice-overs inside video are highly informative of the topic.</li><li>In general, a Word Embedding model is recommended over Bag-of-Words, unless you want to have an easy baseline test or the data contains a lot of domain-specific jargon.</li><li>In our initial experiments, Word2Vec embeddings averaged over a piece of text already gave decent results, and in the next steps we are likely going to see if other embedding models can bring improvements.</li><li>If you want to predict multiple topics for a single video segment, use a sigmoid layer in the classifier instead of softmax.</li></ol><p><em>Written by </em>Timo van Niedek<em>, Machine Learning Expert at Media Distillery.</em></p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/topic-detection-for-video-the-important-points/"><em>https://www.mediadistillery.com</em></a><em> on April 23, 2020.</em></p><p>[1] Allan, J., Carbonell, J. G., Doddington, G., Yamron, J., &amp; Yang, Y. (1998). Topic detection and tracking pilot study final report.</p><p>[2] Dharanipragada, S., Franz, M., McCarley, J. S., Roukos, S., &amp; Ward, T. (1999). Story segmentation and topic detection for recognized speech. In Sixth European Conference on Speech Communication and Technology.</p><p>[3] McCann, B., Bradbury, J., Xiong, C., &amp; Socher, R. (2017). Learned in translation: Contextualized word vectors. In <em>Advances in Neural Information Processing Systems</em> (pp. 6294–6305).</p><p>[4] Devlin, J., Chang, M. W., Lee, K., &amp; Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. <em>arXiv preprint arXiv:1810.04805.</em></p><p>[5] Peters, M. E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., &amp; Zettlemoyer, L. (2018). Deep contextualized word representations. In <em>Proceedings of NAACL-HLT</em> (pp. 2227–2237).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aed156853112" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How do we effectively monitor all the video data we process?]]></title>
            <link>https://mediadistillery.medium.com/how-do-we-effectively-monitor-all-the-video-data-we-process-befc497a5384?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/befc497a5384</guid>
            <category><![CDATA[media-monitoring]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[video-content]]></category>
            <category><![CDATA[development]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Wed, 25 Mar 2020 14:25:07 GMT</pubDate>
            <atom:updated>2020-03-26T12:46:44.515Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Accurately analysing video content requires a vast and diverse deluge of data. Here we’ll walk through how we’ve systematically structured and categorised our services, to get greater visibility and control over what we ship.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0-Dh83SUUOOOwIfLnpWavw.jpeg" /></figure><p>We analyse approximately 20,000 hours of video content per day at Media Distillery. We structure our analysis service by splitting it up into [<a href="https://microservices.io/">microservices</a>], meaning there are many components involved in analysing one piece of video.</p><p>Speech, face, logo, topic recogniser services and more trivial services such as dashboards, authentication interact simultaneously to achieve a single goal: extract valuable information out of video in real-time. At this scale, when things go wrong it’s hard to figure out what exactly went wrong. Reasoning in terms of inputs and outputs is not enough, and in the end we don’t have the full story.</p><p>Did we record the wrong thing or was it the configuration of one of the many services involved in the processing pipeline? Could it have been a bug in the code?</p><p>Manually tracing the pipeline to find the culprit is a time-consuming and scrupulous task. How can we get a full, cohesive picture of what is happening to the video content inside our platform? Finding a systematic way of answering these kinds of questions has become a necessity.</p><p>So we thought about the structure of the content and the processing pipeline. What do we want to get out of it? All content is broken down into three different categories:</p><p><strong>1. Video chunks:</strong> transcoded and stored</p><p><strong>2. Audio chunks: </strong>served to the speech recogniser and stored</p><p><strong>3. Video frames:</strong> served to our analysis services, for instance to the logo recogniser</p><p>In order to find out what happens to the content, we need to keep track of the components above while they pass through relevant points in the system. Such important points are:</p><p>* When the content is broken down into the components above</p><p>* When an error is caught by a service</p><p>* When the speech recogniser returns the analysis results</p><p>* When the logo recogniser returns the analysis results</p><p>* When the content is stored</p><p>Here are some methods we use to keep track of the categories mentioned above (1,2,3) according to the relevant touchpoints (*).</p><h3>1. Content-centric logs</h3><p>Logging is the process of recording each event that a process or service generates as a single line of text in a file, ordered by time. It is a powerful method that helps with debugging services. By shifting the focus away from services and processes, we recorded a line of text every time the content passed the points above. Every line of text reports a single piece of information and has a timestamp attached.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/918/1*ejSMI8-Ka5CjSDtd_BRvyw.png" /><figcaption>Fig.1 Content-centric logs</figcaption></figure><p>The content-centric logs are collected in a central place where we can simply process texts and distill important information such as latency between two operations or the cause of an error. To report the data, we use two different strategies:</p><p>- <strong>Direct reporting:</strong> Data is reported as soon as it’s generated, sending a separate message for every log.</p><ul><li><strong>Piggybacking:</strong> Logging data is attached to the content and propagated along the pipeline. It is reported after processing, where a separate service captures the communication flow and automatically extracts the tracing data.</li></ul><h3>2. Distributed tracing</h3><p>There is significant effort needed to make sense of the content-centric logs. While they are reported to a central place, we still need to connect them using the first part of the logs in order to have an overview. Moreover, with every new type of metric reported or new content structure, we need to deploy new versions of the code that build them accordingly.</p><p>Distributed tracing is a method used to profile and monitor the communication of highly distributed applications, especially the ones built using a microservices architecture. Incoming requests are traced as they hop from one service to another; various metadata is collected along the way in a central place; and the result is an aggregation that tells the whole story about how a request flowed through the system.</p><p>Well, we wanted to tell the exact same story about the video content we analyse. By shifting the focus from the communication flow to the content, we tried to apply distributed tracing.</p><h4>2.1 OpenTracing</h4><p>OpenTracing is the leading standard on [<a href="https://opentracing.io/">distributed tracing</a>] instrumentation. Similarly to the logging method, we had to define the kind of story we wanted to extract out of the system and modify the codebase accordingly. OpenTracing fixed the following:</p><p>* Every operation has some context information attached</p><p>* The API is powerful, allowing us to report any kind of metric</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*1XHu_JTwYHxQPuHKFPjKdw.png" /><figcaption>Fig.2 Sample output using OpenTracing and Jaeger on a piece of audio that goes through speech processing</figcaption></figure><h4>Drawbacks</h4><p>With the two methods described above, the best results come only after a good understanding of the codebase, platform, infrastructure and the business. Moreover, the effort of incorporating these methods is significant. Simply put, tracing or logging content-centric data has got to become part of the development process. This is because engineers know the intricacies of the features they develop while they develop them.</p><h4>2.2 Adapting the processing pipeline</h4><p>All the data we process is fairly standardised. It consists of video data, and variations come from different resolutions and codecs. Therefore, all messages that flow through the system are standardised in the form of an annotation, and all reference a piece of content.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/637/1*P0ISFwL6_oKflfVA6ImzIQ.png" /><figcaption>Fig.3 Sample automaton that applies speech recognition over a piece of video. Notice how frames and videos do not have a result. Instead, they go straight to the ERROR state.</figcaption></figure><p>Incoming data is broken down by the recording services, which is then sent forwards to be processed. To all data packets we attach an automaton which specifies and dictates the analysis it will undergo. It contains a set of states, a collection of transitions between states, a start state and a collection of possible end states.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*vIZilAc8mP0hTWg4Vhz8fA.png" /><figcaption>Fig.4 Sample output using OpenTracing and Jaeger on a piece of audio that goes through speech processing using automata</figcaption></figure><p>Services that receive the packet are responsible for performing the analysis and then updating the automaton to the next state. No matter their state, all packets are picked up by a dispatcher service that routes them to the destination based on the current state of the automaton. This way we extracted the responsibility of knowing the next step in the pipeline from each microservice.</p><p>And to achieve content-centric tracing, we only instrument the point where services change the state of the automaton.</p><h3>Content-centricity distilled</h3><p>Understanding our system’s behaviour was a bigger challenge than we anticipated, but along the journey we gained a lot of valuable insight. The effort showed us the importance of integrating such an initiative from the early stages of the development. Expanding coverage by using OpenTracing or building simple logs was tedious but it inspired us to push forward as results were promising. Having a concise overview of our system’s behaviour was closer, however still incomplete.</p><p>We ended up challenging the core architecture and we not only gained a clear picture of what happens to the video in our system, but ended up with cleaner and more cohesive services that communicate in an even more unified way, while expanding tracing coverage automatically.</p><p><em>Written by </em>Nicolae Marian Popa<em>, Backend Developer at Media Distillery and presented during the Cloudcom 2019 conference in Sydney.</em></p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/smart-tv-advertising-isnt-truly-smart-yet/"><em>https://www.mediadistillery.com</em></a><em> on March 27, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=befc497a5384" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Empowering TV users to find content in the blink of an eye]]></title>
            <link>https://mediadistillery.medium.com/empowering-tv-users-to-find-content-in-the-blink-of-an-eye-31532d1a645b?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/31532d1a645b</guid>
            <category><![CDATA[users]]></category>
            <category><![CDATA[tv]]></category>
            <category><![CDATA[viewers]]></category>
            <category><![CDATA[content-discovery]]></category>
            <category><![CDATA[streaming]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Fri, 06 Mar 2020 12:21:17 GMT</pubDate>
            <atom:updated>2020-03-06T12:21:17.603Z</atom:updated>
            <content:encoded><![CDATA[<p>We’re living deep in the <a href="https://en.wikipedia.org/wiki/The_Experience_Economy">experience economy.</a> In 1999, Joe Pine and Joe Gilmore brought out a seminal book on experience innovation, and two decades later, they’re about to release a new edition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*0EKzeHYrx2A-r87iv_IbUw.jpeg" /></figure><p>Nowhere is this ‘experience economy’ more relevant than in TV and entertainment: Netflix invented binge-watching, setting the standard for time spent digesting new content, episode after episode and for hours on end. But the shift in the experience economy today says that people won’t passively accept the experiences they’re given anymore. They want greater control over what they’re watching — which means more time spent seeing content they love and less time spent searching for it.</p><p>Now the rest of the TV world is catching up. Reducing the time spent searching for content and increasing viewer engagement rates are key KPIs for TV providers. Keeping your user’s eyes on your content is a way for you to measure how satisfied they are — it’s your NPS.</p><p>So how will your experience skip the endless browsing to serve users your unique content seamlessly, and stand out from the rest to bump up your NPS?</p><p>Here are four ways to speed up content discovery for four different user personas — serving them the right recommendations and improving NPS through smooth UX and smartly applied data.</p><h3>1. Stop search dead-ends</h3><p>Your first user is Lucas, and he’s looking for a show, a topic, trend or public figure you don’t have matching content for. He has some time to spare, so he doesn’t want a result dead-end. In fact, no one wants to hit that dreaded ‘zero results found’… if you don’t have exactly what Lucas is looking for, whether it’s a programme or a person, make sure you can take him in a direction he might enjoy — even if you don’t have the exact word-for-word match.</p><p>This is the first crucial step towards salvation for TV service providers. As more content floods the market, it gets more difficult for viewers to find what they’re looking for. Make it your imperative to give Lucas recommendations such as trending topics or the latest news to test and learn how he engages with that content, then improve what you serve — or see him quickly ‘exit’.</p><p>How can you test the stickiness of your content? Through an accurate understanding of what video content is made up of; identifying the faces, objects, and content snippets that users prefer. Machine learning algorithms can help companies with this mission to identify images, audio and text — cross-referencing millions of clips and content sources and combining them into meaningful categories for your user’s content discovery.</p><h3>2. Serve relevant content — in a different way</h3><p>Now you have some insights, don’t only use word-for-word search criteria. Instead, make a metadata search engine that applies the data your user willingly gives to you: dig inside what they watch instead of what they type, and use their search terms to serve up content featuring similar topics, interests, people, story points or moods — detected through speech and advanced moving image recognition.</p><p>Meet Zara, a diehard football fan from Manchester, UK. Based on understanding what’s inside the content Zara watches, you now don’t only know that she engages with the latest news in a broad category like ‘sports’ or even ‘football’. You understand which team or even player she prefers, as well as the specific matches she plays all the way through, and moments she reruns. You start to build a richer picture of what she wants to watch. But that’s not all.</p><p>Based on understanding her watching habits, you also better understand who she is… after all, we are greater than the sum of our postcode, age bracket, and gender.</p><p>Zara isn’t a demographic, she’s a human being. She’s a football junkie, a Liverpool FC fan, a Virgil van Dijk lover and she prefers to skip certain scenes when she watches reruns, rather than sitting through the whole match. She then wants to watch every post-match interview with Virgil; keeping tabs on the non-sports shows where he makes an appearance too.</p><p>She also appreciates the details. As well as enriching search, you can also enrich her watching experience, helping her skim through the programme detail with a film-strip summary so she can easily navigate her personal content highlights over a longer segment.</p><h3>3. Load up ‘binge-watching’ mode</h3><p>But many users won’t even bother to use the search bar before they hit exit, like your third user, Anil. And why should he? Streaming at its best means a) it’s seamless, and b) it intuits what a user wants to watch next — even when they don’t know themselves yet. If you do this well, serving recommendations means not pushing content that you want Anil to watch — but content that he would enjoy.</p><p>So spread your discovery paths beyond the main menu to keep him with you. Is he approaching the end of watching the Oscars ceremony? Serve him a shortlist of similar items he can skip to. These next-best-items can work two-fold, recommending content similar to what Anil has just watched, and content relevant to his profile. Picture this as a next step; a behind-the-scenes interview with the ‘best picture’ winning Director in a talk show that aired a few days ago. .</p><p>Then, does he generally prefer to snack on content rather than binge? Instead of letting him pause and replay small segments of the award ceremony, stitch them together into a string of shorter content snippets you know he will enjoy.</p><h3>4. Make first impressions matter</h3><p>Once you understand what your user wants through tapping into this metadata via their watching habits, you can turn their home-screen into a custom dashboard featuring visual news feeds, snackable content reels, current and undiscovered favourite shows.</p><p>This is a screen of personal highlights that potentially cuts out any unnecessary steps for your final user, Sophie. Instead of a single recommendation on a specific topic, she wants a quick overview of all the news and entertainment she missed whilst away on holiday.</p><p>This home screen doesn’t only dynamically reflect the topics or interests Sophie wants to watch — but in the right sequence and size so she can consume it the way she wants to watch it, right as she switches on. Get this content mix right and she will see your home screen as something of a personalised social media feed — a go-to of what’s going on based on her unique video content fingerprint.</p><p>It follows that visuals on the home screen — in fact, any screen — should be engaging and relevant for Sophie: this is her primary way to decide which video to watch and whether she sticks or switches the channel.</p><p>The next era in the experience economy is already here, and there’s still space for service providers to delve deeper into the TV choices of every user, depending on their unique preferences. This not only means making sure content shows up at any user interaction. It means seeing that their next recommendation mirrors and intuits what they want to watch on a more relevant level.</p><p>Compressing and improving content discovery for TV viewers is not a strategic imperative for tomorrow; it’s a survival need for service providers today.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=31532d1a645b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Turning TV habits into richer user profiles]]></title>
            <link>https://mediadistillery.medium.com/turning-tv-habits-into-richer-user-profiles-65c7bc61bc22?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/65c7bc61bc22</guid>
            <category><![CDATA[television]]></category>
            <category><![CDATA[users]]></category>
            <category><![CDATA[profiling]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Tue, 12 Nov 2019 12:37:01 GMT</pubDate>
            <atom:updated>2019-11-12T12:37:01.522Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>More people are watching OTT than ever, but that doesn’t mean advertisers can reach them in the right way. How can we build a richer picture of who’s watching what, to take the TV value exchange to the next level?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8WFhWSAKyGyyCyPzQVs6uA.jpeg" /></figure><p>The TV industry really needs to reimagine how it reaches audiences and offers service.</p><p>With 72% of SVOD subscribers seeing TV ads as ‘unacceptable’ (via Differentology), parties in the advertising chain need to start fiercely defending and differentiating their revenue streams: leaving users lost in a jungle of mindless spam and multi-screening just won’t work anymore… and neither will follow the status quo to segment in the same ways, or serve up the same old stream of content.</p><p>So let’s start solving this issue right at the heart of the TV-advertiser value exchange; the user profile. Profiles are the fuel for personalised targeting, evolving all the time in their richness and the multi-dimensionality of their attributes — mostly courtesy of digital media.</p><p>It seems ironic that television, the original mass marketing channel, is the last to become a full-fledged citizen of the profile ecosystem. But now, getting a deep understanding of TV watching habits offers unprecedented ways to build both a sustainable source of revenue and superior viewing experiences, through richer profiling.</p><p>This approach to profiling focuses on segmenting your existing audience in new ways, to get more from your current investments. Better yet, it broadens your targeting options so you don’t only have to rely on demographic data. It simply means using ready-made machine learning methodologies to extract valuable metadata and understand users from a different perspective. Let’s walk through how it works.</p><h3>Who’s Sarah?</h3><p>Meet Sarah. Your advertiser can’t trace who she is, but they can determine that she lives with her partner and two kids, has a steady high-level income and even that her car lease is about to expire.</p><p>This means it’s all go for addressable advertising: data on income, postcode, programmes watched, age group and these 3rd party characteristics give the green light for advertisers to target her with a new Mercedes Benz, ready to rent.</p><p>But this heavy reliance on PII (Personally Identifiable Information) doesn’t sit well with tightening consent laws surrounding consumer data, nor does it allow advertisers to monetise the full potential of her profile. There are deep gaps in the data.</p><p>So start to do differently: who really is Sarah, beyond her PII? How about harvesting <em>implicit </em>as well as <em>explicit </em>profile attributes attached to Sarah progressively via understanding exactly what she watches, and how?</p><h3>Interests: Cooking &amp; Climate Change</h3><p>After a month of extensive metadata analysis, you start to understand more about Sarah based not on her personal information, but her personal preferences. TV operators can start to couple viewing behaviour (which programmes she switched or stayed on) with an understanding of actual content: she skipped all sports but remained engaged throughout the UN Summit live stream, paying particular attention to Greta Thunberg’s speech which she replayed several times.</p><p>Through this advanced understanding of Sarah’s interests via facial recognition, image and speech detection within the content she is viewing, you can identify topics she likes and segment content. Then, you can scale targeted advertising to those who are particularly politically engaged or obsessively watch cooking shows.</p><h3>Product Categories: Home &amp; Garden/ Healthy Living</h3><p>Beyond which topics interest her; the actors, politicians and popstars she prefers, service providers can really up the ante when it comes to contextual advertising capabilities.</p><p>Remember the fact you targeted Sarah with a Mercedes ad because her car lease was expiring? What you couldn’t know through demographic data and 3rd party attributes was her interest in climate via her news watching habits.</p><p>Instead, you can relate content she’s watched directly to advertising industry categories (IAB) through logo, object, and topic detection: you know her mind is on cooking and she’s been consuming content around smart homes to reduce her energy usage, so you can pair her current interests with a matching commercial.</p><p>This approach means reducing risk by spreading out inventory: although TV advertising was typically only accessible for large (Fast Moving Consumer Goods) brands seeking mass marketing, more granular targeting capabilities and extended audiences can open up this space to smaller players and offer untapped sales opportunities for service providers.</p><h3>Responding to what’s inside screens with stronger experiences</h3><p>By enriching existing user profiles in these new ways, operators and broadcasters are better positioned not only start adding incremental, refined value for advertisers but also personalise experiences in a whole host of new ways.</p><p>Through activating this scalable metadata, it’s possible to repackage the kinds of content viewers love in ways which surprise, satisfy and delight their consumption needs:</p><ul><li>Correlate and stitch together shorter reels of content which reflect Sarah’s interests from across multiple broadcast channels and sources</li><li>Augment search capabilities by enriching criteria with topics, interests and public figures which go beyond what she’s actively chosen before — and instead serve up new results based on what she’s seen or heard</li></ul><h3>Turning viewing patterns into richer profiles</h3><p>Inside video content, you can find all the hidden answers you need to better serve advertisers and re-imagine experiences. Here, you can go beyond demographic data to determine not only what a viewer wants to watch and how, but their interests, hobbies and product inclinations based on what they’ve seen and heard.</p><p>Through AI-driven insights generated through millions of metadata, TV service providers who build the user profiles of tomorrow will lead the way in reshaping what valuable consumer behaviour actually means.</p><p><em>By Roland Sars, CEO of Media Distillery</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=65c7bc61bc22" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Artificial Intelligence can reshape TV for broadcasters and providers]]></title>
            <link>https://medium.com/swlh/how-artificial-intelligence-can-reshape-tv-for-broadcasters-and-providers-2bdf148ad250?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bdf148ad250</guid>
            <category><![CDATA[broadcasting]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[television]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[netflix]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Thu, 05 Sep 2019 07:46:22 GMT</pubDate>
            <atom:updated>2019-09-06T14:05:56.438Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The ‘Netflix Effect’ is in full swing, with billions of people binge-watching programmes on-demand. But TV isn’t dead — it’s just evolving: what if broadcasters and TV service providers are missing a trick to offer new kinds of value to viewers, simply by reimagining and repackaging the content they already have?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RsFsHr8fSA0K70F0M-nN4w.jpeg" /></figure><p>TV content today is a cultural zeitgeist that unfolds in real-time, to be consumed anytime. People are fed up with 5-to-12-minute unskippable ad breaks, search attempts yielding zero results, and putting up with hours-long programmes for the sake of finding the few minutes they want to watch.</p><p>But although the likes of Netflix and Disney have upended the industry, giving birth to a generation of cord-cutters and cord-nevers, TV viewership is still at a <a href="https://www.digitaltveurope.com/2019/08/05/vod-and-cable-drives-reverse-in-us-tv-viewing-decline/">two-year high in the US </a>, and video services struggle to take hold of programming genres like <a href="https://www.nytimes.com/2019/07/02/business/media/netflix-talk-shows.html?smid=nytcore-ios-share">talk shows and live sports </a>— content which injects a much-needed ‘did-you-see-this?’ watercooler effect among people lost in the midst of a Stranger Things marathon.</p><h3>The topics of TV, the seamlessness of streaming</h3><p>TV’s to-the-moment, topical and snackable nature, its long-standing institutional trust, and its distribution of high quality content from many sources, are its strengths — what sets it apart from challengers. What traditional TV service providers miss and that digital-first does best are the agile, user-driven capabilities to buffer, test and release new services as fast as tastes change. This is a critical blind spot.</p><p>TV service providers still have an opportunity to reinvent themselves and harness the best of both of these worlds, but they need to act fast.</p><p>I’d like to propose 3 ways that TV service providers can play to their strengths and use their resources wisely, to reclaim and sustain a significant slice of the attention economy.</p><p>In doing so, they address three elephants in the TV room:</p><ul><li><strong>Better understanding what’s inside programmes to deliver superior user experiences, by serving existing content in new ways that viewers genuinely want to pay for</strong></li><li><strong>Creating richer, more scalable targeting capabilities</strong> for advertisers, based on user trust and value</li><li><strong>Making smart use of limited resources</strong> by using Artificial Intelligence to interpret vast amounts of video content, and automatically cross-reference it with viewing data for deeper insights.</li></ul><h3>1. Easier in-content search</h3><p>One low-threshold way to add value for viewers is to help them find existing content in new ways, through synthesised metadata. Imagine not only knowing the programmes your viewers previously watched but the topics, TV personalities and even the parts of the episodes they prefer. By using a smart search engine to recognise faces and voices inside video, TV services can offer:</p><ul><li>Deeper search criteria such as subjects and featured people</li><li>The ability to search for segments and snippets of content as opposed to entire programmes, say for example only Formula One highlights within an hour-long sports update</li><li>A treasure-trove of content they wouldn’t normally find, alongside the convenience of finding what they were looking for with greater granularity, by going beyond word-for-word search terms.</li></ul><p>This rich layer of metadata is only a way of reading and surfacing the wealth of licensed TV content service providers already have. Through recognition across faces, texts, speech and object extraction, as well as on-screen overlay elements such as nametags and credits, viewers can tune in anytime to receive the right string of videos or clips, in the right measure.</p><h3>2. Snackable video services</h3><p>the right content through richer metadata can be built into you want to consume it. Have you got five minutes to catch up on Coachella festival highlights, before you head to work? Or do you have three hours to binge-watch all of Beyoncé’s best moments?</p><p>Snackable content capabilities use smart models to scrape the seconds or minutes you don’t want to miss over an entire weekend of live coverage, a series of episodes or a movie marathon, and stitch them back together based on preferred topics, search terms, or previously watched videos.</p><p>TV providers need to forego monolithic experiences where users need to skip scenes or fast-forward to the clip they’re looking for, and instead anticipate and serve right-size content for every occasion — stitched together in real-time based on a user’s ever-shifting behaviours and multi-faceted tastes.</p><p>This is something we don’t see currently in our regular TV subscriptions, but which users gravitate towards YouTube for. What if traditional TV players were to evolve into smart stations for snackable content like YouTube, without the annoying minutes-long broadcast TV ad breaks… what if instead, service providers could go beyond serving demographically-driven ads, to help enrich user profiles based on what was inside the content they consume?</p><h3>3. Topic-based profile enrichment</h3><p>In the struggle to stay relevant, it’s time for TV to offer more meaningful, trust-based ways to target users based on deep content analysis. By creating smart segmentation using implicit metadata like what’s inside content, as opposed to solely relying on explicit Personally Identifiable Information, brands are able to divulge and harness genuine impressions and interactions, profiling by personality and interest — ‘Formula One fans’ or ‘Beyoncé lovers’ — as well as, or instead of, demographics and geolocation.</p><p>This serves as a safeguard from crossing any tricky regulatory red tape, as well as offering a stronger proposition to both advertisers and users based on clear shared value: what is someone watching? What does it really say about them?</p><p>Content-based profile enrichment doesn’t only mean acquiring new, sustainable sources of revenue; it gives TV service providers the opportunity to educate the market about what the TV advertising ecosystem should be all about: content-led consumer behaviour, tailored experiences and applying AI-driven insights across in-video aspects, instead of a spray-and-pray ad-serving approach to entire swathes of the population.</p><h3>From linear broadcasting to user-driven resequencing</h3><p>Broadcasters and TV service providers don’t have to go it alone in their bid to take apart their content, reassemble it, and deliver new kinds of value. Rather than spending billions on proprietary technologies to do this, in-content insights can be brought to scale through working with applied AI specialists. These accessible, purpose-built machine learning models can then retrieve deeper intelligence in an existing backlog of video, by cross-referencing voices, languages, faces, and texts with one another, then activating them where needed.</p><p>It’s time for TV service providers to deconstruct and re-sequence their content, to serve up right-size, on-topic experiences. From smarter search to snackable content and richer profiles, the players who differentiate through their content metadata will manage stay relevant for expectant audiences against the odds, without the need to reinvent the wheel.</p><p><em>This article has been written by Martin Prins, Product Owner at Media Distillery.</em></p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/how-ai-can-reshape-tv/"><em>https://www.mediadistillery.com</em></a><em> on September 5, 2019.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bdf148ad250" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/how-artificial-intelligence-can-reshape-tv-for-broadcasters-and-providers-2bdf148ad250">How Artificial Intelligence can reshape TV for broadcasters and providers</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Smart TV advertising isn’t truly smart…yet]]></title>
            <link>https://mediadistillery.medium.com/smart-tv-advertising-isnt-truly-smart-yet-3b48331e2a06?source=rss-35ebb815930c------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b48331e2a06</guid>
            <category><![CDATA[audience-targeting]]></category>
            <category><![CDATA[tv-advertising]]></category>
            <category><![CDATA[television]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[Media Distillery]]></dc:creator>
            <pubDate>Wed, 24 Jul 2019 13:48:28 GMT</pubDate>
            <atom:updated>2019-08-02T13:35:46.039Z</atom:updated>
            <content:encoded><![CDATA[<p>How can traditional service providers apply deeper understanding to TV experiences, to generate new sources of advertising revenue?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*8G7xsA2D_xdO5ko3at-Jaw.jpeg" /></figure><p>It’s time for TV to reinvent itself. Not only are viewers switching from Linear TV to Over The Top (OTT) video in ever-greater numbers, but consumer behaviours are changing rapidly too- a double-edged sword which will throw broadcaster and operator advertising models into jeopardy.</p><p>You’ve heard it before: streaming services have raised the bar in terms of how they’ve swept up their share of the attention economy. Subscription models and the switch-or-skip capabilities of on-demand services mean that users are fully in control, and they aren’t accepting ads anymore.</p><p>At this tipping point for TV, where can traditional services tap into new and scalable revenue sources, and redefine watching experiences?</p><p>How can linear TV providers save themselves from sinking into irrelevance?</p><p>Here, I explain how TV watching experiences can be redefined to keep up with consumers and keep advertisers sated; an approach which is made possible by Machine Learning, to deliver deeper content recognition and smarter user insights.</p><p><strong>Make as-it-happens TV more addressable</strong></p><p>ProSiebenSat.1 and RTL Deutschland are teaming up to drive the development of addressable TV, whilst Roku and Apple TV are set to drive advertising growth to new heights.</p><p>The future of TV is smart, and advertising strategies need to follow suit. It makes sense for brand-building to digitise and move to hybrid and sustainable monetisation models — and this means user experience (UX) comes first every time.</p><p>UX-first means TV is addressable for both users and advertisers: whilst users sit in the driving seat, advertisers gain a scalable understanding of where their audience is going, and the non-linear roads they take to get there.</p><p>After all, the attractiveness of addressable TV lies in its targeting capabilities. With programmatic technology, this provides an opportunity to approach advertising inventory with immense granularity.</p><p>But let’s look further and add an extra layer: it means in-depth content personalisation, using applied Artificial Intelligence (AI) to peer inside video and understand user decisions. It guarantees end-to-end viewing and enriches user-profiles based on their watching behaviour- not only based on their demographic and geolocation.</p><p>This extra layer can empower advertisers to apply user preferences at scale across anymedia environment, and any connected device. After all, smart TV isn’t truly smart until it has a deep understanding of the content it serves.</p><p>How does it work?</p><p><strong>Time to address content over context?</strong></p><p>With greater data, comes greater responsibility. Regulation has been a big obstacle for TVadvertising and having a more nuanced understanding of consumer context via cookielessTV is at least tricky and at most a GDPR nightmare.</p><p>Service providers need to overcome significant difficulties around data collection and audience targeting, so our most personal entertainment channel can stay relevant, without getting creepy.</p><p>The answer is to enrich profiles with content preferences, so as not to solely rely on collecting audience data.</p><p>A near-future of video watching will force TV broadcasters and operators to think outside of that tightly regulated, digitally-connected box to consider the context of what a viewer is consuming. Rather than their personal information and immediate surroundings, instead think topic detection, facial recognition, watching habits and search behaviour — and not only age, gender and geotargeting.</p><p>What do the closely personalised Total Video experiences of tomorrow look like? What does it mean to closely understand what someone is watching, and what does it really say about them?</p><p><strong>Ads and content which add up</strong></p><p>As our ability to understand video content grows, the visual experiences we deliver toconsumers can get exponentially cooler — and more customised, through applied AI to understand content preferences.</p><p>Your sports-crazy user can not only binge-watch Serena Williams in the run-up to the Wimbledon final by working through her backlog of related catch-up features, but these are also completely tailored to her search behaviour.</p><p>Then, if she shows specific boredom patterns by switching between tennis match clips, and Bachelor? The sequence of her recommendations shifts accordingly, mirroring her attention span and the rhythm of her topic preferences.</p><p>When it comes to watching the Wimbledon final, does your user pause and replay the close-up of Serena serving, as you get a glimpse of the brand and the range of her sportswear? After a couple of replays, can this qualify her profile for potential advertising campaigns based on logo exposure and active interest?</p><p>This really ups the ante on the kind of actionable data TV operators and broadcasters can offer to enrich user profiles.</p><p>Again, deep recognition within video content doesn’t tell you that your Serena fan is in fact female; it doesn’t tell you where she lives, nor does it offer advertisers any information which personally identifies her: nevertheless, an in-depth understanding of what this user watches still happens to be a better reflection of who she is — a more representative attribute than her home address.</p><p>This new world of ‘advertising’ allows TV service providers to not only become premium publishers based on localised audience targeting but opens up doors to reeducate the entire ecosystem about what valuable data and consumer behaviour actually mean.</p><p>To create a TV advertising ecosystem which doesn’t feel like advertising, service providers should use a pinch of explicit insight, but place a larger emphasis on profiling with implicit and inferred metadata; data which is willingly used as currency by the viewer in exchange for superior content experiences.</p><p>In this transaction, the user finds what they are looking for with the ultimate convenience; TVoperators and broadcasters escape extinction, and both brands and traditional players can avoid solely relying on demographic data which has a high regulatory price tag and looming expiry date.</p><p><strong>UX-y and you know it</strong></p><p>As linear television and its stakeholders struggle to stay relevant in an industry now dominated by Total Video, service providers reach an inflection point where finding new ways to monetise users is a must.</p><p>But users are in control now, and changing revenue models need to be on their terms. That’s why we enter a world of TV advertising driven by superior user experience:</p><ul><li>Embracing addressable TV broadens the scope for broadcasters and operators to make the most of deep content analysis</li><li>By making the most of smart devices, service providers can focus on tailoring content experiences and applying AI-driven insights, instead of targeting entire demographics en masse</li><li>This means enriching profiles in creative new ways, and producing audiences based on their preferences whilst steering clear of regulatory red-tape</li><li><em>Don’t go it alone</em>. It’s time for TV providers to forge new partnerships with applied AI specialists, to make sure precise video insights can be extracted and used at scale.</li></ul><p>In the near-future, ads won’t feel like ads anymore. Instead, they will look like personal end-cards or interesting interludes made up of content users genuinely like, want to watch next, and will click.</p><p>The TV ecosystem will know this based on billions of data points about the videos users consume, the topics they love, the parts of the story they pause at, and even what this says about their emotional state. And these insights are going to be greater than the sum of their parts in mirroring our ever-changing behaviour, generating better experiences, and steering more sustainable revenue models: it’s do-or-die.</p><p><em>Written by Stephanie Magnan, Head of Marketing at Media Distillery</em></p><p><em>Originally published at </em><a href="https://www.mediadistillery.com/smart-tv-advertising-isnt-truly-smart-yet/"><em>https://www.mediadistillery.com</em></a><em> on July 24, 2019.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b48331e2a06" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>