Community Detection on Model Explanation Graphs for Explainable AI
Ehsan Moradi
Paper
Details
#ExplainableAI #ModelExplanationGraphs #CommunityDetection

DynBenchmark provides customizable benchmarks for community tracking

DynBenchmark, shown at FRCCS 2025 (May, Bordeaux, pp 74‑85), offers Python libraries and visualization tools to build temporal networks for community‑detection tests. getnews.me/dynbenchmark-provides-cu... #dynbenchmark #communitydetection

Robustness of Graph Neural Networks for Community Detection

A study of six GNN models found the unsupervised DMoN most stable under adversarial attacks, while supervised models lose accuracy when edges are deleted. Read more: getnews.me/robustness-of-graph-neur... #gnn #communitydetection #robustness

Phase Transition for Stochastic Block Models with Many Communities

Research proves low‑degree polynomial estimators fail below a revised threshold for SBMs with >√n communities, but counting cliques enables recovery above it. getnews.me/phase-transition-for-sto... #stochasticblockmodel #communitydetection

Study Reveals Information Loss in Network Embedding Techniques

Researchers found a network‑embedding algorithm fully captures a graph when its mapping is invertible; otherwise it loses edge‑density information, hurting detection. Read more: getnews.me/study-reveals-informatio... #networkembedding #communitydetection

🚨 New paper out from our lab! 🚨

We introduce RC-CCD, a novel framework for community detection in complex networks using rough set theory and consensus clustering.

#CommunityDetection #GraphTheory #RoughSets #ConsensusClustering #ComplexNetworks #AIresearch

Plots showing the mean and range of values derived from increasing amounts of data from single-file movements, scan samples and focal observations. The values from each method converge to similar estimates of the number of distinct communities in the group

However, when we looked at the broader community structure, we found that the single-file movement data and traditional methods converged on similar estimates. This suggests that single-file movement data might be useful for coarse estimation of group-level structure #communitydetection 6/9

Well-connectedness of communities: Park et al show that many communities detected with standard algorithms are not well-connected: by cutting a few edges, such a community breaks into 2. Remedy by post-processing.

https://doi.org/10.1371/journal.pcsy.0000009

#clustering #communitydetection

Our new Local #CommunityDetection in #DynamicGraphs Using #PersonalizedCentrality
http://bit.ly/2vJEOCl
@Algorithms_MDPI @MDPIOpenAccess