Highlights:
- Researchers introduce ‘Modules of Influence’ (MoI), a framework combining graph analysis with model explainability.
- MoI constructs model explanation graphs and applies community detection to expose interacting feature modules.
- The method identifies how feature groups contribute to model biases, redundancy, and causal relationships.
- MoI enhances model debugging and improves trust in AI systems through multi-feature analysis.
TLDR:
Ehsan Moradi’s paper presents ‘Modules of Influence’ (MoI), an innovative framework for explainable AI that employs community detection algorithms to uncover groups of inter-related features affecting AI predictions. The breakthrough deepens interpretability, reveals hidden biases, and improves trust in machine learning systems.
Explainable Artificial Intelligence (XAI) has become a critical field ensuring that machine learning models are transparent, interpretable, and accountable. Traditional feature-attribution methods like SHAP and LIME are well-known for explaining individual model predictions by assigning importance values to input features. However, these approaches often miss complex, higher-order relationships—sets of features that interact with one another to influence outcomes. To address this limitation, Ehsan Moradi (https://arxiv.org/search/cs?searchtype=author&query=Moradi,+E) introduces a novel framework called ‘Modules of Influence’ (MoI), described in the paper *Community Detection on Model Explanation Graphs for Explainable AI*.
The MoI framework reimagines model interpretability through the lens of network science. It begins by constructing a ‘model explanation graph’ that maps features as nodes, with edges representing co-influence relationships derived from per-instance attributions. Once the graph is established, Moradi applies community detection algorithms—commonly used in social network analysis—to identify groups or ‘modules’ of features that act collectively. These modules reveal how intertwined features contribute jointly to predictions and allow practitioners to localize regions of bias, redundancy, or synergy within the model’s behavior.
Through experiments on both synthetic and real-world datasets, MoI demonstrates its ability to uncover meaningful feature correlations that are overlooked by standard attribution methods. The framework supports powerful new diagnostic tools such as ‘module-level ablations’—analyses that selectively remove feature clusters to evaluate their collective impact on model output. The research also introduces stability and synergy metrics to benchmark these discoveries. Beyond explainability, MoI’s ability to map bias exposure at the module level provides important implications for ethical AI design, fairness auditing, and model debugging. By combining graph theory with explainable AI, Moradi’s approach marks a significant leap toward more transparent and trustworthy AI systems.
Source:
Source:
Original research paper: Ehsan Moradi, ‘Community Detection on Model Explanation Graphs for Explainable AI,’ arXiv:2510.27655v1 [cs.SI], https://doi.org/10.48550/arXiv.2510.27655