<?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 Frantisek Nushart on Medium]]></title>
        <description><![CDATA[Stories by Frantisek Nushart on Medium]]></description>
        <link>https://medium.com/@frantiseknushart?source=rss-86cf4d068fc9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xh_nFf8YSDu7GmS-lqf0Yw.jpeg</url>
            <title>Stories by Frantisek Nushart on Medium</title>
            <link>https://medium.com/@frantiseknushart?source=rss-86cf4d068fc9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 14:59:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@frantiseknushart/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[How to Innovate with Data Products While Protecting User Data?]]></title>
            <link>https://medium.com/geekculture/how-to-innovate-with-data-products-while-protecting-peoples-privacy-and-sensitive-information-3f910a349b6a?source=rss-86cf4d068fc9------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f910a349b6a</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Frantisek Nushart]]></dc:creator>
            <pubDate>Wed, 11 Aug 2021 11:36:41 GMT</pubDate>
            <atom:updated>2022-01-14T15:44:13.268Z</atom:updated>
            <content:encoded><![CDATA[<h4>Generate Novel Data with Conditional GANs.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5SGTxPEdw-PPme_9hsuRzA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@davidjorre?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">David Jorre</a> on <a href="https://unsplash.com/@davidjorre?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p><strong>Generative Adversarial Network</strong> (GAN) is a conceptually quite exciting deep learning approach that lets us produce new data samples across various modalities such as images, video, text, sound or structured tabular data.</p><p>While sitting on the backseat in terms of the global business value it currently generates (in comparison to supervised learning solutions), GANs offer means for what one could perhaps call <strong><em>artificial creativity</em></strong><em> </em>and for generating stunning, <strong><em>almost-like-real</em></strong> visual outputs (with a huge body of work done namely in face imaging).</p><p>There’s just something magical about generative networks and in seeing their applications evolve in quality from year to year while entirely novel use cases keep cropping up.</p><p>In the light of those flashy applications, generating artificial tabular data might at first seem like an underwhelming objective. Yet that couldn’t be further from the truth. Data is the single most important thing when it comes to machine learning. Our ability to synthesise artificial distributions that well reflect properties of the original can prove extremely valuable.<br>Not as an end all be all, but a catalyst to innovation and enabler of early data-driven solution prototyping.</p><h3>Goals</h3><p>Some of the main goals for generating new synthetic data are dataset <strong>protection, enhancement, </strong>and<strong> addressing of imbalance</strong>.</p><blockquote><strong>Certain types of data are extremely sensitive and rightfully protected. I‘ve experienced this first-hand while designing product features stemming from user’s banking transaction data analysis. Even just sample data to prototype with can prove to be challenging to acquire.</strong></blockquote><p>This is where generated data can come to the rescue and effectively anonymise records while accurately preserving the original dataset form and properties. The potential of owned data as an asset could be unlocked this way without posing privacy concerns and in turn stimulate innovation efforts based on analysis, insight gathering and machine learning solutions. Outside of the mentioned finance industry, this can greatly impact other areas where we‘re dealing with sensitive information (for instance, healthcare, security, and government).</p><h3>Implementation</h3><p>We are going to synthesize data based on <a href="https://www.kaggle.com/apoorvwatsky/bank-transaction-data">this Bank Transactions dataset</a> with 116000 transactions across ten different accounts and to focus our scope, we are specifically concerned about the fund withdrawal transactions. The validity and evaluation of the original dataset itself aren’t of particular interest for us here, so bear with me when the account balances and numerical values in Indian Rupees (INR) might seem a bit strange.</p><p>Our goal is to simply generate novel data and analyze its properties with a direct comparison to the original.</p><p>We start by downloading the dataset from Kaggle (with our Kaggle API keys located in the connected Google Drive folder) and extracting the file.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/f0e9c974136fe5cb8fab3cd4584a7ecf/href">https://medium.com/media/f0e9c974136fe5cb8fab3cd4584a7ecf/href</a></iframe><p>With Pandas library, we create a data frame for the original data and drop unwanted columns and samples containing no withdrawal amount value (those are the deposits).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b0a83eee8ad615adfe44827348b22ba5/href">https://medium.com/media/b0a83eee8ad615adfe44827348b22ba5/href</a></iframe><p>We install the <a href="https://sdv.dev/SDV/">CTGAN library</a> and import its synthesizer while marking the ‘Account No’ column as discrete.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/9a39d1f2d098b9489eca59db075af0a9/href">https://medium.com/media/9a39d1f2d098b9489eca59db075af0a9/href</a></iframe><p>And instantiate &amp; fit the model with our original dataset, marked discrete columns and run it for 20 epochs.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/925abe81be55fd8d20b048c1f6c7dff7/href">https://medium.com/media/925abe81be55fd8d20b048c1f6c7dff7/href</a></iframe><p>Now everything’s ready for us to draw the number of samples equal to the original dataset, and merge everything into a new data frame.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/94a7044f219d39b8cb481207e2591f55/href">https://medium.com/media/94a7044f219d39b8cb481207e2591f55/href</a></iframe><p>If you want to learn about CTGAN in detail, you can check out the <a href="https://arxiv.org/abs/1907.00503">original paper</a> introducing the architecture.</p><h3>Comparing the Original and GAN Generated Dataset</h3><p>As a first look at how our generated dataset stands against the original, we print out basic statistical summaries with Pandas data frame <em>.describe() </em>function and do a comparison.</p><pre>df.describe()<br>samples.describe()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*Fz5grfrbcjnBjYEnqJ60pQ.png" /><figcaption>Fig. 1— Basic statistical data summary of the <strong>original dataset</strong> (by Author).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/299/1*dDsbA5lBik4Gd16gqyRrfQ.png" /><figcaption>Fig. 2— Basic statistical data summary of the <strong>GAN</strong> <strong>generated dataset</strong> (by Author).</figcaption></figure><p>Outside of some shortcomings in withdrawal amount outliers, the generated dataset seems to be doing pretty well!</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/79f4e1a1e3a3aad62f24d6bae7cdc5b4/href">https://medium.com/media/79f4e1a1e3a3aad62f24d6bae7cdc5b4/href</a></iframe><p>The violin plots below (comparing the two datasets&#39; withdrawal and balance amount feature distributions) support those findings.</p><p>You can use <a href="https://colab.research.google.com/drive/1X67maZlhCfaNVTNHHmjpT9ecE3jKJ2nR?usp=sharing">this Jupyter notebook</a> to inspect the graphs in an interactive form.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*p1zcq6TzlIsS8tINvBB3AA.png" /><figcaption>Fig. 3— Violin plot comparing the continuous feature distribution (‘Withdrawal Amount’) of the original dataset to the GAN-generated one (by the author with <a href="https://plotly.com/">Plotly</a>).</figcaption></figure><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1e79ff0b3f7e7dc75e9b3161a4a4c478/href">https://medium.com/media/1e79ff0b3f7e7dc75e9b3161a4a4c478/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*mRQcFKSA2xaf71Y8prOi4g.png" /><figcaption>Fig. 4— Violin plot comparing the continuous feature distribution (‘Balance Amount’) of the original dataset to the GAN-generated one (by the author with <a href="https://plotly.com/">Plotly</a>).</figcaption></figure><p>To conclude, we employ an amazing library <a href="https://pypi.org/project/table-evaluator/">Table Evaluator</a> that’s been specially designed to evaluate artificial tabular datasets. All we need to do is input our two data frames:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a5d91271bbc10c16d3fabbf052529c16/href">https://medium.com/media/a5d91271bbc10c16d3fabbf052529c16/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/606/1*_C1mTMQy_9MBVFgsDOTKDg.png" /><figcaption>Fig. 5— <strong>Absolute Log Mean</strong> and <strong>STDs</strong> of numerical data (by the author with <a href="https://pypi.org/project/table-evaluator/">Table Evaluator</a>).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/872/1*MEh1tgj7Y8-pvRVnlewIyA.png" /><figcaption>Fig. 6— <strong>Cumulative Sums</strong> per feature (by the author with <a href="https://pypi.org/project/table-evaluator/">Table Evaluator</a>).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/396/1*3_9BUNVcPcTu_1D8GHAM-A.png" /><figcaption>Fig. 7— <strong>Proportional distribution</strong> of the categorical feature ‘Account No’ (by the author with <a href="https://pypi.org/project/table-evaluator/">Table Evaluator</a>).</figcaption></figure><h3>Your Turn</h3><p>The results are again quite impressive! What’s more, with a supervised learning task in mind, Table Evaluator can further help assess predictive performance (with a given target variable).</p><p>Have you ever been dealing with a problem where this approach could help? Where do you think GAN generated data (tabular or not) could bring substantial value?</p><p>Thanks for reading! 🙏</p><blockquote>I’m excited about how can we use machine learning to solve problems that we wouldn’t be able to solve using other means. <em>💭⚙️</em></blockquote><blockquote><em>Check out </em><a href="http://www.deepinnovation.co.uk/"><strong><em>Deepinnovation</em></strong></a><em> for more from the intersection of design, innovation and novel deep learning use cases!</em></blockquote><ul><li><a href="https://www.deepinnovation.co.uk/">Deep Innovation Cards</a></li><li><a href="https://medium.com/geekculture/visualizing-machine-learning-tasks-with-word-embeddings-c044fa5bb5aa">Visualizing Machine Learning Tasks With Word Embeddings</a></li><li><a href="https://medium.com/geekculture/how-can-deep-learning-transformer-models-amplify-positive-impacts-of-journaling-8ad28517ccc5">How Can Deep Learning Transformer Models Amplify Positive Impacts of Journaling?</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f910a349b6a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/geekculture/how-to-innovate-with-data-products-while-protecting-peoples-privacy-and-sensitive-information-3f910a349b6a">How to Innovate with Data Products While Protecting User Data?</a> was originally published in <a href="https://medium.com/geekculture">Geek Culture</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Can Transformer Models Amplify Positive Impacts of Journaling?]]></title>
            <link>https://medium.com/geekculture/how-can-deep-learning-transformer-models-amplify-positive-impacts-of-journaling-8ad28517ccc5?source=rss-86cf4d068fc9------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ad28517ccc5</guid>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Frantisek Nushart]]></dc:creator>
            <pubDate>Mon, 12 Jul 2021 13:10:59 GMT</pubDate>
            <atom:updated>2022-01-14T15:24:01.094Z</atom:updated>
            <content:encoded><![CDATA[<h4>Using Zero-Shot Classification with Hugging Face Transformers 🤗 to Support Emotional Intelligence in a Journaling Context.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DmUf9smSja1BOTHOTtIFew.jpeg" /><figcaption>Fig.1 — Journaling (Photo by <a href="https://unsplash.com/@craftedbygc?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Green Chameleon</a> on <a href="https://unsplash.com/s/photos/green-chameleon?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a>)</figcaption></figure><p>Picking up journaling can bring surprisingly many benefits ranging from improvements in <strong>mindfulness</strong>, <strong>memory</strong>, or <strong>sleep routine</strong> to <strong>reduced stress</strong> and <strong>immune system boost</strong>.</p><p>One of the key ideas is that journaling helps us reflect on thoughts and feelings as we recount and organise past events. And this is where transformer language models come in with the potential to <strong>support engagement, awareness </strong>and <strong>understanding</strong>. We can decipher the meaning encoded in a journal entry and flag it with the most appropriate label from our predefined, use-case specific set. This could be a wider pool of emotion and feeling descriptors or a narrow one, tailored to a specific part of the user journey.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8M2YP1GnmjQZB5eT4i0JAg.jpeg" /><figcaption>Fig.2 — Deciphering emotions in journal entries (Photo by <a href="https://www.pexels.com/@olly?utm_content=attributionCopyText&amp;utm_medium=referral&amp;utm_source=pexels">Andrea Piacquadio</a> from <a href="https://www.pexels.com/photo/collage-photo-of-woman-3812743/?utm_content=attributionCopyText&amp;utm_medium=referral&amp;utm_source=pexels">Pexels</a>).</figcaption></figure><h3>Opportunity</h3><p>Such sentiment analysis variation can take many shapes in terms of purpose-driven integration into the product context.</p><p>For instance, as <strong>empathic probing question prompts</strong> that offer validation, support and further the reflective journaling posture. Perhaps as one-ofs or branching into a dialogue-like back and forth chatbot exchanges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S3MHe_yVUqYjzrFQYTv9PA.png" /><figcaption>Fig.3 — Feedback and probing questions promoting emotional awareness (by author).</figcaption></figure><p>Taking a more passive role, the model could help us <strong>automatically tag journal entries</strong> and consequently use this metadata for <strong>search</strong>, <strong>categorisation </strong>and <strong>analytics</strong> over time (similarly to how many mindfulness products currently incorporate manual mood tracking).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2nFOfMefK9_pXAyER8zpcw.png" /><figcaption>Fig.4 — Automatic ‘Empathic Tagging’ of journal entries (by author).</figcaption></figure><h3>Zero-Shot Classification Prototype</h3><p>In comparison to a sentiment analysis model trained to distinguish a pre-defined set of sentiments, a transformer language model with zero-shot learning capabilities gives us a lot more design freedom.</p><p>We can benefit from the semantic language understanding of the model where it comes to the categories themselves, allowing us to modify them on the fly at an inference time.</p><p>Thanks to this, we can employ the same pre-trained model in various contexts with different set-ups and even use user input or metadata to inform those set-ups.</p><p>In addition, this allows for extremely fast-paced experimentation and fine-tuning of how the model works within our designed product features.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5eSkB8xa8dZeEzRNVc83ag.gif" /><figcaption>Fig. 5 — Web app with the model prototype (by author).</figcaption></figure><h4>Implementation</h4><p>We use <a href="https://huggingface.co/">Hugging Face</a> high-level Pipeline API to query the model and <a href="https://streamlit.io/">Streamlit</a> to quickly build a test-able web application:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/03a245b30180dc7dba4d8a65fda07aec/href">https://medium.com/media/03a245b30180dc7dba4d8a65fda07aec/href</a></iframe><p>And deploy it with a local tunnel:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e50e8c5204e3d923e8706480839432f2/href">https://medium.com/media/e50e8c5204e3d923e8706480839432f2/href</a></iframe><p>This way of on-demand hosting within the <a href="https://colab.research.google.com/">Google Colab</a> instance allows us to conveniently experiment with the model and app with a minimal setup required!</p><p>You can walk through the code and build the prototype directly in your browser! Continue to <a href="https://colab.research.google.com/drive/1LtC3KY2DeGoavKaOnTw8oHD8hq_7mU1F?usp=sharing">this Jupyter notebook</a> 📔 !</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*gSxv-UKZaWY6LLnjbcmtTA.jpeg" /><figcaption>Fig.6 — Jupyter Notebook (photo by <a href="https://pxhere.com/en/photo/764684">Pxhere</a>).</figcaption></figure><h3>What’s Next?</h3><p>Despite hurdles like privacy concerns or incompatibility with a physical form, journaling practice seems ready to benefit from contemporary natural language processing (NLP) capabilities.</p><p>Startups like <a href="https://morningpages.app/">Morningpages</a> or <a href="https://reflectly.app/">Reflectly</a> seem to agree and do some amazing work innovating in this space, yet something’s telling me there’s still a lot to discover.🤔</p><p>What do you think? What does the future journaling experience have to offer?</p><p>Let me know. And thank you for making it this far! 👏</p><blockquote>I’m excited about how can we use machine learning to solve problems that we wouldn’t be able to solve using other means. <em>💭⚙️</em></blockquote><blockquote><em>Check out </em><a href="http://www.deepinnovation.co.uk/"><strong><em>Deepinnovation</em></strong></a><em> for more from the intersection of design, innovation and novel deep learning use cases!</em></blockquote><p><a href="https://www.deepinnovation.co.uk/">Deep Innovation Cards</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ad28517ccc5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/geekculture/how-can-deep-learning-transformer-models-amplify-positive-impacts-of-journaling-8ad28517ccc5">How Can Transformer Models Amplify Positive Impacts of Journaling?</a> was originally published in <a href="https://medium.com/geekculture">Geek Culture</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Visualizing Machine Learning Tasks With Word Embeddings]]></title>
            <link>https://medium.com/geekculture/visualizing-machine-learning-tasks-with-word-embeddings-c044fa5bb5aa?source=rss-86cf4d068fc9------2</link>
            <guid isPermaLink="false">https://medium.com/p/c044fa5bb5aa</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[visualization]]></category>
            <dc:creator><![CDATA[Frantisek Nushart]]></dc:creator>
            <pubDate>Mon, 28 Jun 2021 11:31:49 GMT</pubDate>
            <atom:updated>2021-07-11T11:52:18.648Z</atom:updated>
            <content:encoded><![CDATA[<h4>Organize over 400 concepts in space based on their semantic similarity (using a GloVe model pre-trained on 2 billion Tweets).</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*qz9gA59Iy496tnJ7BL-FnA.gif" /></figure><blockquote><strong>On the second look, the scarry-looking plot uncovers meaningful spatial relationships (eg. tasks of detecting hate speech, abusive language and emotions are embeded close to each other).</strong></blockquote><p>Machine learning is delivering value to a rapidly increasing breadth of industries. Often in a form of clever and novel use cases that leverage its incredibly horizontal application potential. Contrary to intuition, what lies behind the innovation are quite often well-understood tasks and the creative leap is in their identification and impactful integration into the context of the unique problem at hand.</p><h3>Outline</h3><p>In this article, we peek into the landscape of those tasks as they are categorized under machine learning research datasets by <a href="https://paperswithcode.com/">Papers With Code</a> (an amazing resource that compiles research papers along with the supplied code and data).</p><p>We’ve web scraped<strong> well over 400 machine learning tasks</strong> and <strong>embedded</strong> them into a <strong>25-dimensional space</strong> (based on a semantic understanding of words) using a <a href="https://nlp.stanford.edu/projects/glove/"><strong>GloVe</strong></a> (Global Vectors for Word Representation) model, <strong>pre-trained on 2 billion Tweets</strong>. After reducing the dimensionality from 25D to 3D and 2D, using PCA (Principal Component Analysis), we’ve been able to visualize the tasks in a way that captures meaning within their spatial interrelations.</p><p>You can, for instance, explore the neighbours of a chosen task to discover other semantically similar tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*XxknGVrxU72y7UwQnAxaCA.gif" /><figcaption>Fig. 1–3D Scatter plot with tasks and their principal component coordinates (by author).</figcaption></figure><h3>Implementation</h3><p>You can walk through the code and explore the visualizations hands-on in<br><a href="https://colab.research.google.com/drive/1pi2TMNjwLZakNuj9MYDk3KnRLyAWfXl8?usp=sharing">this Jupyter notebook</a> 📔 !</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*gSxv-UKZaWY6LLnjbcmtTA.jpeg" /><figcaption>Fig.2 — Jupyter Notebook (photo by <a href="https://pxhere.com/en/photo/764684">Pxhere</a>).</figcaption></figure><h4>Web Scraping Machine Learning Tasks</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ORvrF6y4H_pcExekJ1CaxA.png" /><figcaption>Fig.3 — Inspecting page elements (by author).</figcaption></figure><p>Firstly we request the datasets webpage at <a href="https://paperswithcode.com/datasets">Papers With Code</a>.<br>The &lt;Response [200]&gt; indicates that the request has been successful.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4fe9093547a788d67dc7e70b5847e0ad/href">https://medium.com/media/4fe9093547a788d67dc7e70b5847e0ad/href</a></iframe><p>Parsing the Html allows us to locate task names and grasp their enveloping structure. We can interactively do the same by inspecting page elements in a web browser (fig.3).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/12c1e3602aaa643bc9c8caa3c534ac1f/href">https://medium.com/media/12c1e3602aaa643bc9c8caa3c534ac1f/href</a></iframe><p>The <em>find_all </em>function below lets us access all anchor tags of the desired class.<br>The tags that we are interested in fall in-between positions [39] and [538].</p><pre>print(soup.find_all(&#39;a&#39;, class_=&quot;filter-item&quot;)[39])</pre><p>We construct a list out of text following those anchor tags while cleaning up redundant spacing before and after task names (using regular expressions).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/381e147f43191d04dcf65c7156b6c80e/href">https://medium.com/media/381e147f43191d04dcf65c7156b6c80e/href</a></iframe><p>Now that we have our data in place, we’re ready to download the GloVe model! 👏</p><h4>Downloading GloVe Model and Preparing the Input</h4><p>We will use the <a href="https://radimrehurek.com/gensim/">Gensim</a> library by Radim Rehurek to download the pre-trained model:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/576ad70c7f72a0fd7727835db28279ee/href">https://medium.com/media/576ad70c7f72a0fd7727835db28279ee/href</a></iframe><p>And preprocess our <em>list of tasks</em> into a <em>list of lists of words</em> contained in every task:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/0e57257b96c1e26b4abbec81c3a09c35/href">https://medium.com/media/0e57257b96c1e26b4abbec81c3a09c35/href</a></iframe><p>This is for us to be able to process the words within a task individually and follow by averaging them out into a resulting task embedding.</p><p>Alternatively, if we were to train the text corpus ourselves (ideally on unstructured data that accurately aligns with our domain of interest), we could utilize approaches for embedding multi-word phrases. Check out the article by Moshe Hazoom below to find out more about this approach.</p><p><a href="https://towardsdatascience.com/word2vec-for-phrases-learning-embeddings-for-more-than-one-word-727b6cf723cf">Word2Vec For Phrases — Learning Embeddings For More Than One Word</a></p><h4>Embedding Tasks Into 25D Space</h4><p>Some of the machine learning tasks contain words that aren’t included in the pre-trained GloVe model. We will replace those in place with empty lists.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1cf86fdb018dc72ce2bce557812e1f98/href">https://medium.com/media/1cf86fdb018dc72ce2bce557812e1f98/href</a></iframe><p>And to ensure somewhat accurate representations, we will exclude those tasks that are left with less than two words (valid in the GloVe model):</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/aaf5f7ba0af62894d94b81aca1da2429/href">https://medium.com/media/aaf5f7ba0af62894d94b81aca1da2429/href</a></iframe><p>The following runs the valid words through the model to obtain a 25-dimensional vector for each and subsequently averages the results out for multi-word tasks.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/90365f177bb63822631905d702ea554b/href">https://medium.com/media/90365f177bb63822631905d702ea554b/href</a></iframe><p>That leaves us with a 25-dimensional vector for each task 🙏.<br>Before we move onto dimensionality reduction, let’s quickly set in place a data frame of labels for our new task vectors:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b852f7dc075c1380984191766970252d/href">https://medium.com/media/b852f7dc075c1380984191766970252d/href</a></iframe><h4>Reducing Dimensionality and Plotting Interactive Visualizations</h4><p>We will use the popular library <a href="https://scikit-learn.org/stable/">Scikit-learn</a> to perform dimensionality reduction with PCA (Principal Component Analysis) and to plot the visualizations, we will use <a href="https://plotly.com/python/plotly-express/">Plotly Express</a>.</p><p><strong>Word 3D Scatter Plot:</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/cef49c3b15022bd0b0900bb94c00e3ad/href">https://medium.com/media/cef49c3b15022bd0b0900bb94c00e3ad/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplotly.com%2F%7EFranta%2F1.embed%3Fautosize%3Dtrue&amp;display_name=Plotly&amp;url=http%3A%2F%2Fchart-studio.plotly.com%2F%7EFranta%2F1%2F&amp;image=http%3A%2F%2Fchart-studio.plotly.com%2Fstatic%2Fwebapp%2Fimages%2Fplotly-logo.8d56a320dbb8.png&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=plotly" width="600" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/e4d4447b7bcb9fed0767a4b7959ff27c/href">https://medium.com/media/e4d4447b7bcb9fed0767a4b7959ff27c/href</a></iframe><p><strong>3D Scatter Plot:</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/7d373bb0e9f8cbff3ee6c6b24992053c/href">https://medium.com/media/7d373bb0e9f8cbff3ee6c6b24992053c/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplotly.com%2F%7EFranta%2F3.embed%3Fautosize%3Dtrue&amp;display_name=Plotly&amp;url=https%3A%2F%2Fchart-studio.plotly.com%2F%7EFranta%2F3%2F&amp;image=https%3A%2F%2Fchart-studio.plotly.com%2Fstatic%2Fwebapp%2Fimages%2Fplotly-logo.8d56a320dbb8.png&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=plotly" width="600" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/4fb3b0e9855d7ddbff62f1bb57106ae0/href">https://medium.com/media/4fb3b0e9855d7ddbff62f1bb57106ae0/href</a></iframe><p><strong>2D Scatter Plot</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/0101e1dece03d8232e2aade2cb812053/href">https://medium.com/media/0101e1dece03d8232e2aade2cb812053/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplotly.com%2F%7EFranta%2F5.embed%3Fautosize%3Dtrue&amp;display_name=Plotly&amp;url=https%3A%2F%2Fchart-studio.plotly.com%2F%7EFranta%2F5%2F&amp;image=https%3A%2F%2Fchart-studio.plotly.com%2Fstatic%2Fwebapp%2Fimages%2Fplotly-logo.8d56a320dbb8.png&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=plotly" width="600" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/883ee26499f76e08f125fddfc6f69b27/href">https://medium.com/media/883ee26499f76e08f125fddfc6f69b27/href</a></iframe><h3>Wrapping Up</h3><p>We’ve looked into how we can utilize the machine learning technique of word embeddings to explore and understand the machine learning tasks landscape. 🙏</p><p>And in the same breath, we’ve covered the approach practicalities in-depth so that you can apply it in different settings to solve interesting problems at hand! 👏</p><blockquote>I’m excited about how can we use machine learning to solve problems that we wouldn’t be able to solve using other means. <em>💭⚙️</em></blockquote><blockquote><em>Check out </em><a href="http://www.deepinnovation.co.uk/"><strong><em>Deepinnovation</em></strong></a><em> for more from the intersection of design, innovation and novel deep learning use cases!</em></blockquote><p><a href="https://www.deepinnovation.co.uk/">Deep Innovation Cards</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c044fa5bb5aa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/geekculture/visualizing-machine-learning-tasks-with-word-embeddings-c044fa5bb5aa">Visualizing Machine Learning Tasks With Word Embeddings</a> was originally published in <a href="https://medium.com/geekculture">Geek Culture</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>