<?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[Ellora AI - Medium]]></title>
        <description><![CDATA[Technical articles on behalf of Ellora AI LLC and Ellora AI Solutions - Medium]]></description>
        <link>https://medium.com/ellora-ai?source=rss----f22a3c540553---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Ellora AI - Medium</title>
            <link>https://medium.com/ellora-ai?source=rss----f22a3c540553---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 17:24:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/ellora-ai" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Understanding and Mastering Your FICO Score]]></title>
            <link>https://medium.com/ellora-ai/understanding-and-mastering-your-fico-score-8d0c3a367ef0?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/8d0c3a367ef0</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Fri, 07 Feb 2025 12:57:25 GMT</pubDate>
            <atom:updated>2025-02-07T12:57:25.030Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Writers:</strong></h3><p><a href="https://medium.com/u/25b22d7e4148?source=post_page---user_mention--8095befbcaf0--------------------------------"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5?source=post_page---user_mention--8095befbcaf0--------------------------------"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><strong><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></strong></p><p><em>Disclaimer: The content is for informational purposes only. You must not construe this information as financial advice</em></p><p><strong>Your credit score can make or break your financial future</strong>.</p><p>But not just any credit score — we’re talking about your FICO® score, the most widely used credit score by lenders across the United States. Whether you’re planning to buy a home, finance a car, or simply want better credit card terms, understanding your FICO score is crucial for your financial success.</p><h3>What Is a FICO Score?</h3><p>Created by the Fair Isaac Corporation in 1956, FICO scores are three-digit numbers ranging from 300 to 850 that summarize your creditworthiness. Think of it as your financial report card: the higher your score, the more trustworthy you appear to lenders. It tells lenders how risky it would be to lend you money.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/0*j6FY2Km9iAVSg9Ic.jpg" /><figcaption><a href="https://www.simmonsbank.com/siteassets/content-hub/learning-center/credit-score-image.jpg">source</a></figcaption></figure><h3>The Three FICO Scores You Actually Have</h3><p>What many people don’t realize is that you don’t have just one FICO score — you have three, one from each major credit reporting agency:</p><ul><li>Equifax: BEACON</li><li>Experian: Experian/Fair Isaac Risk Model</li><li>TransUnion: FICO Risk Score, Classic</li></ul><p>While these scores are calculated using the same method, they might differ slightly because each credit bureau may have different information about your credit history.</p><h3>The Five Pillars of Your FICO Score</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Svn412WPxCjCxAnF.png" /><figcaption><a href="https://www.google.com/url?sa=i&amp;url=https%3A%2F%2Fwww.myfico.com%2Fcredit-education%2Fwhats-in-your-credit-score&amp;psig=AOvVaw2M0LE1mS2Wc-8SJ2M1lFxF&amp;ust=1737664479100000&amp;source=images&amp;cd=vfe&amp;opi=89978449&amp;ved=0CBcQjhxqFwoTCJDtvOSWiosDFQAAAAAdAAAAABAg">source</a></figcaption></figure><h3>1. Payment History (35%) — The Foundation of Your Score</h3><p>The single most crucial factor in your FICO score is how well you’ve paid your bills in the past. This includes:</p><ul><li>Payment record on credit cards, retail accounts, installment loans, and mortgages</li><li>Public records and collection items (bankruptcies, foreclosures, suits, liens, judgments)</li><li>How late any payments were (30, 60, 90 days)</li><li>How many past-due payments you have</li><li>How recently these late payments occurred</li><li>Track record across different types of credit</li></ul><p><strong>Pro Tip</strong>: Even one missed payment can significantly impact your score. Set up automatic payments to never miss a due date.</p><h3>2. Amounts Owed (30%) — Your Credit Utilization</h3><p>This factor looks at your credit utilization — how much of your available credit you’re using. Key considerations include:</p><ul><li>Total amount owed across all accounts</li><li>Amount owed on specific types of accounts</li><li>Number of accounts with balances</li><li>How much of your installment loans you’ve paid down</li><li><strong>Credit utilization ratio on revolving accounts</strong></li><li>Balance-to-limit ratio on credit cards</li></ul><p><strong>Golden Rule</strong>: Keep your credit card balances below 30% of your credit limits. For example, if you have a $10,000 credit limit, try to keep your balance under $3,000.</p><h3>3. Length of Credit History (15%) — Credit Maturity</h3><p>The longer your credit history, the better. FICO considers:</p><ul><li>How long your credit accounts have been established</li><li>Age of your oldest account</li><li>Age of specific accounts</li><li>How long certain accounts have been active</li><li>Average age of all your accounts</li></ul><p><strong>Strategy Tip</strong>: Keep your oldest credit cards open, even if you rarely use them. They’re boosting your credit age!</p><h3>4. New Credit (10%) — Recent Activity</h3><p>Opening several credit accounts in a short period can indicate higher risk. This category examines:</p><ul><li>How many recently opened accounts you have</li><li>Recent credit inquiries</li><li>Time since recent account openings</li><li>Pattern of credit account establishment</li><li>Rate shopping for specific loans</li></ul><p><strong>Important</strong>: Multiple inquiries for the same type of loan (like a mortgage) within 14–30 days count as one inquiry.</p><h3>5. Types of Credit in Use (10%) — Credit Mix</h3><p>Having a diverse mix of credit types can help your score. This includes:</p><ul><li>Credit cards (revolving credit)</li><li>Retail accounts</li><li>Installment loans</li><li>Finance company accounts</li><li>Mortgage loans</li><li>Car loans</li><li>Personal loans</li></ul><p><strong>Smart Move</strong>: Don’t open new accounts just for diversity. Only apply for credit you actually need.</p><blockquote>Important Note About FICO Score Calculation :</blockquote><blockquote>While we’ve outlined the five key components that make up your FICO score, it’s crucial to understand that there isn’t a clear mathematical formula available to the public. The exact calculation method is a proprietary algorithm owned by the Fair Isaac Corporation (FICO), and they don’t disclose the precise calculations.</blockquote><blockquote>Think of it like a secret recipe — while we know the main ingredients (Payment History, Amounts Owed, Length of Credit History, New Credit, and Types of Credit in Use) and their proportions, the exact mixing method remains confidential. This means that while we can understand and work with the general components that influence our scores, we cannot calculate our exact FICO scores using a simple mathematical formula.</blockquote><blockquote>This proprietary nature of the FICO algorithm helps maintain the integrity of the credit scoring system while protecting Fair Isaac Corporation’s intellectual property.</blockquote><p>What FICO Scores Don’t Consider</p><p>It’s equally important to understand what doesn’t affect your FICO score:</p><ul><li>Race, color, religion, national origin, sex, or marital status</li><li>Age</li><li>Salary, occupation, title, employer, employment history</li><li>Where you live</li><li>Interest rates on your credit cards</li><li>Child/family support obligations</li><li>Rental agreements</li><li>Soft inquiries (when you check your own credit)</li><li>Non-predictive information</li></ul><h3>The Impact of Your FICO Score: A Real-World Example</h3><p>Here’s a compelling reason to care about your score: <strong>On a $150,000 30-year fixed mortgage, improving your FICO score from 550 to 720 could save you approximately $165,000 over the life of the loan — that’s $459 less on each monthly payment!</strong></p><h3>Your Action Plan for a Better Score</h3><h3>Immediate Actions (0–3 months)</h3><ul><li>Set up automatic payments for all bills. Or minimum payment to avoid late payments.</li><li>Pay down credit card balances</li><li>Stop applying for new credit</li><li>Check your credit reports for errors</li><li>Dispute any inaccuracies on your credit report</li></ul><h3>Short-term Strategy (3–6 months)</h3><ul><li>Make all payments on time</li><li>Keep credit utilization under 30%</li><li>Don’t close old credit cards</li><li>Dispute any inaccuracies on your credit report</li></ul><h3>Long-term Strategies (6+ months)</h3><ul><li>Keep credit utilization under 30%</li><li>Keep old accounts active with small purchases</li><li>Maintain a mix of credit types</li><li>Build an emergency fund to avoid missing payments</li><li>Monitor your credit reports regularly</li><li>Try to obtain a yearly 3-Bureau credit report.</li><li>You can obtain a yearly free credit report in US from <a href="https://www.annualcreditreport.com/index.action">annualcreditreport.com</a></li></ul><h3>Common Myths Debunked</h3><p>❌ Myth: Checking your own credit score lowers it <br>✅ Truth: Self-checks are “soft inquiries” and don’t impact your score</p><p>❌ Myth: Closing unused credit cards helps your score <br>✅ Truth: This can actually hurt by increasing utilization and reducing average account age</p><p>❌ Myth: You need to carry a balance to build credit <br>✅ Truth: Paying your full balance each month is best for your score. <br>If you cannot, do pay the minimum to avoid late payments and penalties and remain in good standing. Note paying minimum means interest payments will be higher and longer time to pay off the debt.</p><h3>Protecting Your Credit Identity</h3><ul><li>Monitor your credit reports regularly through <a href="https://www.annualcreditreport.com/index.action">AnnualCreditReport.com</a></li><li>Set up fraud alerts</li><li>Keep your Social Security Number card safe (and do not carry it around in wallet)</li><li>Shred documents containing personal information</li><li>Report unauthorized activity immediately</li><li>Keep personal information secure</li></ul><h3>The Bottom Line</h3><p>Your FICO score is more than just a number — it’s a key that unlocks better financial opportunities. While improving it takes time, the potential savings make it worth the effort. Start with one small step today, whether it’s setting up automatic payments or checking your credit report.</p><p>Remember: The journey to excellent credit is a marathon, not a sprint. Focus on consistent, responsible credit management, and your score will reflect your efforts over time. Every financial journey begins with a single step.</p><p>Note: FICO score ranges and their impacts on lending decisions may vary by lender. Always consult with specific lenders about their requirements.</p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d0c3a367ef0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/understanding-and-mastering-your-fico-score-8d0c3a367ef0">Understanding and Mastering Your FICO Score</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PART2 : DT+GNN: A Fully Explainable Graph Neural Network using Decision Trees]]></title>
            <link>https://medium.com/ellora-ai/part2-dt-gnn-a-fully-explainable-graph-neural-network-using-decision-trees-16dcb003a9ae?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/16dcb003a9ae</guid>
            <dc:creator><![CDATA[Dhaval Shah]]></dc:creator>
            <pubDate>Mon, 13 Jan 2025 18:11:09 GMT</pubDate>
            <atom:updated>2025-01-15T01:44:24.935Z</atom:updated>
            <content:encoded><![CDATA[<h3>PART2 : DT+GNN: A Fully Explainable Graph Neural Network using Decision Trees</h3><h3><strong>Writers:</strong></h3><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><strong><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></strong></p><p>While <a href="https://medium.com/ellora-ai/part1-dt-gnn-making-ai-decisions-transparent-with-decision-trees-and-graph-neural-networks-8095befbcaf0">Part 1</a> introduced the DT+GNN approach, here we’ll take a deeper look at the model’s components and its step-by-step interpretability framework, focusing on the Diff-DT+GNN layer and its differentiable mechanisms.</p><p>Traditional GNNs excel at analyzing graph data but are often seen as opaque due to their complex, multi-layer structures. DT+GNN tackles this by replacing the standard multi-layer perceptrons (MLPs) in GNNs with decision trees, which provide interpretable, stepwise decision paths.</p><h3>Building the DT+GNN Model</h3><h3>1. Differentiable Diff-DT+GNN Layer</h3><p>The Diff-DT+GNN layer is the core component of DT+GNN. Unlike traditional GNNs, which rely on continuous states for nodes, Diff-DT+GNN simplifies this by modeling states in a discrete, <strong>categorical space</strong>. <strong>This discrete space aligns with the decision tree structure</strong>, making reasoning processes more interpretable.</p><h3>2. Update Function with Differentiable State Space</h3><p>The <strong>update function</strong> processes messages from neighboring nodes and the current node’s state to produce a new state. To make this process differentiable, it uses the <strong>Gumbel-Softmax</strong> technique, which allows sampling from a categorical distribution in a differentiable manner. This is achieved by introducing controlled randomness via <strong>Gumbel noise</strong> and a <strong>temperature parameter</strong> that adjusts the output’s “discreteness” or “softness.”</p><h4>Understanding Categorical Picking and Its Limitations</h4><p>In many models, we often need to make a discrete choice between categories based on assigned probabilities. For example, imagine we have three options:</p><ul><li>Option A: 0.2 (20% probability)</li><li>Option B: 0.5 (50% probability)</li><li>Option C: 0.3 (30% probability)</li></ul><p>Using <strong>categorical picking</strong>, we select one of these options based on their probabilities. If we sample many times, Option B would be chosen 50% of the time, Option C 30% of the time, and Option A 20% of the time. However, this type of selection is <strong>non-differentiable</strong>. When a categorical choice is made, it produces a hard, discrete output like [0, 1, 0] for Option B, which cannot be smoothly adjusted by gradients.</p><p>Gradient-based optimization, which drives neural network training, relies on small, continuous changes to parameters to minimize a loss function. Since there’s no way to compute meaningful gradients through a discrete sampling process, <strong>categorical picking cannot be used directly in gradient-based optimization</strong>.</p><h4>Introducing Gumbel-Softmax Picking for Differentiability</h4><p>The <strong>Gumbel-Softmax</strong> process solves this issue by creating a “soft” version of categorical picking that remains differentiable. Here’s how it works:</p><p>The Gumbel-Softmax enables differentiable sampling from a categorical distribution, adding a controlled randomness to the output through Gumbel noise (gi). The temperature parameter (τ) controls the smoothness of the output — lower values of τ make the output closer to discrete values, while higher values produce softer, more probabilistic outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*WIXaR6EvSMwJ703MRzZxnQ.png" /></figure><p><strong>Gumbel Noise</strong>: First, random noise from the Gumbel distribution is added to each option’s log-probability. This noise allows a controlled randomness in the selection.</p><p><strong>Temperature Parameter (τ)</strong>: A temperature parameter τ is then applied using a Softmax function. This parameter controls the output’s smoothness:</p><ul><li>Lower values of τ(e.g., 0.1) make the output more discrete, approximating a one-hot vector like [0, 1, 0].</li><li>Higher values of τ (e.g., 1.0) make the output softer and more probabilistic, such as [0.2, 0.5, 0.3].</li></ul><p>For instance, suppose we have the same probabilities as above:</p><ul><li>With <strong>low temperature</strong> τ=0.1: the output might be [0.01, 0.97, 0.02], almost like picking Option B directly.</li><li>With <strong>high temperature</strong> τ=1.0: the output could be [0.2, 0.5, 0.3], more like a &quot;soft&quot; version of the categorical probabilities.</li></ul><p>Since the Gumbel-Softmax outputs a continuous approximation of a categorical choice, <strong>gradients can flow through the selection process</strong>, allowing backpropagation to work.</p><h4>gumble softmax vs softmax :</h4><p><strong>This flexibility is important for DT+GNN, especially if the model needs to adjust how closely it follows discrete decision tree behavior as training progresses</strong>. Normal softmax lacks this control, as it always produces a “soft” probability distribution that can’t become one-hot.</p><h3>3. Retaining Theoretical Expressiveness</h3><p>Despite categorical constraints, Diff-DT+GNN retains expressiveness by enabling nodes to “count” their neighbors’ states using messages limited to a small number of bits. This allows nodes to represent their environment with a manageable level of complexity while remaining interpretable.</p><p>Imagine each node as a person in a neighborhood trying to see the trend in color preferences (e.g., blue, green, or red) around them.</p><ul><li><strong>Limited Information per Message</strong>: Each person can only send a simple, 2-bit message about their preference (e.g., Blue = 00, Green = 01, Red = 10).</li><li><strong>Counting Preferences</strong>: Each person counts how many neighbors like each color. For example, a person with five neighbors might see 2 like blue color, 2 like green color, and 1 likes red color.</li><li><strong>Simple View of the Neighborhood</strong>: This summary (e.g., 2 blue, 2 green, 1 red) gives each person an overview of the neighborhood, helping them decide to follow the majority or blend in.</li></ul><h3>4. Distilling the DT+GNN: From MLPs to Decision Trees</h3><p>After training the model, Multi Layer Perceptrons (MLPs) in each Diff-DT+GNN layer are replaced by decision trees. This transformation involves gathering input-output pairs from the training data and using them to train decision trees that fully replace MLPs with interpretable structures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/685/1*J0yUVoh54_7RmjjR_XcnTA.png" /><figcaption><a href="https://arxiv.org/abs/2205.13234">source</a></figcaption></figure><p><strong>Pairwise Delta Features: </strong>Pairwise delta features Δ are introduced to enhance decision-making by allowing comparisons of node features. This feature enables the decision trees to branch more effectively by understanding relative differences, leading to both improved interpretability and accuracy.</p><p>This mechanism improves both interpretability and predictive accuracy.</p><p><strong>the delta features allow the decision tree branches to capture more nuanced relationships between nodes</strong>. This is important because it means the model doesn’t just look at isolated node features but also considers relational differences, which can improve both interpretability and prediction quality.</p><p><strong>Example of Pairwise Delta Features</strong></p><p>Imagine a social network where each node represents a person, and each person has a <strong>“popularity score”</strong> as one of their features. Nodes are connected to their friends, and we want to predict whether each person is likely to attend a local event. Using pairwise delta features, we can incorporate comparisons between a person’s popularity score and their friends’ scores.</p><p>For example:</p><ul><li><strong>Person A</strong> has a popularity score of 30.</li><li><strong>Person B</strong> has a popularity score of 50 (neighbor of Person A).</li><li><strong>Person C</strong> has a popularity score of 20 (neighbor of Person A).</li></ul><p>With pairwise delta features, we can compute the differences in popularity between Person A and each of their neighbors:</p><ul><li><strong>Δ(A — B) = 30–50 = -20</strong></li><li><strong>Δ(A — C) = 30–20 = 10</strong></li></ul><p>These delta features (e.g., -20 and +10) provide information on whether Person A is more or less popular than their friends. A decision tree can use these deltas to branch on questions like:</p><ul><li>“Is Person A’s popularity significantly lower than their friends?” (if Δ &lt; -15)</li><li>“Is Person A’s popularity higher than at least one friend?” (if Δ &gt; 0)</li></ul><p>This comparison-based branching enables more meaningful decisions, as it considers the person’s popularity <strong>relative to their neighbors</strong>, not just the absolute score.</p><h4><strong>Why Pairwise Delta Features Enhance Interpretability and Accuracy</strong></h4><ol><li><strong>Improved Interpretability</strong>: Pairwise deltas allow us to ask interpretable questions like “Is this node higher or lower than its neighbors on this feature?” which is easy to understand and analyze.</li><li><strong>Enhanced Accuracy</strong>: By using relative comparisons, decision trees can better capture patterns that depend on a node’s relationships within its neighborhood, rather than treating each node’s features in isolation.</li></ol><h3>5. Pruning Techniques for DT+GNN</h3><p>To make the decision trees within DT+GNN more manageable, <strong>pruning</strong> techniques are applied. Pruning helps reduce tree size without sacrificing accuracy, enhancing interpretability. DT+GNN employs two types of pruning:</p><ul><li><strong>Lossless Pruning:</strong> This technique removes nodes that can be replaced by leaves without affecting the model’s accuracy.</li><li><strong>Lossy Pruning:</strong> This more aggressive approach prunes nodes at the cost of a minimal decrease in accuracy. The goal is to make trees smaller and easier to interpret by sacrificing a small amount of precision.</li></ul><p>These pruning steps are performed with both training and validation data to avoid overfitting and maintain generalizability.</p><h3>Process Summary</h3><p>Consider a graph where nodes represent individuals with features like age, occupation, and location, and edges represent social connections. Here’s how DT+GNN would operate:</p><ol><li><strong>Initial Setup:</strong></li></ol><ul><li><strong>Node Features</strong>: Node features (e.g., age, occupation) can be continuous or categorical depending on the dataset. However, the target variable y for classification should be categorical to be interpreted by the decision trees.</li><li><strong>Edge Features</strong>: Edges have features representing connection strength or type.</li></ul><p><strong>2. Message Passing</strong>:</p><ul><li>Nodes exchange categorical information about their states and connections, representing types of interactions (e.g., “Young Professional” or “Student”) and bond strength (“strong bond” or “weak bond”).</li></ul><p><strong>3. State Update</strong>:</p><ul><li>The Diff-DT+GNN layer applies the Gumbel-Softmax function to ensure differentiability, using a temperature parameter to control the transition from probabilistic outputs to discrete states.</li></ul><p><strong>4. MLP Distillation</strong>:</p><ul><li>Once trained, each MLP layer is replaced with a decision tree trained on input-output pairs, which replicates MLP functions with a transparent structure.</li></ul><p><strong>5. Pruning</strong>:</p><ul><li>Decision trees are pruned using both lossless and lossy pruning, making the trees easier to interpret without compromising accuracy.</li></ul><h3>Generating Explanations with DT+GNN</h3><p>DT+GNN generates explanations at a node level, enabling detailed insights into its predictions. This process resembles the forward pass of a GNN.</p><h3>Explanation Propagation through Decision Trees</h3><p>In each layer l , explanations for a node’s state are determined by the importance of its features. DT+GNN utilizes <strong>Shapley values</strong> from the TreeShap algorithm to capture feature importance. The feature importance values are represented as follows:</p><ul><li>I(l,s): Importance of state features</li><li>I(l,m): Importance of message features</li><li>I(l,Δ)​: Importance of delta features</li></ul><p>Each of these values is propagated with different equations:</p><ol><li><strong>State Features Propagation:</strong> This involves a weighted sum based on feature importance</li><li><strong>Message Features Propagation:</strong> This depends on the neighboring nodes:</li><li><strong>Delta Features Propagation:</strong> This compares the states of neighbors</li></ol><p>By summing these components across layers, DT+GNN provides a clear explanation of each node’s prediction.</p><h3>Experiments and Results</h3><p>DT+GNN was evaluated across several benchmarks, including both real-world and synthetic datasets designed for explainability. Key findings include:</p><ul><li><strong>Accuracy:</strong> DT+GNN achieved comparable accuracy to traditional GNNs, demonstrating that interpretability does not come at the cost of performance.</li><li><strong>Tree Pruning Effectiveness:</strong> Both lossless and lossy pruning techniques effectively reduced tree size while maintaining accuracy.</li><li><strong>Explanatory Insights:</strong> DT+GNN could pinpoint nodes and edges crucial to predictions, providing valuable insights into the datasets.</li></ul><h3>Conclusion and Future Directions</h3><p>DT+GNN offers a breakthrough in making GNNs interpretable at every stage, combining differentiable layers, decision trees, and advanced pruning techniques. Future research might explore DT+GNN’s application to datasets with high-dimensional node features, using methods like PCA or specialized MLPs to manage complexity.</p><p><strong>Impact</strong></p><p>DT+GNN could transform how we understand GNN models, especially in sensitive fields where model transparency is crucial. By providing clear explanations for predictions, DT+GNN paves the way for safer and more reliable applications of GNNs in real-world scenarios.</p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16dcb003a9ae" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/part2-dt-gnn-a-fully-explainable-graph-neural-network-using-decision-trees-16dcb003a9ae">PART2 : DT+GNN: A Fully Explainable Graph Neural Network using Decision Trees</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PART1 : DT+GNN: Making AI Decisions Transparent with Decision Trees and Graph Neural Networks]]></title>
            <link>https://medium.com/ellora-ai/part1-dt-gnn-making-ai-decisions-transparent-with-decision-trees-and-graph-neural-networks-8095befbcaf0?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/8095befbcaf0</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Mon, 13 Jan 2025 18:10:45 GMT</pubDate>
            <atom:updated>2025-01-15T01:35:41.134Z</atom:updated>
            <content:encoded><![CDATA[<h3>PART1 : DT+GNN: Making AI Decisions Transparent with Decision Trees and Graph Neural Networks</h3><p><strong>Writers:</strong></p><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><strong><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></strong></p><p>In fields that rely heavily on graph data., such as social networks, finance, and healthcare., AI models are critical for node classification, edge prediction, and graph analysis.</p><p>However, as powerful as Graph Neural Networks (GNNs) are, their decision-making processes are often opaque and difficult to interpret from a human perspective. GNNs analyze complex networks by leveraging deep learning layers, but the resulting predictions usually lack clarity, making it hard to trace back to specific attributes or relationships within the data.</p><p>In many sectors, where transparency and compliance are essential, it’s crucial to understand the reasons behind each AI-driven decision.</p><p>This is where DT+GNN [5] comes in. By combining the pattern recognition capabilities of GNNs with the transparency of decision trees, DT+GNN makes it possible to interpret complex predictions in an understandable, rule-based manner [5].</p><p>Throughout the series of blogs, we’ll walk through a conceptual example of using DT+GNN to analyze a simplified financial assets graph and classify loans based on their network characteristics.</p><h3>A Real-World Scenario: Financial assets and liabilities graph</h3><p>Imagine we have a graph to model a very simplified loan application with the following structure:</p><ul><li>Nodes: Each node represents a financial asset, for example a person’s overall financial snapshot, debts, assets, liabilities &amp; credits</li><li>Node Features: {Debt To Income Ratio: DTI, Loan To Value Ratio : LTV, Credit Score, Net Income, Net Expenses}.</li><li>Edges: An edge between two nodes represents a connection.</li><li>Edge Features: {interaction strength, connection type for example “owns” or “source of funds”}.</li><li>Target: Classify personal credit risk based on background data</li></ul><p>Our goal is to use DT+GNN to interpret and understand each personal credit risk classification based on key parameters.</p><h3>How DT+GNN Works</h3><p>DT+GNN integrates two complementary approaches:</p><ul><li>Graph Neural Network (GNN): The GNN handles the graph-structured data, analyzing nodes and edges to detect patterns and relationships across the network. This allows the model to identify unique characteristics of each node and its connections.</li><li>Decision Tree (DT): After processing the data through the GNN, the decision tree translates these complex embeddings into clear, rule-based paths. While GNNs excel at recognizing patterns, the DT makes these results interpretable by presenting understandable decision rules.</li></ul><p>Let’s walk through the step-by-step process of how DT+GNN operates:</p><h3>Step 1: Embedding Creation with GNN</h3><p>The GNN processes the graph structure, generating embeddings for each node. These embeddings are representations that capture the unique characteristics of nodes and their interactions within the graph.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/746/1*o2408XBm0nnVIO_XLVpfEw.gif" /><figcaption><a href="https://www.google.com/url?sa=i&amp;url=https%3A%2F%2Fwww.linkedin.com%2Fpulse%2Funderstanding-node-embeddings-graphs-part-5-x-my-notes-ajay-taneja-1f0jf&amp;psig=AOvVaw0e6zpwCFhHYy2eVttS_TNJ&amp;ust=1733917552241000&amp;source=images&amp;cd=vfe&amp;opi=89978449&amp;ved=0CBQQjRxqFwoTCNCFhsKQnYoDFQAAAAAdAAAAABAJ">source</a></figcaption></figure><h3>Step 2: Diff-DT+GNN Layer: Categorical State Representation</h3><p>Following the GNN’s message-passing process, the Diff-DT+GNN layer introduces a unique interpretability layer. This layer transforms continuous embeddings into discrete categorical states. Instead of using complex numeric values, the model represents nodes using a limited set of categorical labels that capture key network characteristics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*OWQWWS3JqU_WS-O2UcfJKA.png" /><figcaption><a href="https://arxiv.org/pdf/2205.13234">source</a></figcaption></figure><h3>Step 3: Decision Tree Translation for Interpretability</h3><p>The final step involves translating these embeddings and categorical states into a decision tree structure. The decision tree enables stakeholders to follow the sequence of decisions that lead to each classification. By using categorical labels derived from GNN embeddings, the model presents interpretable, rule-based paths for each prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*fmwKQCYz3W2acmlk07MYOA.png" /><figcaption><a href="https://arxiv.org/pdf/2205.13234">source</a></figcaption></figure><h3>Benefits of DT+GNN</h3><p>By translating complex GNN outputs into interpretable decisions, DT+GNN provides significant benefits:</p><ul><li>Trust and Transparency: The decision tree output enables teams to understand the reasoning behind classifications.</li><li>Improved Compliance: Interpretable AI models assist with documentation and accountability.</li><li>Aligned Decision-Making: Clear classifications allow teams to understand AI insights more deeply.</li><li>Targeted Analysis: By understanding network behavior, organizations can gain more nuanced insights.</li></ul><h3>Conclusion</h3><p>The advancement of AI in graph-based applications depends on both predictive power and transparency. DT+GNN represents a significant step forward, combining the interpretability of decision trees with the pattern recognition capabilities of GNNs. This approach allows businesses to deploy AI models that are both accurate and explainable, leading to more confident and transparent decision-making.</p><h3>Data</h3><p>The data used in this article is a very simplified anonymized publicly available loan data retaining only key columns for risk score/credit classification</p><h3>References</h3><p>We scoured through dozens of peer reviewed research papers from Universities across the world published over past decade. And poured over the latest on XAI. Our research took us into areas of neuro biology like GNN for automatic brain tumour segmentation! The paper references of what we studied towards developing our model is listed here. And also the paper on which we developed our XAI model.</p><p>[1] Scalable Explanation of Inferences on Large Graphs<a href="https://arxiv.org/abs/1908.06482">[paper]</a></p><p>[2] XGNN: Towards Model-Level Explanations of Graph Neural Networks<strong> </strong><a href="https://dl.acm.org/doi/10.1145/3394486.3403085">[paper]</a></p><p>[3] PGM-Explainer: Probabilistic Graphical Model Explanations for Graph Neural Networks <a href="https://arxiv.org/pdf/2010.05788.pdf">[paper]</a></p><p>[4] A New Concept for Explaining Graph Neural Networks <a href="http://ceur-ws.org/Vol-2986/paper1.pdf">[paper]</a></p><p>[5] DT+GNN: A Fully Explainable Graph Neural Network using Decision Trees [<a href="https://arxiv.org/pdf/2205.13234">Paper</a>]</p><p>[6] GraphSVX: Shapley Value Explanations for Graph Neural Networks <a href="https://arxiv.org/abs/2104.10482">[paper]</a></p><p>[7] On Explainability of Graph Neural Networks via Subgraph Explorations<a href="https://arxiv.org/abs/2102.05152">[paper]</a></p><p>[8] GANExplainer: GAN-based Graph Neural Networks Explainer <a href="https://arxiv.org/abs/2301.00012">[paper]</a></p><p>[9] Understanding Node Embeddings in Graph [<a href="https://www.linkedin.com/pulse/understanding-node-embeddings-graphs-part-5-x-my-notes-ajay-taneja-1f0jf/">https://www.linkedin.com/pulse/understanding-node-embeddings-graphs-part-5-x-my-notes-ajay-taneja-1f0jf/</a>]</p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8095befbcaf0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/part1-dt-gnn-making-ai-decisions-transparent-with-decision-trees-and-graph-neural-networks-8095befbcaf0">PART1 : DT+GNN: Making AI Decisions Transparent with Decision Trees and Graph Neural Networks</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GNN-XAI: Putting it All Together (Part 4)]]></title>
            <link>https://medium.com/ellora-ai/gnn-xai-putting-it-all-together-part-4-d88b9ba02c1b?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/d88b9ba02c1b</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Mon, 13 Jan 2025 18:10:10 GMT</pubDate>
            <atom:updated>2025-01-15T02:02:39.182Z</atom:updated>
            <content:encoded><![CDATA[<h3>Writers:</h3><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p>Welcome to the final part of our series on Explainable AI (XAI) for Graph Neural Networks (GNNs). We hope you’ve enjoyed the <a href="https://medium.com/ellora-ai/gnn-xai-evaluating-explainability-in-graph-neural-networks-metrics-methods-and-best-practices-9de222916116">journey</a> and found each part insightful and helpful. If you have any questions or thoughts, don’t hesitate to comment. Your feedback is invaluable and might even inspire a new series of blogs — who knows? 😉</p><p>In the world of artificial intelligence, <strong>making complex decisions transparent is no longer a luxury., it’s a necessity</strong>. Nowhere is this more critical than in financial services, where understanding the reasoning behind decisions for example.,<strong> loan approvals or rejections can mean the difference between fair lending and algorithmic bias.</strong></p><p>Graph Neural Networks (GNNs) combined with Explainable AI (XAI) offer a powerful approach to creating more interpretable machine learning models.</p><p>In this final part, we’ll walk through building a GNN model for loan classification that doesn’t just predict risk but also explains its reasoning.</p><h3>Building an Explainable Loan Risk Classification Model</h3><h4>Step 1: Data Preprocessing and Augmentation</h4><p>Begin by cleaning and preparing your dataset. Perform data augmentation if necessary to enhance the quality and diversity of the data.</p><h4>Step 2: Split the Data</h4><p>Divide your dataset into training and testing sets to ensure robust evaluation of your model.</p><h4>Step 3: Transform Data into Graph Representation</h4><p>When transforming data into a graph representation, carefully select node and edge features.</p><p>For our loan classification example, features such as Channel, Loan Purpose, LTV, DTI, and Credit Score are used</p><pre>def construct_graph(data, edge_feature_columns, node_feature_columns):<br>    # Convert node features to tensor<br>    node_features = torch.tensor(data[node_feature_columns].values, dtype=torch.float)<br>    <br>    # Generate all possible pairs of nodes<br>    num_nodes = len(data)<br>    possible_edges = torch.combinations(torch.arange(num_nodes), r=2)<br>    <br>    # Filter edges based on matching values in edge features<br>    filtered_edges = []<br>    edge_feat_list = []<br>    for i, j in possible_edges:<br>            i, j = i.item(), j.item()<br>            # Create edges where certain features match<br>            if all(data[col].iloc[i] == data[col].iloc[j] for col in edge_feature_columns):<br>                filtered_edges.append((i, j))<br>                edge_features_for_pair = [data[col].iloc[i] for col in edge_feature_columns]<br>                edge_feat_list.append(edge_features_for_pair)<br>        <br>        # Convert edges to tensor format<br>        edge_index = torch.tensor(filtered_edges).t().contiguous()<br>        edge_features = torch.tensor(edge_feat_list, dtype=torch.long)<br>    return node_features, edge_index, edge_features<br><br># Usage<br>node_features, edge_index, edge_features = construct_graph(<br>    train_data,<br>    edge_feature_columns=[&#39;Channel&#39;, &#39;Loan Purpose&#39;],<br>    node_feature_columns=[&#39;LTV&#39;, &#39;DTI&#39;, &#39;Credit Score&#39;]<br>)</pre><h4>Step 4: Train a GNN Model</h4><p>We’ll use a Graph Convolutional Network (GCN) for loan risk classification. Feel free to experiment with other models to find the best fit.</p><pre># Create PyTorch Geometric Data object<br>graph_train_data = Data(<br>    x=node_features,<br>    edge_index=edge_index,<br>    edge_attr=edge_features,<br>    y=torch.tensor(train_data[&#39;loan_risk&#39;].values, dtype=torch.long)<br>)<br><br>class GCNLoanRisk(torch.nn.Module):<br>    def __init__(self, num_node_features, num_edge_features, hidden_channels):<br>        super().__init__()<br>        self.conv1 = GCNConv(num_node_features, hidden_channels, num_edge_features)<br>        self.conv2 = GCNConv(hidden_channels, hidden_channels, num_edge_features)<br>        self.output = torch.nn.Linear(hidden_channels, 2)<br>        self.dropout = torch.nn.Dropout(p=0.5)<br><br>    def forward(self, x, edge_index, edge_attr):<br>        x = self.conv1(x, edge_index, edge_attr)<br>        x = F.relu(x)<br>        x = self.dropout(x)<br>        x = self.conv2(x, edge_index, edge_attr)<br>        x = F.relu(x)<br>        x = self.dropout(x)<br>        x = self.output(x)<br>        return F.log_softmax(x, dim=1)<br></pre><h4>Step 5: Explain Predictions with GNNExplainer</h4><p>Understanding a GNN model’s decision-making process is crucial. The GNNExplainer module for example helps identify which features and subgraphs contribute to a prediction.</p><p>Here is an overview of its usage:</p><pre># Initialize the model and explainer<br>model = GCNLoanRisk(num_node_features=3, num_edge_features=2, hidden_channels=64)<br>explainer = Explainer(<br>    model=model,<br>    algorithm=GNNExplainer(epochs=200, return_type=&#39;log_probs&#39;),<br>    explanation_type=&#39;model&#39;,<br>    node_mask_type=&#39;attributes&#39;,<br>    edge_mask_type=&#39;importance&#39;<br>)</pre><pre># Select a specific node to explain<br>node_to_explain = 0  # First node in the graph</pre><pre># Generate explanation<br>explanation = explainer(<br>    graph_train_data.x,  # Node features<br>    graph_train_data.edge_index,  # Graph edges<br>    edge_attr=graph_train_data.edge_attr,  # Edge attributes<br>    target=node_to_explain  # Node to explain<br>)</pre><pre># Visualize the most important features<br>print(&quot;Top Node Features:&quot;, explanation.node_mask.topk(k=2))<br>print(&quot;Top Edge Features:&quot;, explanation.edge_mask.topk(k=2))</pre><p>Here’s a very simplified example of the output from GNNExplainer: It tells you how much each node feature impacts the prediction, with values normalized to a scale of 1. In this case, none of the node features individually had a significant impact on the prediction. Though Credit Score, Debt-to-Income Ratio (DTI) and Loan To Value Ratio (LTV) showed up as potentials!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*ZmOWiBmgoek6dXFwm_zS-g.png" /></figure><p><strong>Instead, the graph structure played a dominant role, as illustrated by the subgraph below, which had a substantial influence on the prediction.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/873/1*GGEP8hGxbJufhD5hrUjdgQ.png" /></figure><p>One interpretation can be that Credit Score, LTV and DTI play almost equal with a skew towards DTI’s role in the classification of a loan risk.</p><p>This makes sense, if an home loan applicant’s Debt To Income (DTI) is high, even if the applicant has very good credit score and Loan-to-Value (LTV) ratios, the loan classification is risky compared to the background loans.</p><p>This leads to another challenge, the challenge of interpretability.</p><h4>Step 6: Evaluate with Fidelity-Plus</h4><p>Validating the robustness of explanations is essential. Fidelity-Plus measures how much the removal of key subgraphs affects predictions.</p><pre>def calculate_fidelity_plus(model, original_graph, explanation):<br>    # Remove the most important subgraph identified by the explainer<br>    reduced_graph = remove_subgraph(original_graph, explanation)<br>    <br>    # Compare predictions<br>    original_pred = model(original_graph.x, original_graph.edge_index)<br>    reduced_pred = model(reduced_graph.x, reduced_graph.edge_index)<br>    <br>    # Calculate fidelity<br>    fidelity = torch.mean(torch.abs(original_pred - reduced_pred))<br>    <br>    return fidelity</pre><pre># Calculate fidelity<br>fidelity_score = calculate_fidelity_plus(model, graph_train_data, explanation)<br>print(f&quot;Fidelity-Plus Score: {fidelity_score}&quot;)</pre><pre># Interpretation guide<br>if fidelity_score &lt; 0.1:  # Low threshold, adjust based on your specific use case<br>    print(&quot;Explanation is robust: Removing key subgraph minimally affects predictions&quot;)<br>else:<br>    print(&quot;Consider refining the model or explanation method&quot;)</pre><h3>Practical Implications</h3><p>Graph Neural Networks with Explainable AI offer a powerful approach to understanding complex decision-making processes. In loan classification, this becomes crucial, not just predicting risk but understanding why a particular loan might be considered high-risk.</p><h3>Challenges and Considerations</h3><ul><li>Developing robust explanation techniques</li><li>Balancing model complexity with interpretability</li><li>Ensuring explanations are technically accurate and understandable</li></ul><h3>Future Directions</h3><ul><li>Improving explanation algorithms</li><li><strong>Creating more intuitive visualization techniques</strong></li><li>Developing domain-specific XAI approaches</li></ul><h3>Conclusion</h3><p>As AI continues to play a larger role in critical decisions, techniques like GNN-XAI will become increasingly important in building trust and transparency in machine learning models. By combining advanced graph neural networks with explainable AI, we can create more interpretable and reliable predictive systems.</p><p>Happy exploring, and may your AI models always be as transparent as they are powerful!</p><h3><strong>References</strong></h3><ol><li>GNNExplainer: Generating Explanations for Graph Neural Networks [<a href="https://arxiv.org/abs/1903.03894">https://arxiv.org/abs/1903.03894</a>]</li></ol><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d88b9ba02c1b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/gnn-xai-putting-it-all-together-part-4-d88b9ba02c1b">GNN-XAI: Putting it All Together (Part 4)</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GNN-XAI: Evaluating Explainability in Graph Neural Networks: Metrics, Methods, and Best Practices…]]></title>
            <link>https://medium.com/ellora-ai/gnn-xai-evaluating-explainability-in-graph-neural-networks-metrics-methods-and-best-practices-9de222916116?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/9de222916116</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Mon, 13 Jan 2025 18:09:37 GMT</pubDate>
            <atom:updated>2025-01-15T01:53:51.689Z</atom:updated>
            <content:encoded><![CDATA[<h3>GNN-XAI: Evaluating Explainability in Graph Neural Networks: Metrics, Methods, and Best Practices (Part 3)</h3><h3><strong>Writers:</strong></h3><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p>In our previous articles <a href="https://medium.com/ellora-ai/gnn-xai-explainability-in-graph-neural-networks-understanding-the-black-box-part-1-2f8a15b89001">here</a> and <a href="https://medium.com/ellora-ai/gnn-xai-demystifying-graph-neural-networks-a-deep-dive-into-gnnexplainer-part-2-9f6cb78b2581">here</a>, we explored the fundamentals of explainability and dove deep into GNNExplainer. Now, we face a crucial question: <strong>How do we know if our explanations are good?</strong> This part of our series tackles the challenging task of evaluating explanation quality in graph neural networks.</p><h3>The Challenge of Evaluation</h3><p>Evaluating explanations is inherently difficult because:</p><ul><li>Ground truth explanations rarely exist in real-world scenarios</li><li>Multiple valid explanations might exist for a single prediction</li><li>The quality of an explanation often depends on the specific use case</li></ul><p>Let’s explore how we can systematically approach these challenges.</p><h3>Evaluation Frameworks</h3><h3>1. Synthetic Evaluation</h3><p>When working with synthetic datasets (artificially created data that mimics real-world data but is generated under controlled conditions), we can create controlled environments where we know the ground truth explanations.</p><h3>Example: Motif Detection</h3><p><strong>Dataset Construction:</strong></p><ol><li>Generate base graphs (e.g., Barabási-Albert networks or tree graphs)</li><li>Insert specific motifs (recurring patterns in a graph) (e.g., houses, cycles)</li><li>Label nodes based on motif membership</li><li>Train GNN to detect these motifs</li><li>Evaluate if explanations recover the known motifs</li></ol><p><strong>Metrics for Synthetic Evaluation:</strong></p><ul><li>Accuracy = (Correctly identified motif elements) / (Total motif elements)</li><li>AUC = Area under ROC curve for motif recovery</li></ul><p><em>Note: The ROC (Receiver Operating Characteristic) curve shows how well a model can distinguish between different classes. It plots the True Positive Rate against the False Positive Rate at various threshold settings. A higher Area Under the Curve (AUC) indicates better classification performance.</em></p><h3>2. Real-World Evaluation</h3><p>For real-world datasets where ground truth isn’t available, we need to evaluate explanations based on their impact on model predictions.</p><h3>Core Evaluation Metrics</h3><h3>1. Fidelity Metrics</h3><p>Fidelity measures how well the explanation captures the model’s behavior. It has two key components:</p><p>a) <strong>Fidelity-Plus (Necessity)</strong> Fidelity⁺ = P(f(x) ≠ f(x\S))</p><ul><li>f(x) is the original prediction</li><li>f(x\S) is the prediction without the explanation components</li><li>S is the explanation subset</li></ul><p>This measures whether removing the explained components changes the prediction.</p><p>b) <strong>Fidelity-Minus (Sufficiency)</strong> Fidelity⁻ = P(f(x) = f(xS))</p><ul><li>xS contains only the components in the explanation</li></ul><p>This measures whether the explained components alone are sufficient to maintain the prediction.</p><h3>2. Stability Metrics</h3><p>Stability measures how consistent explanations are across different runs or similar instances:</p><p>“Different runs” means running the same explanation method multiple times, potentially with:</p><ul><li>Different random initializations</li><li>Slightly perturbed inputs</li><li>Varying model parameters</li></ul><p>The goal is to check if the explanations remain consistent. If explanations dramatically change for very similar inputs, it suggests the explanation method is unstable and unreliable.</p><p>Stability = 1 — avg_distance(E1, E2)</p><ul><li>E1, E2 are explanations from different runs</li><li>avg_distance measures the similarity between explanations</li></ul><h3>3. Complexity Metrics</h3><p>Complexity = w1 * |edges| + w2 * |features|</p><ul><li>|edges| is the number of edges in the explanation</li><li>|features| is the number of features used</li><li>w1, w2 are weighting parameters</li></ul><h3>Comprehensive Evaluation Framework</h3><h3>1. Combined Fidelity Score</h3><p>To balance necessity and sufficiency: Fidelity_combined = 2 * (Fidelity⁺ * Fidelity⁻) / (Fidelity⁺ + Fidelity⁻)</p><h3>2. Quality-Complexity Trade-off</h3><p>Final_score = Fidelity_combined * exp(-λ * Complexity)</p><ul><li>λ controls the penalty for complex explanations</li></ul><h3>Practical Evaluation Strategies</h3><h3>1. Cross-Method Comparison</h3><p>Compare different explanation methods across multiple dimensions:</p><p><strong>Methods to compare:</strong></p><ul><li>Gradient-based explanations</li><li>Attention-based methods</li><li>Perturbation-based approaches</li><li>GNNExplainer and variants</li></ul><h3>2. Task-Specific Evaluation</h3><p><strong>Node Classification:</strong> When analyzing individual nodes in a graph, we want explanations that:</p><ul><li>Understand local neighborhood influences (k-hop neighborhood)</li><li>Accurately attribute features’ contributions</li><li>Show how nodes are connected in meaningful ways</li></ul><p><strong>Graph Classification:</strong> When classifying entire graphs (like in molecular or social network analysis), we want explanations that:</p><ul><li>Preserve overall graph structure</li><li>Identify key substructures (motifs)</li><li>Ensure meaningfulness (e.g., chemical validity for molecular graphs)</li></ul><h3>Advanced Evaluation Considerations</h3><h3>1. Counterfactual Evaluation</h3><p>A counterfactual explanation shows what would need to change in the input to get a different prediction.</p><p>Assess the quality of counterfactual explanations:</p><p>Counterfactual_score = α * Validity + β * Proximity + γ * Sparsity</p><ul><li>Validity: Does changing the input actually alter the model’s prediction?</li><li>Proximity: How similar is this new input to the original?</li><li>Sparsity: How few changes were needed to modify the prediction?</li></ul><p><strong>Example:</strong></p><ul><li>Original Input: A molecular graph classified as “not toxic”</li><li>Counterfactual Explanation: Removing or changing specific atoms makes the molecule classified as “toxic”</li></ul><h3>2. Domain-Specific Metrics</h3><p><strong>Chemical Graphs:</strong></p><ul><li>Chemical Validity: Ensuring the explained molecular structure makes chemical sense</li><li>Functional Groups: Identifying key chemical substructures</li><li>Chemical Intuition: Explanations align with known chemical principles</li></ul><p><strong>Social Networks:</strong></p><ul><li>Community Preservation: Explanations respect existing social group structures</li><li>Path Importance: Understanding how connections influence predictions</li><li>Temporal Consistency: Explanations remain stable over time</li></ul><h3>Best Practices for Evaluation</h3><h4>Multiple Initialization Testing</h4><ul><li>Run explanations with different random seeds</li><li>Calculate stability across runs</li><li>Report confidence intervals</li></ul><h4>Ablation Studies</h4><ul><li>Test impact of different components</li><li>Vary hyperparameters systematically</li><li>Document sensitivity to parameters</li></ul><h4>Documentation Requirements</h4><ul><li>Dataset characteristics</li><li>Model architecture details</li><li>Explanation method parameters</li><li>Evaluation metric definitions</li><li>Runtime considerations</li></ul><h3>Future Directions</h3><h4>1. Automated Evaluation Frameworks</h4><ul><li>Development of standardized benchmarks</li><li>Integration of multiple metrics</li><li>Automated parameter tuning</li></ul><h4>2. Human-in-the-Loop Evaluation</h4><ul><li>Expert feedback integration</li><li>User studies for explanation utility</li><li>Domain-specific validation</li></ul><h4>3. Dynamic Graph Evaluation</h4><ul><li>Temporal consistency metrics</li><li>Evolution of explanations over time</li><li>Stability across time steps</li></ul><h3>Conclusion</h3><p>Evaluating explainability in Graph Neural Networks remains a complex challenge that requires a multi-faceted approach. By combining quantitative metrics with domain-specific considerations and human expertise, we can work toward more reliable and meaningful explanations.</p><p><strong>Key Takeaways:</strong></p><ul><li>Use multiple evaluation metrics</li><li>Consider task-specific requirements</li><li>Balance fidelity with complexity</li><li>Incorporate domain knowledge</li><li>Maintain consistent evaluation protocols</li></ul><p>As the field continues to evolve, these evaluation frameworks will become increasingly sophisticated, helping us build more trustworthy and interpretable graph neural networks.</p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9de222916116" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/gnn-xai-evaluating-explainability-in-graph-neural-networks-metrics-methods-and-best-practices-9de222916116">GNN-XAI: Evaluating Explainability in Graph Neural Networks: Metrics, Methods, and Best Practices…</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GNN-XAI: Demystifying Graph Neural Networks: A Deep Dive into GNNExplainer (Part 2)]]></title>
            <link>https://medium.com/ellora-ai/gnn-xai-demystifying-graph-neural-networks-a-deep-dive-into-gnnexplainer-part-2-9f6cb78b2581?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/9f6cb78b2581</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Wed, 08 Jan 2025 12:09:56 GMT</pubDate>
            <atom:updated>2025-01-08T17:54:31.341Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Writers:</strong></h3><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><strong><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></strong></p><p>In our previous article <a href="https://medium.com/ellora-ai/gnn-xai-explainability-in-graph-neural-networks-understanding-the-black-box-part-1-2f8a15b89001">GNN-XAI: Demystifying Graph Neural Networks (Part 1)</a>, we explored the fundamental concepts of explainability in graph neural networks and why it matters. Now, let’s dive deep into one of the most prominent methods for explaining GNN predictions: GNNExplainer.</p><p>This post will walk through the technical details, mathematical foundations, and practical implications of this powerful framework.</p><h3>Introduction to GNNExplainer</h3><p>GNNExplainer is a post-hoc explanation method designed to work with any graph neural network after it has been trained. Whether you’re working with financial transaction networks to detect fraud, molecular graphs for toxicity prediction, or social networks for user behavior analysis, GNNExplainer can help illuminate why your model makes specific predictions.</p><h3>The Framework</h3><h3>Key Components</h3><p>At its core, GNNExplainer aims to identify two crucial elements:</p><ul><li>The important subgraph structure that influences predictions</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BPoCnUzRUapvqQBk_hOoYA.jpeg" /><figcaption><a href="https://arxiv.org/pdf/1903.03894">source</a></figcaption></figure><ul><li>The relevant node features that contribute to the model’s decision</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/505/1*mojKuyaXFBOmUIVUJpu_FQ.png" /><figcaption><a href="https://www.researchgate.net/figure/For-vs-explanation-G-S-in-green-GnnExplainer-identifies-what-feature-dimensions-of-G_fig1_331670918">source</a></figcaption></figure><p>For any given prediction, we start with:</p><ul><li>A trained GNN model (φ)</li><li>A computation graph (the neighborhood used for prediction)</li><li>Node features</li><li>The prediction we want to explain</li></ul><p><strong>Note:</strong> In the original GNNExplainer paper, the primary focus is typically on node features and graph structure (adjacency matrix). While edge features can be important, most implementations concentrate on node features and graph topology.</p><p>This is because:</p><ul><li>Node and structural features often carry the most significant information</li><li>Edge features can be more challenging to incorporate into the explanation mechanism</li><li>The adjacency matrix provides a comprehensive view of graph connectivity</li></ul><h3>Mathematical Formulation</h3><p>Let’s break down the mathematical framework behind GNNExplainer:</p><p><strong>Prediction Probability:</strong> The GNN model outputs a probability distribution over classes: P(Y|G) = φ(A, X) where:</p><ul><li>A is the adjacency matrix</li><li>X represents node features</li><li>φ is the trained GNN model</li></ul><p><strong>Explanation Goal:</strong> The explainer seeks subsets of the computation graph (As) and features (Xs) that maximize mutual information with the prediction: <br>max MI(Y, (GS, XS)) = H(Y) — H(Y|GS, XS) <br>where:</p><ul><li>MI is mutual information</li><li>GS represents the selected subgraph structure</li><li>H(Y) is the entropy of the prediction</li><li>H(Y|GS, XS) is the conditional entropy given the selected subgraph and features</li></ul><p><strong>Note:</strong> Entropy measures uncertainty or randomness in a system:</p><ul><li>H(Y) represents the total uncertainty of the prediction</li><li>H(Y|GS, XS) represents the remaining uncertainty after selecting specific subgraph and features</li></ul><p>Lower conditional entropy means the selected features and subgraph provide more precise information</p><p><strong>Optimization Problem:</strong> To make this tractable (making the problem computationally manageable), GNNExplainer approximates the discrete selection of subgraphs with a continuous relaxation: M ∈ [0,1]^(N×N) where M is a mask over the adjacency matrix.</p><ul><li><strong>Discrete Selection:</strong> Originally, selecting exact subgraph components is complex</li><li><strong>Continuous Relaxation:</strong> Converting discrete selection to a continuous optimization problem</li><li>M ∈ [0,1]^(N×N) means creating a mask with continuous values between 0 and 1</li><li>This allows gradient-based optimization instead of combinatorial search</li></ul><h3>The Algorithm in Detail</h3><h3>Step 1: Masking Mechanism</h3><p>GNNExplainer introduces two types of masks:</p><ul><li><strong>Edge Mask (M):</strong> A learnable matrix with values between 0 and 1</li><li><strong>Feature Mask (F):</strong> A vector indicating important feature dimensions, with values typically between 0 and 1</li></ul><p>The masked computation graph is computed as: <br>A’ = A ⊙ M <br>X’ = X ⊙ F <br>where ⊙ represents element-wise multiplication.</p><h3>Step 2: Optimization Objective</h3><p>The full optimization objective includes several components: <br>L = MI(Y, (GS, XS)) — α||M||₁ — β||F||₁<br> where:</p><ul><li>The first term maximizes mutual information</li><li>The second term (α||M||₁) encourages sparsity by penalizing complex explanations in the graph structure</li><li>The third term (β||F||₁) promotes feature selection</li><li>α and β are hyperparameters controlling the trade-off</li><li>||M||₁ and ||F||₁ represent L1 norm (sum of absolute values)</li></ul><p>Prevents overfitting by selecting only the most crucial features/edges</p><h3>Step 3: Training Process</h3><ol><li>Initialize masks with random values</li><li>Compute masked input: (A’, X’)</li><li>Forward pass through the GNN (Gradient descent)</li><li>Compute loss using the objective function</li><li>Backpropagate and update masks (Minimize the loss function through backpropagation)</li><li>Repeat until convergence (typically 50–100 iterations)</li></ol><h3>Practical Implementation</h3><h3>Different Scenarios</h3><p><strong>Node Classification</strong></p><ul><li>Computation Graph: Node’s k-hop neighborhood</li><li>Goal: Identify important neighbors and features</li><li>Example: In a citation network, explaining why a paper is classified as “Machine Learning”</li></ul><p><strong>Link Prediction</strong></p><ul><li>Computation Graph: Combined neighborhoods of both nodes</li><li>Goal: Identify structures suggesting link formation</li><li>Example: Explaining why two users might become friends in a social network</li></ul><p><strong>Graph Classification</strong></p><ul><li>Computation Graph: Entire graph</li><li>Goal: Find important substructures and features</li><li>Example: Identifying toxic molecular substructures in chemical compounds</li></ul><h3>Implementation Considerations</h3><ol><li><strong>Base Value Selection:<br></strong> When dealing with feature importance, we need to consider: <br>Feature Impact = Actual Value — Base Value <br>where the base value represents the expected or default state of a feature.<br>Helps understand how features deviate from the baseline<br>Example: In toxicity prediction, base value might be the average feature value of non-toxic molecules</li><li><strong>Redundancy Management</strong></li></ol><p>Redundancy at multiple levels:</p><ul><li>Node-level: Repeated or highly correlated features</li><li>Graph-level: Multiple features providing similar information</li></ul><p>Goal: Identify and remove redundant features to create a more concise explanation<br>GNNExplainer employs:</p><ul><li>Regularization on mask size</li><li>Thresholding to remove weak connections</li><li>Constraints to maintain connected subgraphs</li></ul><p><strong>3. Size Control</strong></p><p>The size of explanations is controlled through:</p><p>size_constraint = λ * ||M||₁</p><p>where λ is a hyperparameter controlling explanation complexity.</p><h3>Practical Examples</h3><p>Molecular property prediction uses graph analysis to highlight key functional groups and atomic features for toxicity prediction, while <strong>financial network analysis focuses on identifying suspicious transaction patterns and user characteristics to detect potential fraud</strong>, with both approaches generating concise, interpretable explanations (around 10 key nodes) that reveal the critical factors underlying complex predictive models.</p><h3>Limitations and Considerations</h3><p><strong>Computational Overhead:</strong></p><ul><li>Requires optimization for each instance</li><li>Can be mitigated using deep explainers for batch processing</li></ul><p><strong>Multiple Valid Explanations:</strong></p><ul><li>Different initializations might lead to different explanations</li><li>Solution: Multiple runs with ensemble averaging</li></ul><p><strong>Feature Interaction:</strong></p><ul><li>Zero-valued features might still be important</li><li>Need for careful baseline selection</li></ul><h3>Looking Ahead</h3><p>While GNNExplainer provides a powerful framework for understanding GNN predictions, its effectiveness needs to be evaluated systematically. In our next article, we’ll explore various metrics and approaches for evaluating explanation quality, including fidelity, stability, and conciseness.</p><p>The power of GNNExplainer lies in its flexibility and principled approach to generating explanations. By providing insights into both structural and feature importance, it helps bridge the gap between complex graph neural networks and human understanding, making these powerful models more trustworthy and actionable in real-world applications.</p><h3>References :</h3><p><a href="https://arxiv.org/pdf/1903.03894">https://arxiv.org/pdf/1903.03894</a></p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f6cb78b2581" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/gnn-xai-demystifying-graph-neural-networks-a-deep-dive-into-gnnexplainer-part-2-9f6cb78b2581">GNN-XAI: Demystifying Graph Neural Networks: A Deep Dive into GNNExplainer (Part 2)</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GNN-XAI: Explainability in Graph Neural Networks: Understanding the Black Box (Part 1)]]></title>
            <link>https://medium.com/ellora-ai/gnn-xai-explainability-in-graph-neural-networks-understanding-the-black-box-part-1-2f8a15b89001?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/2f8a15b89001</guid>
            <dc:creator><![CDATA[yasmine karray]]></dc:creator>
            <pubDate>Wed, 08 Jan 2025 12:08:36 GMT</pubDate>
            <atom:updated>2025-01-08T17:48:09.750Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Writers:</strong></h3><p><a href="https://medium.com/u/25b22d7e4148"><strong>yasmine karray</strong></a><strong> &amp; </strong><a href="https://medium.com/u/23cc8ade9ab5"><strong>Dhaval Shah</strong></a><strong> for Ellora AI</strong></p><p><strong><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></strong></p><p>In the rapidly evolving landscape of machine learning, the need for understanding how our models make decisions has become increasingly crucial. This is particularly true in the domain of Graph Neural Networks (GNNs), where the complexity of graph structures adds another layer of intricacy to the already challenging task of model interpretation. In this first part of our three-part series, we’ll dive deep into the concept of explainability and its various problem settings.</p><h3>Why Do We Need Explainability?</h3><p>Imagine you’re a bank using a machine learning model to approve or deny loan applications. When the model denies a loan, both the bank and the applicant need to understand why. This isn’t just about transparency — it’s about trust, fairness, and the ability to provide actionable feedback to customers.</p><p>Similarly, in healthcare or drug discovery, when a model predicts that a molecule is toxic, scientists need to understand which molecular substructures contribute to this toxicity. This understanding isn’t just academic; it’s crucial for developing safer drugs and ensuring the reliability of our AI systems.</p><h3>The Black Box Problem</h3><p>Traditional machine learning methods often provided clear insights into their decision-making process:</p><ul><li>In linear regression, the slope of each feature directly indicates its importance</li><li><strong>Decision trees present a transparent, step-by-step logic for classifications</strong></li><li>Dimensionality reduction techniques clearly show how data points relate to each other</li></ul><p><strong>However, deep learning models, including GNNs, operate as “black boxes.” </strong>Their complex architectures and non-linear transformations make it challenging to trace how input features influence the final predictions. This opacity becomes particularly problematic when these models are deployed in critical applications like healthcare, financial services, or criminal justice systems.</p><h3>Types of Explainability</h3><h3>1. Instance-Level vs. Model-Level Explanations</h3><ul><li><strong>Instance-Level Explanations</strong>: Focus on understanding why a specific prediction was made. For example, explaining why a particular financial transaction was flagged as fraudulent by highlighting suspicious patterns in the transaction network.</li><li><strong>Model-Level Explanations</strong>: Aim to understand the overall behavior of the model across all predictions. This might involve identifying what general patterns the model has learned about toxic molecules across an entire chemical database.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UKOgpqL65ZHhL_QJ.png" /><figcaption><a href="https://ar5iv.labs.arxiv.org/html/2012.15445">Source</a></figcaption></figure><h3>2. Ante-hoc vs. Post-hoc Approaches</h3><ul><li><strong>Ante-hoc Explainability</strong>: Built into the model from the beginning, making the model inherently interpretable. This approach often involves architectural choices that prioritize transparency.</li><li><strong>Post-hoc Explainability</strong>: Applied after model training to explain predictions. This is more flexible as it can be applied to any existing model but may require additional computational overhead.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/788/0*AsJpIqoQIXvQavMw.png" /><figcaption><a href="https://medium.com/@sparsha.stars/explainable-artificial-intelligence-technical-perspective-part-1-6eec91b4cc60">source</a></figcaption></figure><h3>3. Model-Specific vs. Model-Agnostic Methods</h3><ul><li><strong>Model-Specific</strong>: Tailored to particular types of models, leveraging their unique architectures for better explanations.</li><li><strong>Model-Agnostic</strong>: Can be applied to any model type, offering more flexibility but potentially less detailed insights.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*be33xb2yXmfjUHbY-lbQ1g.png" /><figcaption><a href="https://www.researchgate.net/figure/XAI-Explanation-Model-Specific-vs-Model-Agnostic_fig7_375598223">source</a></figcaption></figure><h3>The Dual Nature of Graph Explanations</h3><p>What makes explainability in GNNs particularly interesting is its dual nature. When explaining predictions on graphs, we need to consider:</p><ol><li><strong>Feature Importance</strong>: Which node or edge attributes are crucial for the prediction?</li><li><strong>Structural Importance</strong>: Which connections and subgraphs in the network are influential?</li></ol><p>For instance, in a social network fraud detection system, we might need to understand both:</p><ul><li>The important user features (transaction amounts, frequency)</li><li>The crucial relationship patterns (unusual connection patterns, cyclic transactions)</li></ul><h3>Challenges and Considerations</h3><h3>1. Protected Characteristics</h3><p>Explainability helps identify when models might be using sensitive attributes (like gender or race) in their decision-making process, even when these aren’t explicitly provided as features. This is crucial for ensuring ethical AI deployment.</p><h3>2. Model Transferability</h3><p>Understanding what features drive predictions helps assess how well a model will transfer to new scenarios. For example, if a dog classification model relies heavily on the presence of humans in photos, it might fail when applied to images without people.</p><h3>3. Balancing Detail and Simplicity</h3><p>Explanations must be detailed enough to be meaningful but simple enough to be understood. This is particularly challenging in graph contexts, where both node features and network structure must be considered.</p><h3>Looking Ahead</h3><p>As we continue to deploy GNNs in critical applications, the need for reliable explainability methods becomes increasingly important. In the next part of this series, we’ll dive into GNNExplainer, a powerful framework for generating explanations for graph neural networks.</p><p>The field of explainable AI in graphs is still evolving, with new challenges and solutions emerging regularly. As we work to bridge the gap between model complexity and human understanding, we’re not just making our models more transparent — we’re building trust in AI systems that will increasingly shape our future.</p><p><em>This article is brought to you by ElloraSolutions.</em></p><p><em>@2024–2025 Ellora Solutions, Ellora AI and its affiliates. All rights reserved.</em></p><p><strong>Authored by:</strong><br><strong>Yasmine Karray</strong></p><ul><li><em>Data and Decision Systems Engineering Student, University of Sfax (ENET’Com)</em></li><li><em>Intern at ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/yasmine-karray"><em>LinkedIn</em></a><em> | </em><a href="https://medium.com/@ykarray29"><em>Medium</em></a></li></ul><p><strong>Dhaval Shah</strong></p><ul><li><em>CTO &amp; Founder, ElloraSolutions</em></li><li><a href="https://www.linkedin.com/in/dhaval-shah"><em>LinkedIn</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2f8a15b89001" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/gnn-xai-explainability-in-graph-neural-networks-understanding-the-black-box-part-1-2f8a15b89001">GNN-XAI: Explainability in Graph Neural Networks: Understanding the Black Box (Part 1)</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to co-author on Medium]]></title>
            <link>https://medium.com/ellora-ai/how-to-co-author-on-medium-7bb3da484fff?source=rss----f22a3c540553---4</link>
            <guid isPermaLink="false">https://medium.com/p/7bb3da484fff</guid>
            <category><![CDATA[publications-on-medium]]></category>
            <category><![CDATA[co-authoring]]></category>
            <category><![CDATA[publishing]]></category>
            <dc:creator><![CDATA[Dhaval Shah]]></dc:creator>
            <pubDate>Mon, 02 Dec 2024 23:56:13 GMT</pubDate>
            <atom:updated>2024-12-12T20:12:29.640Z</atom:updated>
            <content:encoded><![CDATA[<p>Co-authoring on Medium is notoriously difficult. There are not many medium articles that tell you how to co-author OR there are many articles but the search terms on co-author do not bring up many OR the search terms I am entering is not clear to the Medium’s search feature.</p><p>After considerable search, as the above paragraph suggests, I figured out that this might be the possible way to publish a medium article with multiple co-author(s). Note this is work in progress, and implementing this, you may still find yourself ending up <a href="https://en.wikipedia.org/wiki/Schitt%27s_Creek">Schitt’s Creek</a>.</p><p>As I understand, co-authoring can only be done as part of Medium Publishing.</p><h4>Familiarize yourself with Medium Publishing</h4><ol><li>Familiarize yourself with Medium Publishing : <a href="https://help.medium.com/hc/en-us/articles/115004681607-Getting-started-with-a-Medium-publication"><strong>https://help.medium.com/hc/en-us/articles/115004681607-Getting-started-with-a-Medium-publication</strong></a></li></ol><p>Repeating from above:</p><p>a. <strong>Medium publications are shared spaces</strong> for stories written around a common theme or topic, usually by <strong>multiple authors</strong>.</p><p>b. You must have a Medium account with an active membership subscription in order to create a Medium publication. <a href="https://help.medium.com/hc/en-us/articles/115004545567">Learn more about becoming a Medium member</a>.</p><p>c. Setup your publication. Viz: Steps 1 and 2 in “<a href="https://help.medium.com/hc/en-us/articles/115004681607-Getting-started-with-a-Medium-publication">Getting started with medium publication</a>” in the link above.</p><p>d. <a href="https://help.medium.com/hc/en-us/articles/115004681647-Manage-publication-editors-and-writers"><strong>Step 3: Add publication editors and writers</strong></a><strong>.</strong></p><p>Now you are ready to write your first article for the publication. To do just that, I am following the links:</p><p>i. <a href="https://help.medium.com/hc/en-us/articles/213904978-Add-a-draft-or-post-to-publication">https://help.medium.com/hc/en-us/articles/213904978-Add-a-draft-or-post-to-publication</a></p><p>^ Above is actually not very helpful. Since the documented menu items is not seen on my desktop browser!</p><p>Try to create a draft and publish it in the publication you created as part of steps above.</p><h4>Drafts in Publication</h4><p>I created a publication and a draft article within that and sent it to my co-author. My co-worker did the same on another draft article.</p><p>Good part is, you will see the draft articles in the notification bell icon on top right. Left of your profile icon. On right of your profile icon, you will see the publication(s) you have created. I created only one, and have no intention to create many at this point. So I see only one.</p><p>Anyway, above is only the good part. Once you click on the draft, you can only read it. You cannot edit it. You do not have any documented way of editing it as well.</p><p>At this point, I will pause unless a better suggestion comes along.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7bb3da484fff" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ellora-ai/how-to-co-author-on-medium-7bb3da484fff">How to co-author on Medium</a> was originally published in <a href="https://medium.com/ellora-ai">Ellora AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>