Highlights:
- Introduces ‘Modules of Influence’ (MoI) for analyzing model explanation graphs.
- Uses community detection algorithms to identify groups of interacting features.
- Improves model interpretability, debugging, and bias localization.
- Offers open-source implementation with new stability and synergy metrics.
TLDR:
Ehsan Moradi’s new framework, ‘Modules of Influence’ (MoI), leverages graph-based community detection to uncover groups of features that work together in AI models, helping researchers understand complex interactions, reduce bias, and improve model transparency.
In a recent advancement in the field of Explainable Artificial Intelligence (XAI), researcher Ehsan Moradi has introduced a novel framework titled ‘Modules of Influence’ (MoI), as detailed in the paper ‘Community Detection on Model Explanation Graphs for Explainable AI’ (arXiv:2510.27655). While popular feature-attribution methods such as SHAP and LIME focus on explaining individual model predictions, they often overlook how combinations of features interact to influence outcomes. MoI addresses this gap by applying community detection principles from network science to uncover high-level feature collaborations within model explanation graphs.
The MoI framework begins by constructing a model explanation graph using per-instance attribution data. Each node corresponds to a feature, and weighted edges represent relationships or co-influences discovered across instances. By applying community detection algorithms to these graphs, MoI automatically identifies ‘modules’—groups of features that act together to sway a model’s predictions. This modular perspective enables researchers to understand higher-order patterns of influence, making AI models more interpretable at a systemic level rather than just locally.
One of MoI’s key strengths lies in its capacity to reveal correlations, redundancies, or biases embedded within models. By isolating these feature communities, practitioners can perform module-level ablations to study how removing or altering entire groups impacts model behavior. Furthermore, the framework includes new quantitative metrics—stability and synergy—to measure the consistency and cooperative effects of identified modules. According to Moradi, experiments on both synthetic and real-world datasets show that MoI not only improves interpretability but also serves as a robust diagnostic tool for model debugging and fairness analysis.
The release of the reference implementation, evaluation protocols, and performance benchmarks marks a significant step toward standardizing how the research community explores multi-feature interactions in explainable AI. This innovation is set to become a valuable addition to the AI interpretability toolbox, especially in domains where transparency and trust are critical, such as healthcare, finance, and social decision-making systems.
Source:
Source:
Ehsan Moradi, ‘Community Detection on Model Explanation Graphs for Explainable AI’, arXiv:2510.27655v1 [cs.SI], DOI: https://doi.org/10.48550/arXiv.2510.27655