arxiv_ai 95% Match Research Paper AI Researchers,Machine Learning Engineers,Data Scientists,AI Ethicists,Model Developers 2 days ago

Community Detection on Model Explanation Graphs for Explainable AI

ai-safety › interpretability

📄 Abstract

Abstract: Feature-attribution methods (e.g., SHAP, LIME) explain individual predictions but often miss higher-order structure: sets of features that act in concert. We propose Modules of Influence (MoI), a framework that (i) constructs a model explanation graph from per-instance attributions, (ii) applies community detection to find feature modules that jointly affect predictions, and (iii) quantifies how these modules relate to bias, redundancy, and causality patterns. Across synthetic and real datasets, MoI uncovers correlated feature groups, improves model debugging via module-level ablations, and localizes bias exposure to specific modules. We release stability and synergy metrics, a reference implementation, and evaluation protocols to benchmark module discovery in XAI.

Authors (1)

Ehsan Moradi

Submitted

October 31, 2025

arXiv Category

cs.SI

arXiv PDF

Key Contributions

Proposes Modules of Influence (MoI), a framework that constructs model explanation graphs from per-instance attributions, uses community detection to find jointly affecting feature modules, and quantifies their relation to bias, redundancy, and causality. It enables model debugging via module-level ablations and bias localization.

Business Value

Enhances the trustworthiness and reliability of AI models by providing deeper insights into their decision-making processes, aiding in debugging, fairness assessments, and regulatory compliance.

Paper Metadata

Innovation Type

Framework/Methodology

Deployment Feasibility

Moderate. Requires per-instance attributions from existing models and computational resources for graph construction and community detection. The release of a reference implementation aids feasibility.

Limitations Addressed

Feature attribution methods often miss higher-order feature interactions,Difficulty in understanding how groups of features collectively influence predictions,Challenges in debugging models based on feature correlations,Lack of methods to localize bias to specific feature sets

Performance Gains

Improves model debugging, bias localization, and understanding of feature interactions. Specific quantitative gains are not detailed.

Technical Tags

Explainable AI (XAI)Feature AttributionCommunity DetectionModel Explanation GraphsModules of Influence (MoI)Bias DetectionCausalityModel Debugging

Research Topics

Explainable AIMachine Learning InterpretabilityAI FairnessModel DebuggingCausal Inference

Methods & Architectures

Construction of model explanation graphsCommunity detection algorithmsModule-level ablationsQuantification of module relationships (bias, redundancy, causality) Models explained by feature attribution methods (e.g., SHAP, LIME)

Applications & Tasks

Machine Learning Model Development AI Auditing AI Ethics Identifying higher-order structure in feature attributionsDiscovering sets of features that act in concertDebugging models based on feature interactionsLocalizing bias and understanding causality patterns Model InterpretabilityBias DetectionCausal AnalysisModel DebuggingFeature Interaction Analysis

Datasets & Benchmarks

Datasets

Synthetic datasets, Real datasets

Stability metricsSynergy metrics

Related Fields

Explainable AIMachine LearningData MiningGraph TheoryAI EthicsCausal Inference

Keywords

Explainable AIXAIFeature AttributionCommunity DetectionModel Explanation GraphsModules of InfluenceMoIBias DetectionCausalityModel DebuggingFeature InteractionsSHAPLIME

Academic Context

#Explainable AI#Machine Learning Interpretability#AI Fairness#Model Debugging#Causal Inference

Technology Stack

Frameworks & Libraries

MoI (Modules of Influence)

Commercial Potential

Potential Products

AI model debugging toolsExplainability platformsBias detection and mitigation softwareTools for analyzing feature interactions

Target Industries

TechnologyFinanceHealthcareAny industry using ML models

Use Case Examples

Identifying why a loan application was denied based on feature groupsDebugging a medical diagnosis model by analyzing feature module behaviorUnderstanding how sets of features contribute to bias in a hiring algorithm

Competitive Edge

Goes beyond per-instance explanations (like SHAP/LIME) to uncover higher-order, group-level feature interactions and their implications for bias and causality.

Resource Requirements

Deployment Constraints

Computational cost of generating explanation graphs,Dependence on the quality of initial feature attributions

Scalability

Scalability depends on the size of the explanation graph and the community detection algorithm used. The paper provides metrics to benchmark module discovery, implying scalability considerations.

Regulatory Considerations

AI fairness regulationsGDPR (right to explanation)

Production Readiness

Maturity Level

Research

View Full Paper Back to Papers