Hackathon XAI — Alessandro Carella

In an era where AI influences decision-making across numerous domains, the lack of transparency in black-box models presents a significant challenge to understanding and trusting their decisions. This hackathon project explores how Explainable AI (XAI) enables comprehension of complex machine learning models and augments human understanding of AI-driven decisions.

Our team implemented a gradient boosting classifier for uplift modeling in marketing campaigns and applied three XAI techniques, LIME , LORE , and SHAP , to uncover how the model arrives at its predictions. The project focuses on identifying which customers are most likely to respond positively to marketing promotions, categorizing them into four segments: Sure Things, Persuadables, Sleeping Dogs, and Lost Causes.

By evaluating these explanations across three critical dimensions, content, presentation, and user experience, we aimed to demonstrate how diverse XAI methods can serve stakeholders with varying technical expertise, from executives to data engineers.

Me and my team achieved third place while competing with other master and PhD students.

This hackathon provided invaluable insights into the practical application of Explainable AI and its critical role in modern machine learning workflows.

I discovered that even high-performing models can make correct predictions for potentially wrong reasons. Through systematic analysis of 10,000 correctly classified instances, we found that when considering only the top three features, nearly half of the predictions (4,820 instances) might have been correct but relied on unexpected feature combinations. This reinforced the importance of not just achieving high accuracy but understanding the reasoning behind model decisions.

The uplift modeling approach taught me the importance of proper problem framing. By creating customer segments based on treatment response patterns and addressing class imbalance through SMOTE, I had additional proof that thoughtful data preparation directly impacts model interpretability. The decision to focus on "Persuadables", customers who buy only when receiving promotions, demonstrated how domain knowledge shapes both model design and business value.

The hackathon emphasized evaluating explainability along multiple dimensions. Correctness and comprehensiveness ensure explanations capture all relevant factors. Continuity validates that similar customer profiles receive consistent explanations. Compactness and composition affect how easily users can digest information. Context and coherence determine whether explanations align with real-world expectations. This evaluation framework showed me that explainability is as much about user experience as it is about technical accuracy.

Perhaps the most valuable lesson was understanding that effective XAI must be tailored to its audience. Executives need high-level insights for strategic decisions, marketing agents require actionable customer insights, and data engineers need detailed technical validation. By providing explanations suited to each user group's expertise level, XAI transforms from a technical exercise into a practical tool that drives trust and adoption across an organization.

The dataset focused on uplift modeling for marketing campaigns, with the primary goal of identifying customers most likely to respond positively to promotional offers. The analysis centered on two critical variables: the offer type sent to customers and whether they ultimately made a purchase (conversion).

Decision tree approach for categorization of customers

Using a decision tree approach, we categorized customers into four distinct segments that form the foundation of uplift modeling. Sure Things represent customers who purchase regardless of promotional intervention. Persuadables are the target segment, customers who only buy when receiving promotions. Sleeping Dogs are counterintuitively harmed by promotions and are less likely to purchase when contacted. Lost Causes will not convert regardless of marketing efforts.

This segmentation revealed that marketing promotions should be targeted exclusively at Persuadables, as they are the only segment where treatment yields better outcomes than no treatment. This insight drove the creation of our binary target variable "treat", indicating whether a customer should receive promotional offers.

The initial dataset exhibited significant class imbalance with 42,694 positive cases and only 21,306 negative cases. To ensure robust model training and prevent bias toward the majority class, we implemented the SMOTE (Synthetic Minority Over-sampling Technique) algorithm. This approach synthetically generated samples for the minority class, resulting in a balanced distribution of 42,614 instances for each class, providing a more reliable foundation for model development.

Categorical variables required transformation for model compatibility. The offer column was encoded with "Buy One Get One" as 1, "Discount" as -1, and other offers as 0. Channel variables were mapped with Web as 1, Phone as 2, and other channels as 3. Geographic information in the zipcode column was transformed with Suburban as 1, Urban as 2, and other regions as 3. This systematic encoding preserved ordinal relationships where appropriate while ensuring numerical compatibility with machine learning algorithms.

To predict customer treatment eligibility, we evaluated three ensemble learning algorithms: Gradient Boosting, Random Forest, and AdaBoost. Each model underwent rigorous hyperparameter optimization using 7-fold cross-validation to identify optimal configurations while preventing overfitting.

Gradient Boosting emerged as the superior performer with 79.3% accuracy, configured with a learning rate of 0.5, maximum depth of 10, subsample ratio of 1, and 75 estimators. Random Forest achieved 72.5% accuracy using 50 estimators, maximum depth of 10, and considering 5 features at each split. AdaBoost demonstrated the lowest performance at 66.4% accuracy with a learning rate of 0.9 and 15 estimators.

The Gradient Boosting classifier's balanced performance across both classes (81% precision and 77% recall for class 0, 78% precision and 82% recall for class 1) made it the ideal candidate for explainability analysis. Its sequential learning approach and ability to capture complex patterns in customer behavior provided a robust foundation for understanding feature importance and decision-making processes.

The selected model demonstrated consistent performance with macro and weighted averages of 79% across precision, recall, and F1-score metrics. This balanced performance indicated the model was neither overfitting to one class nor showing bias toward majority predictions, making it an excellent subject for explainability analysis where understanding nuanced decision boundaries is crucial.

To achieve comprehensive understanding of our black-box model, we implemented three complementary explainability techniques, each offering unique perspectives on model decision-making processes.

SHAP provides theoretically grounded explanations based on game theory principles. It calculates feature contributions by considering all possible feature combinations, ensuring consistent and fair attribution. The implementation revealed that for many instances, a single feature predominantly drove predictions, demonstrated by numerous instances having only one non-zero SHAP value. The beeswarm plots generated by SHAP offer intuitive visualizations showing both feature importance and directionality of influence, making it ideal for high-level stakeholders seeking quick insights.

LIME creates locally faithful approximations of the black-box model by training interpretable surrogate models around specific predictions. Our implementation typically generated explanations using seven features, providing detailed local insights into individual customer classifications. The bar chart visualizations clearly indicate positive and negative feature contributions, making it particularly valuable for understanding specific customer segments and informing targeted marketing strategies.

LORE generates decision rules that explain individual predictions through logical conditions. Analysis of 10,000 explanations revealed rule lengths averaging six conditions, with the longest containing eleven rules and the shortest just two. This rule-based format provides transparent logic chains showing exactly which feature value ranges led to specific predictions. LORE excels at helping technical users identify logical inconsistencies and verify model reasoning, though its effectiveness depends on rule complexity remaining manageable.

The three methods proved highly complementary. SHAP identified the most influential features globally and locally. LIME revealed how specific feature values affected individual predictions. LORE provided explicit logical rules for validation and debugging. Together, they offered multiple lenses through which to understand model behavior, addressing the needs of diverse stakeholders.

We assessed explainability methods across three critical dimensions, content, presentation, and user experience, ensuring comprehensive evaluation of both technical accuracy and practical usability.

Correctness and comprehensiveness examined whether explanations captured all relevant factors influencing predictions. For uplift modeling, this included purchase history, acquisition channel, recency, past promotion usage, and geographic classification. Completeness verified that highlighted features sufficiently explained decisions without omitting crucial factors. Continuity testing compared similar customer profiles (instances 792 and 10301) to ensure consistent explanations for analogous cases. LIME and SHAP both demonstrated strong continuity, identifying similar feature importance patterns for customers with comparable characteristics.

Compactness measured explanation conciseness. SHAP showed exceptional compactness with only one non-zero value for specific instances, indicating single dominant features. LIME explanations typically involved seven features, providing moderate compactness while maintaining informativeness. LORE rules averaged six conditions, with consistent median and mode values across 10,000 explanations. Composition analysis evaluated visualization formats: SHAP's beeswarm plots effectively showed feature contributions, LIME's bar charts clearly indicated directional influence, and LORE's decision rules provided logical walkthrough of reasoning.

Coherence evaluation analyzed whether explanations aligned with human intuition. Testing on a customer with web channel, median purchase history, no referral, recent purchase, BOGO (Buy One, Get One) usage, no discount usage, and urban location confirmed that all three methods produced explanations consistent with qualitative assessment. Context relevance ensured explanations addressed stakeholder needs, providing actionable insights for marketing strategy, model debugging, and business decision-making. Visual clarity through plots and rules enhanced interpretability across technical expertise levels.

A critical investigation examined instances where models make correct predictions but potentially based on inappropriate reasoning. By analyzing 10,000 correctly classified instances, we identified top features most frequently appearing in explanations. LIME and SHAP both highlighted is_referral, used_discount, channel, used_bogo, and zip_code as primary decision drivers with frequencies ranging from 4,820 to 5,750 appearances.

Iteratively reducing the feature set revealed concerning patterns. When considering only top three features, 4,820 instances were correctly classified but without relying on features like zip_code or specific BOGO/discount patterns. This suggests nearly half of correct predictions might stem from incomplete feature consideration. LORE analysis on a smaller sample of 250 instances identified history, recency, and used_discount as most frequent rule components, though rule-based analysis proved less suitable for this specific investigation.

Multiple explainability methods significantly enhance usability by addressing three key dimensions. Broad accessibility ensures both technical and non-technical stakeholders find suitable explanation formats. SHAP provides simple summary plots for executives alongside theoretically grounded values for data scientists. LIME offers probability-based visualizations accessible to marketing professionals. LORE delivers logical rules familiar to engineers and analysts.

Diverse visualization properties cater to different cognitive preferences. Rule-based visualizations suit users comfortable with logical reasoning. SHAP's dependence plots and force diagrams reveal global patterns and feature interactions. LIME's bar charts provide straightforward contribution indicators. Different error detection aspects emerge from each method: SHAP identifies potential biases through feature importance patterns, LIME reveals local anomalies in specific predictions, and LORE exposes logical inconsistencies through contradictory rules.

This hackathon project successfully demonstrated that Explainable AI is not merely a technical enhancement but a fundamental requirement for deploying trustworthy machine learning systems. By implementing LIME, SHAP, and LORE on a gradient boosting model for uplift modeling, we created a comprehensive framework for understanding black-box model decisions.

The multi-method approach proved essential for thorough model understanding. SHAP provided intuitive global insights ideal for strategic decision-makers. LIME delivered actionable local explanations supporting tactical marketing decisions. LORE offered detailed logical validation enabling technical quality assurance. Each method addressed distinct user needs while contributing to holistic model transparency.

The research revealed critical insights about model reliability. Nearly half of correct predictions potentially relied on incomplete feature consideration, highlighting that accuracy metrics alone provide insufficient assurance of model quality. Explainability methods enable detection of such issues, supporting more robust model development and deployment practices.

Ultimately, this hackathon reinforced that transparency and performance are not competing objectives but complementary requirements for responsible AI. By making model reasoning accessible and understandable, we enable organizations to leverage AI's capabilities while maintaining accountability, fairness, and trust in automated decision-making systems.