<?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 Rosie Hood on Medium]]></title>
        <description><![CDATA[Stories by Rosie Hood on Medium]]></description>
        <link>https://medium.com/@rkhood?source=rss-1b2e95541e1b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*0Da5064k4ccA2XqUV3zb9g.jpeg</url>
            <title>Stories by Rosie Hood on Medium</title>
            <link>https://medium.com/@rkhood?source=rss-1b2e95541e1b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 17:23:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rkhood/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[Looking to the future: Introducing EDITED Women]]></title>
            <link>https://rkhood.medium.com/looking-to-the-future-introducing-edited-women-6e6b1a6fa1b7?source=rss-1b2e95541e1b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e6b1a6fa1b7</guid>
            <category><![CDATA[inclusion]]></category>
            <category><![CDATA[women]]></category>
            <category><![CDATA[community]]></category>
            <category><![CDATA[empowering-women]]></category>
            <category><![CDATA[women-in-tech]]></category>
            <dc:creator><![CDATA[Rosie Hood]]></dc:creator>
            <pubDate>Tue, 16 Mar 2021 14:58:36 GMT</pubDate>
            <atom:updated>2021-03-19T14:53:28.440Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*g_H4D9o0nmUrmMsJQwk0pg.png" /></figure><p>We’ve launched EDITED Women this month, a new employee resource group. And it has been a journey.</p><p>Inspired by the creation of EDITED’s first ERG, <a href="https://inside.edited.com/introducing-edited-black-an-erg-committed-to-supporting-the-inclusion-of-our-black-employees-88ef347a0ddb">EDITED Black</a>, we formed a planning committee using our hack days in October. The goal of the hack days was to focus on projects that could improve the working from home experience. It opened the opportunity to conceptualize EDITED Women and simultaneously formed the largest hack day group project at EDITED.</p><p>It took 6 months to get to launch. Between mission statements, charters and panel reviews to get through, we also needed to take the time to share our own individual journeys. During our first meeting, we shared experiences of how gender had shaped and affected each one of our careers. It was empowering to hear the strength of these amazing women. While EDITED has a predominantly female representation (67% of our team are women, with 71% in senior management), that one meeting gave us a safe space to have these types of discussions. From that day forward, we decided EDITED Women would first and foremost be a support group and an open forum for anyone to join, and second, a foundation to see real change.</p><p>We dedicated 6 months to make an ERG based on our own experiences, that felt constructive and fostered a comfortable space for anyone to join. So, what are we here to do?</p><p>EDITED Women strives to <strong>Support</strong>, <strong>Promote</strong>, and <strong>Hire</strong> women in tech at EDITED, and encourage these practices to be shared in the wider tech community. We seek to educate peers and senior management to understand unconscious bias and push for a zero tolerance attitude towards gender discrimination. EDITED Women will be a safe and empathetic environment for all women.</p><p>With our mission statement, let’s break that down into 3 objectives:</p><ul><li><strong>Support</strong>: We will foster a fair and safe environment at EDITED for women in tech. We will provide a forum for women to discuss related issues. We will encourage the wider company to engage with our community. We will put practices in place that foster a safe, empathetic and constructive chain of communication around issues of gender equality.</li><li><strong>Promote</strong>: We will work towards progression and fair promotion of women at EDITED. We will move away from gender-biased words amongst peers and managers, as well as help to educate around language, culture and stereotypes.</li><li><strong>Hire</strong>: We will actively work towards increasing women representation at EDITED by attending events, collaborating on the interview processes and encouraging the next generation of women to pursue careers in tech. We want EDITED to innovate and be an example for the promotion of women in tech.</li></ul><p>Beyond supporting each other in our daily roles, we want tangible growth for ourselves and new women joiners. ERGs have the potential to inspire and innovate what may feel like an unmovable workplace. They encourage people to speak out by influencing a roadmap, shaping recruitment in their team, elevating a voice they feel isn’t being heard and reversing the direction of a conversation. This makes for a more diverse, inclusive and ultimately, stronger team.</p><p>As we commemorate International Women’s Day, we’re celebrating the achievements of women at EDITED and looking to the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e6b1a6fa1b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Knowledge graphs applied in the retail industry]]></title>
            <link>https://medium.com/data-science/knowledge-graphs-applied-in-the-retail-industry-ecac4e7baf8?source=rss-1b2e95541e1b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ecac4e7baf8</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[retail]]></category>
            <dc:creator><![CDATA[Rosie Hood]]></dc:creator>
            <pubDate>Tue, 07 Jan 2020 16:18:34 GMT</pubDate>
            <atom:updated>2020-01-07T18:15:50.626Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1013/1*Ng6rQIWY1k-tP970rGhorA.png" /></figure><p><strong>TL;DR</strong>:<em> Knowledge graphs are becoming increasingly popular in tech. We explore how they can be used in the retail industry to enrich data, widen search results and add value to a retail company.</em></p><p><a href="https://edited.com/">EDITED</a> ended 2019 with its annual Hack Week, the perfect opportunity to test new ideas with minimal risk, along with promoting collaboration across teams and specializations. My team worked on an area of research I’ve been keen to explore this last year: knowledge graphs. How can they be leveraged and optimized in the retail industry?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/370/1*7LA-16wjJXQWPjneLWSCNA.png" /><figcaption>An example of a Google knowledge panel</figcaption></figure><h3>The rise of knowledge</h3><p>Knowledge graphs are <a href="https://trends.google.com/trends/explore?date=today%205-y&amp;q=knowledge%20graph">exploding</a> in the tech industry. They’re used by voice assistants, online retail recommenders and search engines, and count tech giants such as Facebook, Google and Netflix as some of their biggest advocates. You most likely see them everyday in the form of a box or “knowledge panel” that appears when searching with Google. Google uses data scraped from websites like Wikipedia that use a <a href="https://en.wikipedia.org/wiki/JSON-LD">JSON-LD</a> schema to declare a page’s structured data, as well as its own understanding of user behavior to expand your search results. Take this search for <em>David Bowie </em>(left). Google returns summary information from Wikipedia such as birthplace, date of birth and spouses, as well as popular songs and movies and the user-defined “People also search for”. This results in a more informative panel of information, where data is entity-centric aka “<a href="https://www.blog.google/products/search/introducing-knowledge-graph-things-not/">things not strings</a>.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mzb2NjqWWPa_xpWRI85fJA.png" /><figcaption>David Bowie as a graph</figcaption></figure><p>If we were to view David Bowie’s filmography data as a graph it would look something like the above: David Bowie (represented as a person node in red) connected to various movies (movie nodes in blue) via an ACTED_IN relationship. We can extend this further with data related to these nodes (such as other actors from the films) or data associated with the person node (such as birth year in yellow).</p><p>This simple graph is able to store data in an intuitive way, allowing for queries of a kind (such as following links) that would be difficult for a conventional RDBMS. We can see relationships where each path from a node has an associated meaning like “acted in”. This forms an <a href="https://en.wikipedia.org/wiki/Ontology_(information_science)">ontology</a> where we have domain-specific knowledge represented as properties, connections and relationships between entities, concepts and data.</p><h3>A retail product graph</h3><p>Lots of data fit into ontologies, and retail is no exception. Let’s take a look at how a product from EDITED can be viewed as a product graph. For ease, we have only included a subset of product data and fields. There are many packages available to construct graphs, we use the popular <a href="https://neo4j.com/">Neo4j platform</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/385/1*vmhmwkFjJHfHKz8t4KTi_Q.png" /><figcaption>An EDITED product as a graph using Neo4j</figcaption></figure><p>Here we have a product node (orange) named with a unique identifier. From this product node we have several relationships pointing to other nodes including: HAS_ATTR (green), COLOR (blue), CARE (pink), BRAND (red) and CATEGORY (beige). HAS_ATTR refers to an attribute, e.g. this product contains a fastening, and along with CARE is found using a simple <a href="https://en.wikipedia.org/wiki/Tf%E2%80%93idf">TF-IDF</a> word extraction (using weights to isolate words that are important attributes) on the name, description and care of a product. Brand is directly taken at the time of crawl. While other relationships, CATEGORY and COLOR, are found using text and image-based machine learning.</p><p>Storing retail data in this format creates a parsable data structure, highlighting both relationships to a product as well as other products. For example, other products that have a “ruffle” feature will also point to the “ruffle” node, as shown below. This structure also allows data to be continually ingested from a pipeline as we crawl, as the graph is able to adapt. By storing data as a graph, we can also apply algorithms formulated using graph theory such as utilizing shortest paths for nearest neighbors (which may aid <a href="https://towardsdatascience.com/unravelling-product-matching-with-ai-1a6ef7bd8614">product matching</a>) and link prediction, where we are able to predict relationships between nodes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1009/1*mNss0_QRhVGzSi3nH1knvg.png" /><figcaption>Extracting other products that share a “ruffle” feature with our product</figcaption></figure><h3>Getting caught up in semantics</h3><p>So far we’ve talked a lot about representing our data as a graph, but have we been discussing knowledge graphs? There’s some mixed opinion on this topic, but I prefer to define a knowledge graph as an ontology where domain-specific knowledge has been added (see for example, <a href="http://ceur-ws.org/Vol-1695/paper4.pdf">Ehrlinger et al 2016</a>). Though our product data uses machine learning to extract product attributes we use as nodes in a graph, to me a knowledge graph uses external data or an understanding of retail to construct ontologies.</p><p>For example, EDITED uses Burberry as a brand node in our product graph. If this were to exist in a knowledge graph it would have information stored in nodes such as when and where the company was founded, along with certain attributes associated with the brand such as heritage, print, luxury, Riccardo Tisci (the chief creative officer) and so on. This requires orchestration and thorough research by retail experts. To make this knowledge graph distinctly separate from our product graph the nodes are colored in yellow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yfphWfCXTAgQwPVgZF9veA.png" /><figcaption>Example subset of a retail knowledge graph using words related to Burberry. The node types include year, company, country, person and associated words.</figcaption></figure><h3>Query expansion — enhancing a search engine</h3><p>Now we have a retail knowledge graph (where yellow nodes are entities in the retail industry with semantic connections) and an EDITED product graph (multi-colored nodes of product data like color and brand), we can utilize the two to do some advanced query expansion.</p><p>We use semantic information in the knowledge graph to get words that are related by cultural context. Take our above example of words related to Burberry, specifically the nodes connected to “Britain” via “BRAND_FROM”. We can use this cluster of the knowledge graph to return British brands like Burberry and Paul Smith.</p><p>Below we have a Burberry product and a Paul Smith product in our product graph related to each other by their connection in the knowledge graph. They are also related to each other in the product graph by some shared nodes (e.g. dress, shirt, print, etc). So if we were to search for “British designer shirt dress”, our knowledge graph would return results based on British brands and a simple search (like fuzzy) ensures we focus on shirt dresses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/677/1*2CdiuG4lpN8alK4WelQPMQ.png" /><figcaption>Relating products by their brand’s country of origin</figcaption></figure><p>Semantic searches are common in retail, particularly something like “sassy dress” or “dress for dancing”, will occur more than a search for a specific product name like “black asymmetrical sequin mini dress”. We can also use a knowledge graph based on product attributes and semantic entities to associate words together, not so much an ontology as keyword association, to construct simple rule-based classification allowing for quick scaling. Again, industry experts are crucial to this process to choose those words based on their knowledge of retail. Below we have a subset of our knowledge graph — a curated list of words related to “party”. We can use this to expand a customer’s search.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/885/1*r8qK76Mrb54c-IiL9ZFx3w.png" /><figcaption>A subset of a retail knowledge graph with words related to party</figcaption></figure><p>By searching “sassy dress” we return all curated words related to “sassy” from the knowledge graph within two “path lengths” (i.e. arrows). We can also extend this to path lengths &gt;2 to tap further into the knowledge graph connections, for example, to return brands that are known for these attributes. This returns a “party” cluster to query our product graph, returning all products that have party-related words as attributes (i.e. via the HAS_ATTR relationship), thereby expanding our original query. We can repeat this process using any part of our knowledge graph, for example, “leather jacket” uses a fabric ontology (such as leather and suede), and “sustainable” uses a sustainable ontology (linking brands, materials, associated words, etc).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*fpWGRWgrFYmI-zUvg0i2mw.png" /><figcaption>Using a retail knowledge graph to query our product graph using query expansion</figcaption></figure><p>This method of connecting a knowledge graph to a product graph also allows us to probe our retail data further. We can begin to see trend clusters forming by observing increased connections to certain words, as well as brand clusters where there are an increased number of shared nodes between brands.</p><h3>Final thoughts</h3><p>Knowledge graphs are proving to be a big contributor to the future of tech and soon the next big revolution in retail. We covered just a small part of their potential in retail over the course of a couple of days. However, they can add serious value to retail companies looking into improving search expansion, personalizing online recommendations, predicting trends, product matching and much more. We’re looking forward to exploring all of these possibilities at EDITED.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ecac4e7baf8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/knowledge-graphs-applied-in-the-retail-industry-ecac4e7baf8">Knowledge graphs applied in the retail industry</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unravelling product matching with AI]]></title>
            <link>https://medium.com/data-science/unravelling-product-matching-with-ai-1a6ef7bd8614?source=rss-1b2e95541e1b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a6ef7bd8614</guid>
            <category><![CDATA[retail]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[retail-technology]]></category>
            <dc:creator><![CDATA[Rosie Hood]]></dc:creator>
            <pubDate>Tue, 26 Nov 2019 13:37:23 GMT</pubDate>
            <atom:updated>2019-11-26T15:19:08.827Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/996/1*qzs_Uowaoi2L1upiNRqOzQ.png" /></figure><h3>Unravelling product matching in retail with AI</h3><p><strong>TL;DR</strong>: <em>Product matching, the ability to match the same product as sold by multiple retailers, is integral to the progression of retail. We explore a machine learning approach to provide high accuracy and scalability with the world’s single biggest source of real-time retail data.</em></p><p>Product matching is an important area of research in retail technology. From a retailer perspective such as ASOS and Zalando, ensuring consistent pricing and discounting amongst competitors is a priority, from the single product level to whole assortment overlaps. A retailer must remain competitive in its market. However, an awareness of products that retailers are not selling or exclusively promoting, also leads to a significant impact on profit. Retailer exclusivity (i.e. when there are no matches) can be lucrative in retail. For any established brands such as Nike and Burberry who are expanding into different markets, or selling through multiple retailers, it can be difficult to track a product’s pricing and desired description. Brands want their products represented uniformly, and retailers may be underselling them in both senses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*80B91U31cqDooQB-WZNEPQ.png" /><figcaption>Matching products can have different images, names and descriptions</figcaption></figure><p>At <a href="https://edited.com/">EDITED</a> we have data on over 500 million retail products. These products are sold by numerous retailers in multiple languages and markets. They can be sold at different times, under different branding (Tommy Hilfiger vs. Tommy Hilfiger Underwear) and with retailer-chosen images and text. Their product information can range from an extensive narrative to the singular word description “dress” or a numerical identifier. Looking at the above example, based on real data, each product shows a variation in name, description and image. In a dataset of 500 million, these products are not unique or similar enough to make product matching a simple task. How can we match these products exactly? More importantly, how can we have confidence in these matches? Let’s take a look at how this can be achieved scalably and with high accuracy.</p><h3>Every byte of data helps</h3><p>To understand whether two products are the same we need as much overlap in product data as possible. EDITED crawls multiple text fields, images and metadata (e.g. time of crawl) from retailer web pages. All of these are considered as “features” in our models. For example, some common text-based features include the product’s name, the product description provided and the product’s brand.</p><p>We need to convert this data into something a computer can process. First, let’s take a look at our text-based features. Natural language processing (NLP) is a subfield of machine learning, which aims to understand how computers analyze human language. One common NLP task is word embedding, mapping words to vectors that represent them, and many models exist to perform this task. At their simplest, this can mean a count of the words in the text, where the length of the vector is equivalent to the vocabulary size with each element representing a word. At their most complex, this can be a model trained on millions of words or phrases to produce fixed length vectors based on features learned during training, where the representation of a word is distributed across several elements in a vector (e.g. <a href="https://arxiv.org/abs/1810.04805">BERT</a> and <a href="https://arxiv.org/abs/1906.08237">XLNet</a>). This complexity of word embeddings must be considered in tangent with the ability to scale when designing a model infrastructure.</p><p>Next, we need to reimagine this process so that our images are represented by vectors. This is a large area of machine learning research, known as computer vision. At its most basic, we can simply check if two images are identical by passing them through a hash function. However, if there is any discrepancy, even at the single pixel level, we will be unable to match the images. Instead, we use a convolutional neural network (CNN), which is a deep learning algorithm for our image recognition. CNNs use moving filters on images to find its characteristics, whilst reducing the amount of data that needs to be processed. At the lowest levels, these filters extract features such as edges. The network then learns how these features can be combined to form complex structures, gaining more insight into the components that define a product. It then uses this to classify a product, for example, as a top or dress. These are called “feature vectors”, where a vector’s elements correspond to learned features from the CNN.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FD6AAQqGqeO9tr-DAohfhg.png" /><figcaption>Our model architecture using word embeddings derived from text and feature vectors learned from images</figcaption></figure><h3>Mapping the world’s largest real-time retail dataset</h3><p>With these two vectors per product gained from text and images, we can visualize the relationships in EDITED’s database. We do this by using our vectors as a high-dimensional coordinate system, where each element corresponds to a coordinate in a separate axis. For example, imagine a vector with a length of 2 elements. This would mean one element in an <em>x</em>-axis and another in a <em>y-</em>axis, like a simple graph, as shown in the above figure’s scatter plot. Expanding to 3 elements, the graph now has a <em>z-</em>axis, and can be visualized by a 3D plot. Continue this process until we have a few hundred or thousand elements in a vector, and with that many dimensions it becomes difficult to conceptualize.</p><p>This is a high-dimensional mapping of products, where products near each other will share similar characteristics due to their feature vectors and word embeddings. This translates to a distance between the product we wish to query and its surrounding data points or products, where shorter distances are closer matches. Collapsing to 2D (using a dimension reduction technique, such as <a href="https://en.wikipedia.org/wiki/Singular_value_decomposition">SVD</a>), we can see the complexity with just a handful of data points below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*ZKee6LGMWUnSyETz-DnbJg.png" /><figcaption>A 2D representation of mapping for a subset of EDITED data, color-coded by brand</figcaption></figure><p>At EDITED we have a high-dimensional mapping of 500 million products! We call this an “index”. Checking each product in our index for one query product is unrealistic in terms of storage and the time taken to build and query the index. How can we realize which products are near each other and crucially, do it quickly?</p><p>Enter another thriving field in data science, Approximate Nearest Neighbors (ANN), involving contributions from some of the world’s largest tech companies, including Spotify’s <a href="https://github.com/spotify/annoy">Annoy</a> and Facebook’s <a href="https://github.com/facebookresearch/faiss/">FAISS</a>. These algorithms balance accuracy with speed by providing a guess of a product’s nearest neighbors. At their core, ANNs split data randomly and repeat the process until a binary tree structure is created, where each leaf is a subset of data after each split and branches are used to recall neighbors. By randomizing the splits and creating repeated binary trees, we obtain a good estimate of the nearest neighbors.</p><p>The latest developments in ANNs use a graph-like structure for the data, known as <a href="https://arxiv.org/abs/1603.09320">Hierarchical Navigable Small World graphs</a>, which are leading the <a href="http://ann-benchmarks.com/">benchmarks</a> for accuracy and speed. Here, a query begins a traversal of the index, where at each stage the distance between the query data point and its neighbors are noted, per node of the graph. The next node to traverse is selected based on minimizing this distance, and the best neighbors again noted. This process finishes when a certain criterion is met defined by the user, such as the number of calculations to perform, returning a better estimate of the nearest neighbors.</p><h3>Evaluation, Evaluation, Evaluation</h3><p>Finally, we want to understand how well our system performs. When we query the neighbors of a product from our high-dimensional index, we need to check whether our neighbors are indeed matches and if we have returned all possible matches. This is the same as precision — the proportion of neighbors that are correct — and recall — the proportion of matches that are returned. Further to this, we need matches and more relevant neighbors returned at higher ranks, so there is confidence in the index.</p><p>By returning matches based on distance to a query product, we are using a “ranked retrieval” system. As we return a high number of neighbors (e.g. 100) we should be able to achieve total recall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/703/0*L3uOn5C7oRX-0AYu" /><figcaption>Paul Verhoeven’s Total Recall</figcaption></figure><p>However, we also want to assess the accuracy of queries where total recall cannot be achieved. For this we modify our metrics to “precision at rank <em>k</em>”, the proportion of <em>k </em>neighbors that are correct, and “recall at rank <em>k</em>”, the proportion of matches that are in our <em>k</em> neighbors, where <em>k</em> is the number of neighbors returned (e.g. 100). As we filter our neighbors by distance, <em>k</em> will vary per query. We need to ensure precision is maintained for changes in <em>k.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_aHKw8DMe9XYiclnJglltA.png" /><figcaption><em>Calculating average precision in a ranked retrieval system for a query with four matches</em></figcaption></figure><p>For this we can calculate the “average precision”, which does not require a fixed <em>k. </em>The methodology is practicable — we iterate through our returned <em>k</em> neighbors filtered by distance, and at each iteration if the product is a match we calculate the precision at that rank. We then use the average of these precision values, as demonstrated in the diagram above. This allows us to penalize non-matches with a higher rank more severely, while also considering non-matches at lower ranks. Finally, we calculate the “mean average precision” for multiple queries, which is the mean of the average precision calculated for each query. This has proven to be an invaluable metric at EDITED, as it demonstrates understanding of precision and recall, while prioritizing user-focused features such as ranking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ohLLYJvQ2Od_ST-QkvPvCA.png" /><figcaption>Ensuring consistent pricing and discounting amongst competitors. These matched products have different names, brands, descriptions and images</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iLaTrZ1MHs-3Ct68jUXEaQ.png" /><figcaption>Tracking consistency in a brand’s pricing and descriptions across markets and retailers</figcaption></figure><h3>Key Takeaways</h3><p>Our model uses text and image-based features to create a high-dimensional mapping of how products are related, providing product matches despite varying descriptions and images. By using the latest machine learning advances, we are able to extract the same product across multiple retailers, languages and markets with precision and confidence and at a scale unprecedented in retail.</p><p>Product matching is a difficult challenge to tackle in retail, fundamentally due to the variance and sheer size of data. By sharing our methodology at EDITED, we hope to have provided insight into how to approach this formidable task.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a6ef7bd8614" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/unravelling-product-matching-with-ai-1a6ef7bd8614">Unravelling product matching with AI</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to become an AI-driven company]]></title>
            <link>https://medium.com/data-science/how-to-become-an-ai-driven-company-e8816453c786?source=rss-1b2e95541e1b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8816453c786</guid>
            <category><![CDATA[retail-technology]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[retail]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rosie Hood]]></dc:creator>
            <pubDate>Tue, 30 Jul 2019 12:32:05 GMT</pubDate>
            <atom:updated>2019-07-30T20:31:20.702Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sr2pEEYASyeoJ049Kqekxw.png" /></figure><p><em>AI is transforming how we do business at an unprecedented pace, but the transition to becoming AI-driven is easier than you think. Now is the time to invest and remain at the top of your game.</em></p><p>A few weeks ago Artificial Intelligence was thrown into the spotlight as the winners of this year’s Turing award, widely known as computing’s “Nobel Prize”, were announced. Yoshua Bengio, Geoffrey Hinton and Yann LeCun were recognized for their pioneering work over a 30-year period on AI. In the 80s, when AI research was very unfashionable, the trio were motivated by the idea of creating algorithms that could operate, at least at the level of metaphor, like a brain. Neurons in the brain work as an ensemble, together learning internal representations of the input they observe, and these researchers believed their artificial neural networks could do the same. More than 30 years later their combined research has led to revolutionary advancements in computer vision, natural language processing and robotics, forming an impressive array of new technologies, many of which we use at <a href="https://edited.com/">EDITED</a>, a retail decision platform.</p><h3>AI and industry</h3><p>Significantly, this work was created with industry, and not solely academic, applications in mind. Hinton is currently a VP and Engineering Fellow at Google, LeCun is a VP and Chief AI Scientist at Facebook. Indeed, LeCun’s breakthrough work trained a CNN on images of handwritten digits to read bank cheques — it’s estimated that his system read over 10% of all the cheques in the US in the late 1990s and early 2000s. AI has a rich heritage in industry with innumerable practical applications. Although it is hard to ignore its more eye-catching use cases — Amazon’s warehouse robotics, or Tesla’s self-driving cars — arguably AI’s most impactful success stories have been in consumer technologies. AI powers translation engines, recommender systems, image classification and forecasting — all of which we employ at EDITED — and critically, supports their potential to scale. Researcher Andrew Ng has a maxim, “if a typical person can do a mental task with less than one second of thought, we can probably automate it with AI either now or in the near future”. This provides a useful way to identify AI applications in a domain.</p><h3>AI-driven means data-driven</h3><p>A combination of cheaper computation and masses of data has been integral to advances in AI, and it will lead to the success of companies introducing AI now. By 2020 <a href="https://www.domo.com/learn/data-never-sleeps-6"><em>Data Never Sleeps</em></a> — the yearly infographic from the cloud-based platform provider Domo — estimates that for every person on Earth 1.7 MB of data will be created every second. Harnessing data is crucial to the success of AI, and luckily it’s created in abundance. The majority of this is harvested by large corporations and it can be hard to tap into. So it’s important to first scope what data you need, whether this is customer-focused and/or product-related, including considerations around the ethics behind it. At EDITED we make use of both as product data drives our company and user data gives us insights on our customers. We use crawlers and AI to harvest commercial data provided by the online apparel industry and convert this into something tangible for the user. This creates the single biggest source of real-time retail data in the world. We are then able to understand the success of our features through user interaction and make our own data-driven decisions on new features.</p><h3>Scaling and upskilling</h3><p>AI-powered technology is most useful when it is leveraged to deliver actionable insights at scale, but developing sustainable and scalable infrastructure can be challenging. At EDITED we strive to exponentially increase our crawling and to add more intelligence to each product, whilst maintaining its accuracy. With the rise of machine learning and AI-focused degrees across institutions worldwide, a growing talent pool has emerged. Combine this with extensive online and course-based resources to upskill your current workforce, implementing tailor-made AI solutions is now achievable for many organizations.</p><p>For the more AI-averse, several machine learning platforms exist requiring little domain knowledge to execute. However, the best data-driven decisions come from a workforce of data scientists and engineers that understand the technology and how to drive improvement with it. Still better is a combined effort with data science, product management and leadership to assess opportunities where AI can improve the value chain, as well as the appropriate timescale for implementation.</p><h3>Be brave</h3><p>For AI to succeed, an experimental culture needs to become the standard, with a consistent emphasis on upskilling of the workforce. When Bengio, Hinton and LeCun began their pivotal work on AI, the culture was stagnant and investment low. Their research took decades to see fruition. It is important we realize the potential of AI, and the need to repeatedly invest and learn. At EDITED we are continually introducing new technologies and competencies into our data science team. We encourage others to work on new AI applications and promote the sharing of knowledge, while avoiding a reduction in investment by managing expectations amongst users and the company of how AI can be leveraged. At EDITED our foundational value is “Be Brave”, which has been key to our AI success. AI has proved its worth within industry, so if you are not making the switch, you may find your competitor already has.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8816453c786" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/how-to-become-an-ai-driven-company-e8816453c786">How to become an AI-driven company</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>