<?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 Lokesh Prakash Manohar on Medium]]></title>
        <description><![CDATA[Stories by Lokesh Prakash Manohar on Medium]]></description>
        <link>https://medium.com/@lokeshonai?source=rss-c97c85517f85------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*kV1PloPolixBIEzyQK14BA.png</url>
            <title>Stories by Lokesh Prakash Manohar on Medium</title>
            <link>https://medium.com/@lokeshonai?source=rss-c97c85517f85------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:43:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lokeshonai/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[What is Model Drift in AI & ML? And why should you care?]]></title>
            <link>https://medium.com/@lokeshonai/what-is-model-drift-in-ai-ml-and-why-should-you-care-087e4347ac49?source=rss-c97c85517f85------2</link>
            <guid isPermaLink="false">https://medium.com/p/087e4347ac49</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[model-drift]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Lokesh Prakash Manohar]]></dc:creator>
            <pubDate>Fri, 07 Mar 2025 16:32:21 GMT</pubDate>
            <atom:updated>2025-03-07T16:32:21.801Z</atom:updated>
            <content:encoded><![CDATA[<p>Decision-makers should carefully consider the cost of maintaining artificial intelligence and machine learning models since they require constant monitoring, tuning, and retraining to remain effective. Compared to maintaining software applications, how do maintaining AI and ML models differ?</p><p>The answer lies in the why. Software applications need to be maintained to account for changes in user behavior, business needs, and technology. AI and ML models also need to be maintained for the above reasons but, in addition, they need to be maintained to account for Model Drift.</p><h3>What is Model Drift?</h3><p>Model drift is a phenomenon that can occur in machine learning when the statistical properties (distribution) of the data changes over time. This can cause a model to become less accurate, degrade in performance and produce unexpected results. The distribution of data can change for a variety of reasons, including changes in the environment (increase in commute time as the population of a city grows), changes in the way data is collected or labeled (using cctv cameras to measure footfalls versus satellite imagery of the store’s parking lot), or a change in user behavior (avoiding or reducing screen time before going to bed).</p><p>Sometimes, the original data may not change much but, the feature derived by applying mathematical transformations may change meaningfully impacting the model performance.</p><h3>What are the two main types of model drift?</h3><p>Concept drift refers to the changes in the statistical properties of the dependent variable, the datapoint that the model was built to predict. A model that predicts the number of felonies in a city will become inaccurate if, for example, some felony offenses are downgraded to misdemeanors after a law is passed. A model that predicts app ratings would become inaccurate if, for some reason, the app store (or Google Play) switches to a ten-point scale instead of five stars.</p><p>Data drift occurs due to the changes in the statistical properties of the independent variables or the relationship between them. A model that predicts the daily electricity consumption of a city will become inaccurate if the gradual increase in temperature due to climate change is not taken into account. It is possible for a model that predicts EV sales to underestimate them if it does not take into account the effect of expanding charger networks.</p><h3>How to address model drift?</h3><p>There are a few things you can do to avoid or mitigate model drift:</p><ul><li>monitor the statistical properties of raw data and features for changes</li><li>retrain your models on a regular basis</li><li>make sure your training data is representative of the data you’ll be using to make predictions</li><li>use human feedback loops to validate your predictions</li></ul><p>Since model drift is a continual problem, it’s important to have a process to monitor distributional shifts over time and be proactive in retraining models. The good news is, most modern AI application management platforms offer out-of-the-box model monitoring services.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=087e4347ac49" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why do ML product managers and stakeholders care about Feature Importance?]]></title>
            <link>https://medium.com/@lokeshonai/why-do-ml-product-managers-and-stakeholders-care-about-feature-importance-271fea14e801?source=rss-c97c85517f85------2</link>
            <guid isPermaLink="false">https://medium.com/p/271fea14e801</guid>
            <category><![CDATA[interpretable-ai]]></category>
            <category><![CDATA[feature-engineering]]></category>
            <category><![CDATA[ai-product-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Lokesh Prakash Manohar]]></dc:creator>
            <pubDate>Fri, 07 Mar 2025 16:30:56 GMT</pubDate>
            <atom:updated>2025-03-07T16:30:56.372Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The purpose of this article is to explain Feature Importance in simple terms and demonstrate how ML product managers use it to make more informed choices while building models.</em></p><h3>What is Feature Importance in machine learning?</h3><p>Feature Importance in machine learning is a measure of how important each feature is in predicting the outcome of the model. For example, if we were trying to predict how much money restaurants make, Feature Importance would show which variables — like location, menu items, price or reviews — contribute the most to predicting accurate sales figures. Feature Importance does not denote the magnitude of contribution but rather the relative contribution of each feature to the accuracy (or any desired performance metric) of the model. In the example above, location might contribute the most to sales (indicated by its coefficient in the model), but price could be the top feature for predicting sales accurately.</p><h3>How is Feature Importance computed?</h3><p>Feature Importance is usually expressed as a score for each feature, with higher scores indicating greater importance. The scores are calculated using a variety of techniques depending on the type of model being used. When using decision tree models, one popular way to measure Feature Importance is called “Gini Importance”. Gini Importance measures how well a single feature can split data into classes. When using regression models, Feature Importance is usually computed by measuring things like “Permutation Feature Importance”. This method finds out how much the accuracy of the model changes when you shuffle the values of one specific feature. Model agnostic methods such as LIME and SHAP are frequently used to explain complex models and calculate Feature Importance.</p><p>The ranking of a feature’s importance changes depending on which performance metric you are trying to optimize. For example, if you want to prioritize precision over accuracy, the feature scores would be different.</p><h3>What are the benefits of Feature Importance scores?</h3><p>Feature Importance helps to explain which features have the most influence on the model’s performance. This is helpful for product managers and stakeholders because it provides useful insights that can help with making decisions about building models. In particular, by analyzing Feature Importance scores, ML product managers can influence decisions in the following areas of model building:</p><ul><li>Feature engineering (For instance, if you’re trying to predict used car sales, the model’s accuracy will probably improve by using price ranges as an input feature instead of utilizing the actual prices.)</li><li>Data gathering (If the ranking of features based on their importance is not as expected, then gathering more data for training could be a possible solution.)</li><li>Model optimization (Features of lesser importance may be omitted to reduce overfitting and to reduce data and computational overhead during inferencing.)</li></ul><p>Additionally, Feature Importance can provide a clearer picture of why certain predictions are made and help explain complex models to stakeholders. For example, understanding which features are most important in a clustering model can help explain why data points are being grouped together. This information can be used for tasks such as labeling clusters or devising optimal marketing campaigns for different groups of customers.</p><p>Feature Importance is a very important tool for building explainable, efficient, and effective ML models. Additionally, Feature Importance can help identify opportunities for key goals such as sales growth and cost reduction.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=271fea14e801" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>