Master Thesis — Alessandro Carella

As artificial intelligence systems become increasingly integrated into critical decision-making processes across healthcare, finance, justice, and other high-stakes domains, the need for transparent and interpretable explanations of machine learning predictions has become clear. My master's thesis project addresses the fundamental challenge of transforming complex explainability method outputs into intuitive visual representations that support exploration, confirmation, and understanding.

The proposed framework enables users to iteratively explore how local decision boundaries are captured by synthetic neighborhoods and how these neighborhoods translate into interpretable decision rules. Rather than relying on static textual representations that place significant cognitive burden on users, my system implements interactive visualizations that bridge the gap between algorithmic explanations and human understanding.

The system focuses on local explainability methods that generate synthetic neighborhoods around explained instances and extract rules through surrogate models. While implemented using the LORE sa (Stable and Actionable Local Rule-based Explanations) method, the framework's architecture could easily accommodate other state-of-the-art methods.

Through the thesis I gained insights into the intersection of explainable AI, information visualization, and human-computer interaction. I discovered that effective explanation communication requires more than sophisticated algorithms, it demands carefully designed interactive systems that support diverse analytical workflows and accommodate different cognitive preferences.

I learned the importance of coordinated multiple views in complex data exploration. By implementing bidirectional coordination between spatial neighborhood projections and hierarchical rule structures, I learned how enabling users to transition between spatial intuition and logical reasoning creates more powerful analytical capabilities than either representation alone.

Working with dimensionality reduction techniques deepened my understanding of their trade-offs. I learned that UMAP excels at both local and global structure preservation while maintaining computational efficiency, making it ideal for interactive applications. PCA 's linear transformation preserves decision boundaries, enabling direct visualization of surrogate model regions. Each technique offers distinct advantages for different analytical scenarios.

Example of UMAP output on a generated neighborhood

Example of PCA output on a generated neighborhood

The development of three alternative decision tree visualizations, Tree Layout, Rule and Counterfactual Rules Centered, and Rule Centered, taught me that there is no single "best" visualization approach. Different layouts serve different cognitive approaches and analytical needs.

The Voronoi tessellation implementation for decision boundary visualization taught me sophisticated geometric algorithms and their application to machine learning visualization. Balancing computational efficiency with visual accuracy through user-configurable granularity settings demonstrated the importance of providing users control over performance-quality trade-offs.

Perhaps most importantly, I learned that effective XAI tools must address cognitive load through visual connection mechanisms, enable confirmation through interactive exploration, and support diverse mental models through multiple representation formats. The documented use cases validated that well-designed coordination mechanisms transform explanation consumption from passive acceptance into active analytical investigation.

The system architecture consists of two primary components: a progressive configuration interface and a coordinated visualization system. The configuration interface implements a seven-stage workflow including dataset selection, classifier configuration, instance specification, explanation parameters, dimensionality reduction settings, and visualization selection.

The spatial neighborhood analysis component employs scatter plot projections with four dimensionality reduction techniques: UMAP, t-SNE, PCA, and MDS. Each method is accessible through radio button controls, enabling rapid switching between projection methods for comparative analysis.

Example of t-SNE output on a generated neighborhood

Example of MDS output on a generated neighborhood

For PCA projections, the system implements Voronoi tessellation-based decision boundary visualization. This approach naturally accommodates irregular decision boundaries produced by decision trees. The algorithm constructs a dense mesh grid sampling the 2D visualization space, applies inverse PCA transformation to project grid points back to the original feature space, obtains surrogate model predictions for each point, and generates Voronoi diagrams with graph-based merging of adjacent regions sharing identical class predictions.

Each of the three decision tree layouts serves different analytical needs. The Tree Layout implements the traditional hierarchical structure using the Reingold-Tilford algorithm, providing comprehensive views of entire surrogate model structures with clear parent-child relationships.

The Rule and Counterfactual Rules Centered visualization employs depth-aligned positioning where nodes at equivalent depths align vertically in columns. This layout positions the explained instance path at the bottom for immediate identification, with alternative decision paths sorted by divergence points and positioned above. Rectangular nodes maximize information density while straight-line connections with sample-proportional stroke widths reveal instance flow patterns.

Rule and Counterfactual Rules Centered Layout

The Rule Centered visualization implements adaptive spacing based on subtree complexity, with horizontal distances between path nodes calculated proportionally to off-path subtree sizes. The layout features interactive subtree collapse/expand functionality, enabling progressive disclosure where users focus on explanation-relevant paths while selectively exploring alternatives through right-click context menu operations.

Bidirectional coordination serves as the fundamental interaction paradigm. When users click spatial neighborhood points, the system identifies corresponding feature vectors, traces decision paths through tree structures, and propagates highlighting. Conversely, tree node interactions identify synthetic neighborhood instances satisfying node conditions and highlight corresponding scatter plot points. This coordination maintains visual consistency through synchronized color schemes across all visualization components.

Rule and Counterfactual Rules Centered Layout on point click

Rule and Counterfactual Rules Centered Layout on leaf click

The backend employs Python with scikit-learn for machine learning workflows, DEAP framework for genetic algorithm implementation, and optimized CART decision tree construction. The frontend utilizes D3.js for interactive visualizations, implementing custom layouts, Voronoi tessellation algorithms, and sophisticated coordination mechanisms. The system integrates with Jupyter notebooks, supporting both standalone web interfaces and embedded notebook views for data science workflows.

This work makes several contributions to the field of explainable artificial intelligence and visual analytics. The primary contribution is the dual-representation coordination strategy that integrates spatial neighborhood analysis with hierarchical rule visualization through bidirectional interaction mechanisms. This approach addresses fundamental limitations in current XAI presentation formats by making the relationship between synthetically generated neighborhoods and resulting explanations explicitly visible and explorable.

The development of three alternative decision tree layouts advances visualization design for explanation scenarios. The Rule and Counterfactual Rules Centered visualization's depth-aligned positioning enables direct comparison of decision criteria across branches at equivalent depths. The Rule Centered visualization's adaptive spacing and progressive disclosure supports efficient exploration of complex decision structures while maintaining focus on explanation-relevant paths.

The system demonstrates practical value across diverse application scenarios. For educational contexts, the progressive disclosure workflow supports scaffolded introduction to XAI concepts, while multiple tree layouts enable demonstration of how different visual representations reveal complementary insights. For professional data science workflows, the iterative refinement capabilities through parameter adjustment enable thorough investigation of explanation quality and decision space coverage.

The work validates key design principles for XAI visualization: information integration reduces cognitive load through coordinated views, interactive confirmation supports explanation trust through bidirectional exploration, and multiple representation formats accommodate diverse user mental models and analytical preferences.

Several enhancement opportunities could extend the system's capabilities while maintaining its core coordination principles. Visual encoding enhancements include multi-colored edge encoding displaying class distribution proportions within decision tree branches, as implemented in BaobabView, one of the surveyed visualization systems. This would provide immediate visibility of class composition without requiring tooltip interactions, though careful visual hierarchy management would be necessary to maintain clarity.

Adaptive interaction mechanisms could provide more nuanced complexity control. A hybrid visualization mode enabling progressive transformation of individual counterfactual paths from circular to rectangular representations would support natural analytical workflows where users start with focused examination and gradually expand scope based on discovered patterns. Threshold-based leaf filtering through user-adjustable sliders would enable dynamic hiding of low-sample leaves, managing visual complexity while maintaining access to complete information through expandable stub branches.

Clustering-based alternative coloring schemes would enable investigation of intrinsic neighborhood structure beyond class-based organization. Toggle controls switching between class-based and clustering-based coloring would support cluster-class correspondence verification, within-class pattern discovery, and projected space artifact detection. This would create rich exploration spaces when combined with the existing dimensionality reduction technique selection capabilities.

Textual rule export functionality would address scenarios requiring communication through channels not supporting interactive visualization. Right-click context menu operations on tree nodes could generate formatted textual representations in multiple output formats including natural language for non-technical audiences, Python-style conditionals for code integration, formal logic notation for academic manuscripts, and LaTeX table formats for document insertion.

This thesis demonstrates that effective explainable AI requires more than sophisticated algorithms, it demands carefully designed interactive visual systems that bridge computational outputs with human understanding. The developed framework successfully transforms complex explainability method outputs into intuitive visual representations through coordinated multiple views, enabling users to transition fluidly between spatial intuition and logical reasoning.

The system's dual-representation strategy, combining scatter plot projections with alternative decision tree layouts through bidirectional coordination mechanisms, addresses critical gaps in current XAI visual analytics approaches. By making the relationship between synthetic neighborhoods and extracted rules explicitly visible and explorable, the framework transforms explanation consumption from passive acceptance into active analytical investigation.

The documented use cases showcase the framework's effectiveness across diverse scenarios, from pedagogical applications introducing XAI concepts to professional data science workflows investigating complex decision patterns. The iterative development process establishes methodological patterns for future enhancement evaluation.

While implemented using the LOREsa explainability method, the framework's architecture and coordination mechanisms could readily accommodate other state-of-the-art techniques generating synthetic neighborhoods and extracting symbolic rules. This extensibility, combined with the system's demonstrated effectiveness and identified future enhancement opportunities, positions this work as a foundation for continued advancement in explainable AI visual analytics.