<?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 Alex Teodor on Medium]]></title>
        <description><![CDATA[Stories by Alex Teodor on Medium]]></description>
        <link>https://medium.com/@alexteodor?source=rss-84b01c952586------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*nO-YH19R29z-nnB2_rhltA.jpeg</url>
            <title>Stories by Alex Teodor on Medium</title>
            <link>https://medium.com/@alexteodor?source=rss-84b01c952586------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:31:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexteodor/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[Towards Reusable Entity Resolution]]></title>
            <link>https://medium.com/peak-product/towards-reusable-entity-resolution-eed1c6ee4a14?source=rss-84b01c952586------2</link>
            <guid isPermaLink="false">https://medium.com/p/eed1c6ee4a14</guid>
            <category><![CDATA[entity-resolution]]></category>
            <category><![CDATA[siamese-networks]]></category>
            <category><![CDATA[variational-autoencoder]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Alex Teodor]]></dc:creator>
            <pubDate>Thu, 29 Apr 2021 11:56:57 GMT</pubDate>
            <atom:updated>2021-04-29T12:23:55.568Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SB-INOvsgPtgsV2SHr7-ig.jpeg" /><figcaption>Image source: <a href="https://www.verdict.co.uk/reusability-consumer-technology/">https://www.verdict.co.uk/reusability-consumer-technology</a></figcaption></figure><p>Success in the world of data science often comes not only from mastering the science behind machine learning algorithms but also from mastering the reusability of the resulting solutions, so that they can be quickly and easily adapted to similar data-related problems. At <a href="https://peak.ai/">Peak</a>, we frequently face building new or adapting existing AI solutions for various such problems. One recent example is <a href="https://en.wikipedia.org/wiki/Record_linkage"><em>entity resolution</em></a><em> (ER).</em> This blog is about our research journey towards a reusable entity resolution solution and about how we’ve managed to use deep learning to help us perform entity resolution efficiently and, crucially, with high reusability potential. First, though, let’s define entity resolution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/684/1*yrR0TNCud26fO-Lj_c81uQ.png" /></figure><blockquote>Entity resolution (ER), a.k.a. de-duplication, record linkage, or entity reconciliation, is the task of identifying all the representations that refer to the same real-world entity from a given (set of) dataset(s), where an entity is an object, individual, unit, place, or any other item that should be unique within a domain.</blockquote><p>To make things a little more concrete, let’s take an example of a case where we want to find all records from a table that refer to the same company across lists of companies taken from two different sources. Figure 1 shows an example of how a pair of such duplicates might look in practice.</p><figure><img alt="Example of duplicate table rows." src="https://cdn-images-1.medium.com/max/1024/1*Cn4F0h2ZaoHVrxVIpHxXaA.png" /><figcaption><strong>Figure 1: </strong>Example of duplicate table rows</figcaption></figure><p>The aim, then, is to create an association between elements in each row that refer to the same entity and, collectively, denote the same company details. You might be able to see that this can get complicated very quickly: subtle differences in how entities are represented can wreak havoc on naive approaches to matching such as string similarities.</p><p>Furthermore, the process of performing ER is often highly data and use case-specific. In other words, even for the most experienced data scientists, entity resolution can be an expensive problem to address, both in terms of time and domain knowledge. The cost in time is typically determined by the need for frequent user intervention to build new or adapt existing ER tools to the case at hand. The cost in knowledge is determined by the need for familiarity with the data and the data domain. There is no single recipe for a successful ER process. Some examples of use cases that often require user guidance and data familiarity include:</p><ul><li>Matching different representations of the same entity to create complete and unique representations before analyzing the data</li><li>Cleaning up the duplicates in training data before using machine learning and predictive analytics</li><li>Filtering entities from one source based on another source, e.g. compare customer lists against prospect lists to avoid marketing to people who are already customers</li><li>Matching entities against reference types to auto–label training data</li><li>Matching incomplete entity representations against reference entities to fill in missing values</li></ul><p>With several such real-life ER use cases under our belt, we quickly realized that the current offering of ER tools and methods comes short in the area of <em>reusability</em>. For instance, traditionally, ER is approached as a similarity problem, where entity representation, such as table rows, are compared using several similarity measurements. Examples here include Jaccard similarity for strings or Cosine similarity for numbers. But such measurements may not properly catch the notion of similarity in every data scenario or use case.</p><p>More recent ER solutions employ the use of machine and/or deep learning to automatically learn such similarity functions, tailored for the current use case and relying on enough training data in the form of positive and negative examples. <a href="https://github.com/dedupeio/dedupe"><em>DEDUPE</em></a><em> </em>and <a href="https://github.com/anhaidgroup/py_entitymatching"><em>py-entitymatching</em></a><em> </em>are good examples of such solutions. More complex variants often employ deep learning to reason about entity similarities. Examples here include <a href="https://github.com/megagonlabs/ditto"><em>DITTO</em></a><em> </em>and <a href="https://github.com/anhaidgroup/deepmatcher"><em>DeepMatcher</em></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dblW3YQrOo6hHrKP" /><figcaption><strong>Figure 2: </strong>ML-based ER architecture</figcaption></figure><p>Such solutions usually follow a common architecture, illustrated in Figure 2. Behind the scenes, the rows of one or more tables are initially processed to extract specific representations (usually vectorised representations). This step is commonly called <em>Feature Engineering.</em> Then a <em>Blocking</em> process removes obvious non-duplicates and forwards groups of duplicate candidates to a <em>Matching </em>process where a supervised classification algorithm performs pairwise comparisons within each group. The <em>Blocking </em>step is optional and aims to ease the comparison effort of <em>Matching</em> by quickly filtering out obvious non-duplicates.</p><p>These types of learning-based ER tools score more points in the area of reusability than non-learning ones. However, the process is still not pain-free. In practice, adapting an ER solution that follows the architecture from Figure 2 to a new use-case often translates to:</p><ol><li>Ensuring that the types of features that <em>Feature engineering</em> extracts are suitable for the comparison of two entities in the context of the new use-case.</li><li>Identifying local examples of duplicates and non–duplicates that reflect the new intended notion of similarity and can be used to train the <em>Matching </em>algorithm.</li><li>Training the <em>Matching </em>algorithm to learn a similarity function that discriminates between duplicates and non-duplicates.</li></ol><p>In practice, we observed that (1) tends to be time-consuming and that it is tightly coupled with (3), with both relying on enough training data from (2). As a consequence, if either of (1) and (3) is unsuitable for the current ER task the tool used has to be adapted or changed. In other words, the ER task becomes even more time — and knowledge — costly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/732/1*IEJQ-ANIAimxGf3g9LiQFQ.png" /></figure><p>This motivated us to <em>decouple Feature engineering </em>from <em>Matching</em> and to devise a feature engineering method that is independent of training data and, crucially, of the matching algorithm. As a bonus, this decoupling allowed us to introduce a notion of <em>transfer learning </em>to the problem of ER. This is because the independent representations proved to be transferable to other ER tasks performed on data from different domains.</p><p>As you probably suspect, the devil is in the details. So, let’s dive into some of these!</p><h4><strong>Representation learning for entities</strong></h4><p>We want to give our entity representations freedom from the matching algorithm and its training process. If we consider each entity to be a single table row, one way of obtaining independent representations is to follow a process borrowed from the <em>representation learning </em>research area, as illustrated in Figure 3.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F2n6ix9bhbm6Z6c1" /><figcaption><strong>Figure 3: </strong>Autoencoder-based entity representation model</figcaption></figure><p>Each row is firstly mapped to some vectorized representation. This can be based on word-embeddings or simple TF/IDF scores of the constituent words. However, such representations are deterministic, fixed vectors in a potentially irregular multidimensional space with reduced control over how close duplicates end up. In other words: we can do better. To make the representations more specific to the input data, we can use the representation learning paradigm illustrated above: an encoder-decoder mechanism (or autoencoder) where the encoder is a neural network that squeezes the input to a lower-dimensional space, while the decoder is a neural network that aims to reconstruct the original input given the reduced, often called <em>latent</em>, output of the encoder. When trained in tandem, these two networks can produce high-quality feature vectors that retain much of the richness of the original inputs.</p><p>Following a similar approach, one can transfer all the entity representation reasoning into a probabilistic space by using variational autoencoders (VAEs) instead of plain ‘vanilla’ autoencoders (AEs). In such a case, for each row, instead of a simple latent representation, the encoder generates a multidimensional probability distribution. Then, the decoder inputs are generated at random from anywhere inside this distribution, exposing the decoder to a range of variations of the encoding of the same input. The decoder, therefore, learns that not only is a single point in the latent space referring to a row, but all nearby points refer to that row as well — accounting for variation and uncertainty across the attribute values of duplicates.</p><blockquote>If you haven’t come across them before, VAEs can be a little intimidating. If you are interested in learning more, <a href="https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73">this post</a> does a great job of giving an overview of the basics. For their use in entity resolution, we leave the investigation as homework for the particularly keen readers and recommend as reference our recent <a href="https://arxiv.org/pdf/2011.10406.pdf">research paper</a> from IEEE’s ICDE 2021 on how to put all these representation learning ideas into a practical ER solution.</blockquote><p>A nice feature of the VAE-based method above is its ability to generate our representations for table rows in an <em>unsupervised</em> fashion. As you’re probably aware, vectorized representations such as these are compatible with most modern machine learning algorithms you might choose to use and this comes in very handy when we turn our attention to the <em>Matching </em>phase of ER. Secondly, such latent representations are more expressive than the initial auto-encoder inputs, meaning that they can encode additional useful information that can in turn be exploited in subsequent ML steps. In particular, this can better enable us to determine the similarity between entities represented in this way. Thirdly, in cases where the deduplication task is simple enough not to require supervised matching, we can use clustering or other unsupervised learning approaches on our latent representations to determine the potential duplicates, too.</p><h4>Supervised matching</h4><p>So far, we have discussed a representation learning method for table rows. Since the representations are generated in a manner that allows us to measure the similarity between duplicates, you may be wondering what is the purpose of a subsequent matching step. To answer this, we refer back to Figure 3. The illustrated input to the whole mechanism for generating representations is a collection of table rows. In the figure, we illustrate two entities denoting the same song by the same artist released as part of two different albums. Whether or not the two rows are duplicates depends on the use-case and the unsupervised representations may not reflect that decision. There is, therefore, a need for aligning the notion of similarity conveyed by the row representations with the use-case intent. Often, this is done using supervised machine learning and training data. That’s where the <em>Matching </em>step comes in.</p><p>First, visualize that a training instance in our context is a pair of rows with an associated duplicate/non-duplicate label. With that in mind, consider that the label is assigned with a certain probability, as in any binary classification scenario. In other words, this probability can be seen as a similarity score between the two rows and the label assignment can be done by a similarity function which the matching model has to learn.</p><blockquote>One famous proposal for learning similarity functions between two input entities is through the use of <a href="https://en.wikipedia.org/wiki/Siamese_neural_network"><em>siamese neural networks</em></a><em>.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_QSLpi7P5PRvVCJP" /><figcaption><strong>Figure 4: </strong>Siamese entity matching network</figcaption></figure><p>Now consider Figure 4 where we illustrate such a siamese neural network model with each head being represented by an <em>already trained representation encoder. </em>You’ve met them in Figure 3. The latent representations resulting from these siamese encoders are combined into a single input that is fed into a matching neural network layer responsible for assigning a duplicate/non-duplicate label.</p><p>Broadly, the potential loss in feature expressiveness determined by the unsupervised nature of the representation model is attenuated by the siamese matching model from the figure above. Its purpose is to adjust the parameters of the representation encoders and of the matching layer in light of training data — so this matching process is <em>supervised</em>. This adjustment is also efficient since most of the encoding optimization work has already been done by the unsupervised training process.</p><h4><strong>Enabling transfer learning in entity resolution</strong></h4><p>Let’s go back to our <em>decoupling </em>idea from before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*3juS-T3o-gGxKNGBjCL2iA.png" /></figure><p>The most important consequence of decoupling entity representation learning from similarity learning when performing entity resolution is that the representation model trained during one entity resolution task can be reused in other entity resolution tasks, in principle <em>regardless of the data domain</em>. In other words:</p><blockquote>Learning entity representations in an unsupervised fashion, independent from the Matching training data, allows for a form of transfer learning when addressing future ER problems. In other words, it boosts reusability.</blockquote><p>Concretely, because the VAE model operates on vectorised representations of rows, its outputs are independent of domain-specific aspects, such as the domain vocabularies (i.e. domain-specific jargon). Consequently, once the parameters of the autoencoder are trained, any new initial representation can be accurately encoded, regardless of the domain of the original data. This is significant as it unlocks the idea of transfer learning and boosts the reusability of the representation model across different data domains.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aQPMnH1_x2CZDJVl" /><figcaption><strong>Figure 5: </strong>Transfer learning experiment results for cross-domain ER</figcaption></figure><p>We discuss all these notions in more detail in our recent ICDE 2021 <a href="https://arxiv.org/pdf/2011.10406.pdf">paper</a>. In Figure 5 we use one of the experiments from the paper that attests to the reusability potential of our entity representation model.</p><p>For this experiment, we considered eight data domains with information about restaurant locations (<em>Rest.)</em>, research paper citations (<em>Cit. 1)</em>, cosmetic products (<em>Cosm.</em>), software products (<em>Soft.</em>), songs (<em>Music</em>), beer products (<em>Beer</em>), company available stocks of goods (<em>Stocks</em>), and client contact details <em>(CRM</em>). Most of the datasets have been borrowed from <a href="https://github.com/anhaidgroup/deepmatcher/blob/master/Datasets.md">here</a>. With this data in hand, we first train a VAE-based representation model on each of the eight data domains. Let’s call these the <em>local representation models (LRMs). </em>Similarly, we trained another VAE-based representation model on a ninth data domain, different from the ones above. Let’s call this one the <em>transferred representation model (TRM)</em>. We then use the representations given by each of the LRMs and perform supervised matching on their respective domains using some training and test data available at the same link above. We measure the F1 score of the matching task (in blue in Figure 5). Similarly, as you probably guessed, we use the representations given by the TRM and re-run matching for each of the domains. For this last part, we simply used the same TRM in each case as an already trained representation model. The corresponding F1 scores are in red in Figure 5. The conclusion is somewhat unexpected and yet quite valuable:</p><blockquote>The TRM, although originating from a different data domain, proved capable of replacing each of the LRMs with minimal impact on the matching results.</blockquote><h4>Concluding remarks</h4><p>We started this post commenting on the lack of reusability of most entity resolution solutions out there and about the challenges faced in attempting to decouple feature engineering from matching when using such solutions. This post has hopefully given you an idea about the benefits of decoupling the two steps in practice and you may now have the basics for achieving some of these benefits yourself.</p><p>If you’re interested in finding out more about our approach we’d encourage you to check out our ICDE 2021 paper on the topic (pre-print version linked below). If you have any questions or comments, feel free to reach out — we’re always happy to chat!</p><h4>Before wrapping up</h4><p>Thanks to <a href="https://medium.com/u/fba1125456c6">Mark Douthwaite</a> for reviewing and editing this story.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/176/1*ka63auhJdWDtYP4ALAjsPg.png" /><figcaption><a href="https://arxiv.org/pdf/2011.10406.pdf">https://arxiv.org/pdf/2011.10406.pdf</a></figcaption></figure><p>Check out our <a href="https://arxiv.org/pdf/2011.10406.pdf">recent paper</a> from ICDE 2021 with an in-depth technical analysis of the ideas discussed in this post.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/280/1*ZlwWlTLCz-6rOnUN9XrzpQ.png" /></figure><p>We’re looking for brilliant minds to join us on our AI journey and help us reach new heights. Here’s a <a href="https://lnkd.in/dmpa9KH">selection of roles</a> we’re currently looking to fill at Peak.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eed1c6ee4a14" width="1" height="1" alt=""><hr><p><a href="https://medium.com/peak-product/towards-reusable-entity-resolution-eed1c6ee4a14">Towards Reusable Entity Resolution</a> was originally published in <a href="https://medium.com/peak-product">Peak</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Format Transformation — synthesizing from custom examples]]></title>
            <link>https://medium.com/the-data-value-factory/format-transformation-synthesizing-from-custom-examples-22bff0b1fbd7?source=rss-84b01c952586------2</link>
            <guid isPermaLink="false">https://medium.com/p/22bff0b1fbd7</guid>
            <category><![CDATA[data-wrangling]]></category>
            <category><![CDATA[format-transformation]]></category>
            <category><![CDATA[data-preparation]]></category>
            <dc:creator><![CDATA[Alex Teodor]]></dc:creator>
            <pubDate>Mon, 27 Apr 2020 08:01:02 GMT</pubDate>
            <atom:updated>2020-04-27T08:01:02.591Z</atom:updated>
            <content:encoded><![CDATA[<h3>Format Transformation — synthesizing from custom examples</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*DbXsRxEZSm838khH-2KgQw.jpeg" /><figcaption>Source: <a href="https://pixabay.com/photos/swallowtail-butterfly-cocoons-larva-329376/">https://pixabay.com/photos/swallowtail-butterfly-cocoons-larva-329376/</a></figcaption></figure><p>Data analysis is hard! Practitioners have testified time and again that extracting the much-promised value out of data through analysis is no free lunch. There is a famous 80/20 rule of working with data that describes the state of affairs in the real-world quite well:</p><blockquote>data scientists tend to spend up to 80% of their time preparing the data and only 20% analysing it.</blockquote><p>Although preparation here includes everything from getting familiar with the data to cleaning, standardising and structuring unruly data to match the needs of an analysis or visualisation package, in this blog we focus on format transformations.</p><p>We have already covered <a href="https://thedatavaluefactory.com/format-transformation-in-data-preparation/">how format transformation can be performed in practice</a>, and one of the challenges that emerged from the past discussion was that transformation often requires at least some programming knowledge. In this blog, we go a step further and explore two potential approaches to save the user from the need to write code.</p><h4>FlashFill and transformation learning</h4><p>Let’s start simple and note that the need for format transformation is older than the need for data wrangling. For instance, think Microsoft Excel, the swiss-army-knife of data standardisation and structuring, and consider the following example where the task is to extract the country name from the <em>Participants</em> column.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/301/1*za1IaE3sHDqZMM66ql6TOQ.png" /><figcaption>Format transformation example in Excel 2013 using FlashFill</figcaption></figure><p>Surely, Excel users are not always experienced programmers and transforming the format representation of columns does not always involve writing complex VBA programs. Indeed, it doesn’t, or at least since Excel 2013. This is thanks to a feature called FlashFill: a special tool that analyzes the information users enter and automatically fills data when it identifies a pattern.</p><p>Performing the task from our example above in Excel using FlashFill is straightforward, as the tool can <em>automatically</em> <em>learn</em> the pattern transformation from the given examples and fill in the rest of the column as suggested in the picture above. These can be considered custom examples, as they are provided by the user.</p><h4>Wrangler and predictive interaction</h4><p>Getting back to the data analysis world, performing format transformations using some Excel features, be it a VBA script, is infeasible since most of the data preparation tasks require the power and expressiveness of some programming libraries, e.g., in Python. We have seen <a href="https://thedatavaluefactory.com/format-transformation-writing-rules/">before</a> how Trifacta’s Wrangler aims to blend the ease of use of tools such as Excel and the power of programming languages such as Python, albeit still requiring some written code, at least until its predictive feature learns enough transformations. This predictive feature, much like FlashFill in Excel, intelligently assesses the data and synthesizes a transformation. As opposed to FlashFill though, this transformation stems from a combination of past user decisions and data formatting. Nevertheless, both approaches limit the need for written code for performing format transformations.</p><p>Here is an example of extracting the city name from a city-country pair in Trifacta’s Wrangler. For this relatively simple example, the tool is able to synthesize the required transformation automatically. For more complex pattern transformations, the tool needs experience, in the form of previously used similar transformations, to be able to suggest a relevant one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/479/1*d0oWY9f4XosgUnAsB7qikw.png" /><figcaption>Format transformation example in Wrangler using predictive interaction</figcaption></figure><h4>Conclusions</h4><p>So there is life for format transformation tasks outside the, often inaccessible, world of programming. We have seen two examples of performing transformations only by providing custom examples. But looking at the need for such transformations from the perspective of data wrangling reveals one important shortcoming: the user has to be familiar with the data and able to identify different formatting patterns for which to provide examples: at least one, often more, examples per pattern to transform. This makes such tools suitable for spreadsheet-like scenarios where data tends to be of manageable sizes and the user can relatively quickly explore the data. For larger datasets, where there are more than a few transformation cases and spotting all these cases is hard, tools such as the ones discussed above come up short. A potential solution in such cases is <a href="https://thedatavaluefactory.com/approach/">Data Preparer</a>, a data preparation tool that adopts a broader view on data wrangling and performs format transformation without assuming programming knowledge or high levels of data familiarity from the user. The next blog of this series will introduce Data Preparer and exemplify how format transformation can leverage existing data to discover the much needed examples.</p><blockquote>This post was originally posted on <a href="https://thedatavaluefactory.com/format-transformation-synthesizing-from-custom-examples/">The Data Value Factory blog</a>.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22bff0b1fbd7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-data-value-factory/format-transformation-synthesizing-from-custom-examples-22bff0b1fbd7">Format Transformation — synthesizing from custom examples</a> was originally published in <a href="https://medium.com/the-data-value-factory">The Data Value Factory</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The one with the FRIENDS language models]]></title>
            <link>https://alexteodor.medium.com/the-one-with-the-friends-language-models-a91a7cc273c5?source=rss-84b01c952586------2</link>
            <guid isPermaLink="false">https://medium.com/p/a91a7cc273c5</guid>
            <dc:creator><![CDATA[Alex Teodor]]></dc:creator>
            <pubDate>Thu, 23 Apr 2020 18:08:21 GMT</pubDate>
            <atom:updated>2021-04-27T09:47:59.303Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Udacity Data Science Nanodegree project</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DYW0ghKxeOBYM7hTRdretw.jpeg" /></figure><p>There have been many analysis reports and blogs written on hit sitcom shows such as FRIENDS. Most of them have been analytical, such as <a href="https://yashuseth.blog/2017/12/29/data-analysis-lead-character-of-friends-data-science/">this one</a> that aims to identify the lead character based on the episode scripts from <a href="https://fangj.github.io/friends/">this</a> amazing fan website. But such data allows us to do much more than analysing line/word counts and screen appearances. It gives us a window into the typologies embodied by each character.</p><p>In the field of social sciences, it is believed that every time we open our mouths, we unwittingly reveal our personalities. Following a similar, albeit much simpler, hypothesis, in this blog, I set out to synthesize the language of each FRIENDS character by using NLP deep learning models. This could be used, for instance, given a character’s line, to predict that character. Although character prediction is discussed in this project, more complex applications could easily be accommodated, such as generating text that expresses <em>Chandlerian sarcasm </em>or <em>Phoebeian quirkiness.</em></p><h4>Project overview</h4><p>So the task discussed here is to predict the speaker from a given FRIENDS quote using their language. The solution would have to be able to accurately guess the speaker only from the text quote, without any other information provided. To this end, some latent features present in each character’s lines would have to be extracted.</p><p>Concretely, I start from <a href="https://fangj.github.io/friends/">this</a> data, conveniently curated and merged into a single CSV file <a href="https://github.com/shilpibhattacharyya/Friends_Analysis/blob/master/friends_dataset.csv">here</a>. I then use part of the lines available in the data to create an LSTM-based language model for each character. With these models in hand, I then move to train another model that is able to accurately predict the friend who said a given line. This quote prediction model can potentially be created without the language models, directly from the given lines. In practice though, such an approach resulted in poor accuracy, e.g., 20–25 % accuracy with multiple types of features tested.</p><h4>Data exploration and cleaning</h4><p>A few stats about the data I think are useful. Firstly, data seems to be clean enough, with few missing values which I will remove later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/853/1*mrTkbyP6kq7tuBKl-u4_cA.png" /></figure><p>The distribution of line counts attributed to each character suggests a relatively balanced contribution of each friend, with Ross and Rachel leading in the dialogue contributions, while Phoebe contributing the least.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/410/1*72KvwnXMdNpaDyN3PV3Rcg.png" /><figcaption>Total line count/character</figcaption></figure><p>Drilling down per season shows a similar trend.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/469/1*MwrbrgCSLFaNvk1QhATVVg.png" /><figcaption>Line count/character/season</figcaption></figure><p>Data presents itself quite clean so not much pre-processing was needed, apart from the usual tokenization and lemmatization. I chose to keep the stop words as their over/underuse can sometimes reflect an important characteristic of the language.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e7c3c8c82debcedf0d9a392e5c08a74e/href">https://medium.com/media/e7c3c8c82debcedf0d9a392e5c08a74e/href</a></iframe><h4>Language models</h4><p>The intuition behind the language models is based on the hypothesis that</p><blockquote>the choice of words and their context is specific to each character.</blockquote><p>This is akin to the hypothesis behind word embeddings. So the aim of the per-friend language models is to generate specific word representations for each character. To this end, I used <a href="http://keras.io/">Keras</a> to define an LSTM-based architecture that looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/284/1*DiX8R02MiyT1J2bDoMCCkg.png" /><figcaption>The neural network architecture used</figcaption></figure><p>The initial version of this architecture included 2 LSTM layers with one Dropout between them, but this resulted in a significant increase in training time without much return in accuracy.</p><p>Another important characteristic here is that the Embedding layer follows a CBOW architecture where the context of a word is given by its surrounding words. Here is the code for generating CBOW-like sequences.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4c1f498f38a55ab5782e701e276e9cb1/href">https://medium.com/media/4c1f498f38a55ab5782e701e276e9cb1/href</a></iframe><p>Furthermore, the Embedding layer builds on top of pre-learned, general-purpose embeddings from <a href="https://fasttext.cc/docs/en/english-vectors.html">FastText</a>. These pre-trained embeddings are modified during the learning process to reflect the specific use of words per character. The intuition is that starting from general-purpose embeddings and tailoring them for the case at hand will give a more accurate representation of words and their context. This proved true, since trying to predict quotes (as described later) by using only pre-trained embeddings from FastText resulted in 20% less accuracy. Similarly, learning FRIENDS-specific embeddings without the transfer learning from FastText resulted in 34% less accuracy. This is not surprising since learning embedding requires large training corpora.</p><p><em>More details about the values chosen for hyperparameters can be found in the code linked at the end of the article.</em></p><p>Using the above LSTM-based architecture and data from Seasons 1–8, I trained a sequential model for each FRIENDS character. I then used these models to generate embeddings to be used as features in the classification process described next.</p><h4>Quote prediction</h4><p>Given a set of quotes-character associations, learning a model that can then predict the speaker of new quotes is straightforward with the right feature extraction method. I started simple with TF/IDF weights and general-purpose word embeddings but none of these came close to the results obtained with the FRIENDS-embeddings reported below. So given a quote, a feature vector is obtained by extracting the average embeddings from each of the six models created above and concatenating them. This results in 1800 features in each vector (the FastText models return 300-feature embeddings). I used lines from Seasons 9 and 10 to generate training and test data used to train and test a Random Forest classifier.</p><p>Below is the code for the custom transformer that performs the above-mentioned transformation for each new quote.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/bee8a8576cca36f780f96a7d9e4b9dec/href">https://medium.com/media/bee8a8576cca36f780f96a7d9e4b9dec/href</a></iframe><h4>Evaluation</h4><p>Given that, in the end, this is a classification project, precision, recall and F-score have been the chosen metrics in evaluating the performance of the solution. Below is the classification report for the quote prediction part.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/536/1*0tTQHn2fw5DcKbcC_E7izw.png" /><figcaption>The classification report for quote prediction</figcaption></figure><p>There is also the confusion-matrix that shows correct discrimination between characters most of the time, with only a few percents of the quotes misclassified for each character.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/582/1*UnCAZihug7QpKmZyZb1JlQ.png" /><figcaption>The confusion matrix for quote prediction</figcaption></figure><p>It is interesting to note that, although Ross has the most number of lines across all seasons, its F-score is the lowest. This could be because there is more overlap between the choice of Ross’s words and the choice of others or because the apparent small imbalance in available lines per character has a bigger impact than initially thought. This part deserves further investigations.</p><h4>Conclusions and possible improvements</h4><p>Using embeddings specific to each FRIENDS character was not the obvious winner in this task of <em>guess the FRIEND </em>but, in the end, it returned the best results. Such an approach can be extended and used in more complex tasks. For example, Chandler has become the image of funny sarcasm. Similarly to Chandler’s language model, a more general sarcasm language model can be created from sarcastic sentences that can then be used in sentiment analysis tasks.</p><p>Lastly, there are quite a few improvements that could be added to the current implementation. For example, experimenting with other neural network architectures for creating language models. Or combining the embeddings in less naive ways than averaging them when generating features for the quotes. Feel free to try your own ideas based on the <em>guess the FRIEND </em>full code available <a href="https://github.com/alex-bogatu/guess_the_FRIEND">here</a>.</p><p>So guess the friend who said this:</p><pre>Don&#39;t you put words in people&#39;s mouths. You put turkey in people&#39;s mouths.</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*MapAU6YaQB8A1L5yByjsUg.gif" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a91a7cc273c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Objectively choosing where to live]]></title>
            <link>https://alexteodor.medium.com/objectively-choosing-where-to-live-2c505885fd6b?source=rss-84b01c952586------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c505885fd6b</guid>
            <dc:creator><![CDATA[Alex Teodor]]></dc:creator>
            <pubDate>Sat, 04 Apr 2020 14:34:42 GMT</pubDate>
            <atom:updated>2020-04-04T14:34:42.777Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aprCEAj2M9enj9XR3iSIog.jpeg" /><figcaption>Source: <a href="https://www.expatica.com/de/moving/location/best-cities-in-germany-107869/">https://www.expatica.com/de/moving/location/best-cities-in-germany-107869/</a></figcaption></figure><h3>Introduction</h3><p>Socio-economic globalization has offered a privileged category of people the luxury of choosing the country they would like to live in. But when it comes to making such a decision the choice becomes harder than it looks. Whether it is family roots, happiness levels, overall financial stability, climate, or something completely unrelated to any of these suggestions, the definition of the number one country to live in is strongly subjective. Nevertheless, I believe there is an objective component to this question that, with the right information, can be sorted out. This is what I set out to do in this post: using data from reliable sources, understand what could that data objectively tell me about the countries I would consider living in.</p><h3>The data</h3><p>The keywords here are <em>the right information from reliable sources</em>. Like many <a href="http://hdr.undp.org/en/content/human-development-index-hdi">other surveys</a> conducted regarding the overall consensus as to which countries are superior in terms of livability, in this post I rely on data from the <a href="https://datacatalog.worldbank.org/dataset/world-development-indicators">World Bank</a> and <a href="https://www.numbeo.com/">Numbeo</a>.</p><p>The World Bank’s development indicators (WDI) is a collection of statistics compiled from officially-recognized international sources. It presents the most current and accurate global development data available and includes national, regional and global estimates.</p><p>Numbeo is the world’s largest database of user-contributed data about cities and countries worldwide. It provides current and timely information on world living conditions including cost of living, housing indicators, health care, traffic, crime and pollution.</p><p>Now that you are acquainted with the data let’s dive right in and see, for 13 countries (my personal choices) some economical-, health care- and general wellbeing-related statistics from the last 8 years that may make some places more livable than others.</p><h3>Economy</h3><p>One factor that I consider important when choosing the country to live in is the economical one. Of course, many criteria can be considered here to analyse a country’s economy. In fact, WDI data offers a plethora of economical indicators that can be used for economic analysis. Two such indicators got my attention. Firstly, the <strong>GDP per capita</strong> as a measure of the overall wealth of the country. Then, compare this against the <strong>percentage share of income or consumption that accrues to subgroups of population</strong> indicated by deciles or quintiles. Since I consider myself being part of the middle class, I will use here the indicator for the share of income held by the <strong>third 20%</strong> of the population. This last indicator can be seen as a measure of inequality or inconsistency between the apparent wealth of a country and the wealth of its middle-class inhabitants.</p><p>Each point on the graph below represents a country and its size is given by the population size of that country. This plot style and similar ones from the rest of this post are inspired by the awesome <a href="https://www.gapminder.org/">Gapminder</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6gBqVvqX1uSHndd-19PHaQ.png" /></figure><p>As expected, there seems to be a tendency of uniformity amongst Western European countries (centre of the graph). USA stands out as having a high GDP per capita but a higher level of inequality than the others as less than 15.5 % of the wealth is held by the middle class.<br> Romania (an eastern European country) scores well at inequality but the average GDP over the last 8 years is not great.<br> The winners for me here (apart from Norway) are any countries from the central cluster: Belgium, Austria, Netherlands, etc.</p><p>Keeping the ball in the economy court, I wonder if the <em>cost of living</em> and/or <em>purchasing power</em> have anything to say in this analysis. Here it is with data from Numbeo. Details about the indices can be found <a href="https://www.numbeo.com/cost-of-living/cpi_explained.jsp">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d757FNzMm16991KizZFPiA.png" /></figure><p>Ideally, here I would pick a country in the bottom-right corner, such as Germany, where the two measures are relatively balanced or there is a purchasing power advantage.</p><h3>Healthcare</h3><p>The economical analysis can continue with many other criteria but the decision we want to make here is not an economical one. For instance, the “mighty” GDP does not count the health of country inhabitants, or their life expectancy. But we may be able to use GDP-related information if we look at the proportion of GDP being dedicated to the health system and the life expectancy in each country. Surely, the connection between the two is much deeper and more complex to analyse than the next graph, but this could be a good start.<br> As before, each point on the graph represents a country but this time its size is given by the GDP per capita of that country.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RarYUaidCQotPN4PdCUprQ.png" /></figure><p>The Western European uniformity is present in this analysis as well, with most of the countries in that part of Europe investing around 10% of their, often large, GDP per capita in health. Romania, on the other hand, has still some investments to make in the health system. USA clearly treats its health system seriously but there are likely other reasons for the lower life expectancy figures.</p><p>On the same note of health system analysis, <a href="https://www.numbeo.com/">Numbeo</a> offers more general indices. For example, a <strong>health care index</strong> as an estimation of the overall quality of the health care system, health care professionals, equipment, staff, doctors, cost, etc., and a <strong>pollution index</strong>, including air, water, etc. More details about the indices can be found <a href="https://www.numbeo.com/health-care/indices_explained.jsp">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zc_NSyUzJ23R6pazymuD_g.png" /></figure><p>As before, there seems to be a “good-enough” cluster amongst Western European countries and two ideal choices: Norway and Switzerland. The surprises appear to be Belgium and Romania with an unexpectedly high average pollution over the measured period.</p><h3>General wellbeing</h3><p>Lastly, I wondered what does a more general analysis of quality of life and safety would tell us about the countries under analysis. <em>Quality of Life Index</em> (higher is better) is an estimation of overall quality of life by using an empirical formula which takes into account purchasing power index (higher is better), pollution index (lower is better), house price to income ratio (lower is better), cost of living index (lower is better), etc. <em>Safety index </em>(higher is better) is part of that formula and represents the opposite of the overall level of crime in a given country. More information can be found <a href="https://www.numbeo.com/quality-of-life/indices_explained.jsp">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RdvK7KUwQR2fn-9-PdGpnA.png" /></figure><p>Romania Switzerland, Denmark, and Austria stand out as quite safe countries. Once again, apart from idealistic Switzerland, a handful of Western European countries such as Germany, Austria, or Nederlands seem like a good choice when considering the overall quality of life.</p><h3>Conclusion</h3><p>I started from the belief that choosing a country to live in is a highly subjective process. But there is an objective component to it that can be analysed in terms of various development and quality of life indicators. This is what I set out to do in this post and here are the overall key takeaways:</p><ul><li>Switzerland consistently ranks high in my (far from complete) analysis, although the cost of living matches its rank.</li><li>The results for Northern and Western European countries suggest that there is uniformity across the EU (EEA) countries for most of the measured aspects.</li><li>Romania, although part of EU, has still some way to go to match Western EU levels.</li><li>USA is definitely a rich country but the average middle-class inhabitants enjoy less of that wealth than in the case of European countries. Conversely, USA matches the overall quality of life of most European Countries but it is perceived as less safe.</li></ul><blockquote>Surely you have your own view about where to live.</blockquote><p>There are many more relevant aspects that I do not cover here, such as quality of education or climate. To learn more about this analysis and to tailor it for your own view, see the link to my Github available <a href="https://github.com/alex-bogatu/where2live">here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c505885fd6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Format Transformation — writing rules — The Data Value Factory]]></title>
            <link>https://medium.com/the-data-value-factory/format-transformation-writing-rules-the-data-value-factory-2b7d8ea39084?source=rss-84b01c952586------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b7d8ea39084</guid>
            <category><![CDATA[data-wrangling]]></category>
            <category><![CDATA[data-preparation]]></category>
            <dc:creator><![CDATA[Alex Teodor]]></dc:creator>
            <pubDate>Thu, 27 Feb 2020 11:02:42 GMT</pubDate>
            <atom:updated>2020-03-02T13:36:39.539Z</atom:updated>
            <content:encoded><![CDATA[<h3>Format Transformation — writing rules — The Data Value Factory</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*CAg1L4HX8Hr8gGL8.jpg" /></figure><p>Format transformation is the process of carrying out changes to the representation of data values, such as values of a table column, with a view to reducing some of the existing inconsistencies. In this blog, we set out to show one approach of how to format transformation is being carried out in practice. As an example, consider the scenario exemplified below (the full data set can be checked <a href="https://public.enigma.com/datasets/the-tate-collection-artists/14e97d55-e411-4455-a544-638189b72dce">here</a>), where the values for <em>Name </em>have to be formatted from their existing structure: <em>last_name, first_names </em>to a new format, say <em>first_name last_name</em>.</p><p>Adams, Robert → Robert Adams</p><p>Blais, Jean-Charles → Jean-Charles Blais</p><p>Churchill, Sir Winston → Sir Winston Churchill</p><p>Cooper, Richard, Junior → Richard Cooper Junior</p><p>Such a task is commonly performed programmatically by writing rules that either rely on some programming libraries, e.g. string manipulation libraries, or use a specialized language, e.g., regular expressions. In this process, often performed by data scientists, the following steps are carried out:</p><ol><li>Become familiar with the data and the patterns present on the <em>Name </em>column.</li><li>Create a program/script that extracts each component, e.g., <em>first_name</em>, <em>last_name</em>, and concatenates them in the appropriate order. Another option is to split the column by comma and concatenate the resulting two columns in inverted order.</li><li>Validate the result by checking the correctness of resulting values for some rows, especially for ones that represent special cases, e.g., the last row in the table above.</li></ol><p>Note that, in performing Step 1 above, the data scientist has to identify the major patterns that inform the subsequent steps. This rapidly becomes infeasible if the dataset exceeds a few hundred rows. Step 2 requires knowledge of string manipulation techniques, e.g., based on regular expressions, and knowledge of a language in which transformative rules can be written and executed. This makes format transformation accessible only to technical users. Furthermore, such a program has to be created every time a different column is transformed. Finally, in Step 3, another, potentially time-expensive, exploratory analysis is required to identify corner cases that have not been caught by the previous rules. Such challenges are widespread in data preparation, and dealing with them often delays or even hinders successful data analysis.</p><h3>Format Transformation Rules in Python</h3><p>At its core, format transformation relies on transformation rules that are either written directly by a programmer in some programming language or built through the interaction of the user with a graphical interface provided by a specialized data-preparation tool. The code snippet below is a Python program that, given the values of <em>Name</em>, uses regular expressions to obtain the values of <em>New Name</em></p><pre><em>def transformation_rule(pattern, in_value):<br>    matches = pattern.findall(in_value)<br>    if matches is not None and len(matches) &gt; 1:<br>        return matches[1][1] + “ “ + matches[0][1] + “ “ +<br>             “”.join([t[1] for t in matches[min(2, len(matches)):]])<br>    return in_value</em></pre><pre><em>in_values = [‘Adam, Robert’, ‘Blais, Jean-Charles’, ‘Churchill, Sir 8. Winston’, ‘Cooper, Richard, Junior’]<br>pattern = re.compile(r’(,\s?)?([\w\-\s]+)’)<br>transformed = [transformation_rule(pattern, value) for value in in_values]<br>print(transformed)</em></pre><p><em>Output: </em>[‘Robert Adam ‘, ‘Jean-Charles Blais ‘, ‘Sir Winston Churchill ‘, ‘Richard Cooper Junior’]</p><p>The program, through the <em>transformation_rule </em>function, transforms each value that matches the <em>pattern </em>to its corresponding output. <em>pattern</em> is a regular expression that is consistent with any value consisting of one or more words separated by a comma and/or space and/or hyphen. Identifying such a regular expression pattern or a set of such patterns that cover all input values is infeasible when there are many values. Furthermore, automatically inferring one or more such regular expressions from a given collection of matching strings is a non-trivial computing problem.</p><h3>Format Transformation Rules in Trifacta</h3><p>An alternative to writing transformation rules in the form of programs or scripts is to use a data wrangling tool such as Wrangler, commercialized by Trifacta. Most of the advantages offered by such tools over the more conventional programmatic transformations lie with Steps 1 and 3 above. For instance, wrangling tools offer data preview features that allow the identification of existing representation patterns automatically and flagging of problematic strings that cannot be transformed by the current transformations. With respect to Step 2 above, some data wrangling tools offer domain-specific transformation languages that are more grounded into natural language, or even go as far as to provide GUI representations of common transformation functions, therefore, making format transformation accessible to less technical users. Despite their simpler transformation language though, in performing data wrangling using specialized tools, Step 2 from above can still become infeasible quickly. This is because, even if sometimes the system is capable of providing suggestions regarding potentially suitable transformations, the user still has to know how to express the transformations using the underlying language (or what combination of GUI representations achieves the task).</p><p>As an example, consider the transformation script specific to Data Wrangler below.</p><pre><em>splitpatterns</em> col: Name on: /,\s/ ignoreCase: true<br><em>replacepatterns</em> col: Name1 with: “ on: `{start}”|”{end}` global: true<br><em>replacepatterns</em> col: Name2 with: “ on: `{start}”|”{end}` global: true<br><em>merge</em> col: Name2,Name1 with: ‘ ‘ as: ‘Full Name’</pre><p>The script splits the column into two sub-columns: <em>Name1 </em>and <em>Name2</em>, removes the quotes characters added when ingesting the data and concatenates the values of the two sub-columns separated by space. Note that the script above will not correctly transform the last row of our example. So the user has to incrementally edit the script to include all the formats in the <em>Name </em>column. Note also that a more complex script, e.g., based on regular expressions that match most of the existing formats, can be created using data wrangling tools. In such a case the complexity of Step 2 above would be very similar to performing format transformations using Python.</p><p>It transpires that, although successful in speeding up format transformation, data preparation tools such as Data Wrangler do not eradicate the need for specialized knowledge of data and/or the language used in writing transformation rules. Subsequent posts of this series will discuss alternative approaches to format transformation that aim to facilitate transformations even in the absence of data and programming knowledge.</p><h3>Summary</h3><p>The traditional programmatic approach to data wrangling offers efficient and flexible transformation methods for experienced users familiar with both the data and the programming language used. As data wrangling has become a necessity for many analysis scenarios, specialized wrangling tools have been proposed, with the aim of enabling data wrangling by less-technical users. As shown in this post, when using such tools, data wrangling tasks, such as format transformation, require less familiarity with the data, but similar levels of knowledge of the necessary transformations and of how to express them in the underlying language.</p><p>Subsequent blogs in this series will discuss proposed ideas, most of them actively researched, that aim to alleviate the user from having to be an experienced programmer in order to write transformation rules.</p><p><em>Originally published at </em><a href="https://thedatavaluefactory.com/format-transformation-writing-rules/"><em>https://thedatavaluefactory.com</em></a><em> on February 27, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b7d8ea39084" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-data-value-factory/format-transformation-writing-rules-the-data-value-factory-2b7d8ea39084">Format Transformation — writing rules — The Data Value Factory</a> was originally published in <a href="https://medium.com/the-data-value-factory">The Data Value Factory</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>