An Enhanced Dynamic Network Community Identification by Utilizing Deep Learning, Fuzzy Clustering, and an Evolutionary Algorithm
MSc Thesis Proposal by: Kimia Shahmiri
Date: Friday May 22nd, 2026
Time: 9:30AM
Location: Microsoft Teams
Abstract:
Dynamic community detection in temporal networks remains a challenging problem, as real-world networks continuously evolve over time through changes in node connections, community merges, and splits. Existing methods either treat each time snapshot independently, losing temporal continuity, or rely on rigid hard-assignment clustering that fails to capture nodes belonging to multiple communities simultaneously. This thesis proposes CIDFE (Community Identification using Deep learning, Fuzzy clustering, and Evolutionary algorithm), an enhanced framework built upon the DLEC algorithm (Pan et al., 2024) that addresses these limitations through three integrated contributions. First, a deep autoencoder learns low-dimensional structural embeddings from the community similarity matrix at each time step. Second, Fuzzy C-Means clustering replaces hard assignment to produce soft membership matrices, enabling nodes to belong to overlapping communities, a pattern commonly observed in real-world networks such as organizational hierarchies. Third, an evolutionary algorithm enforces temporal smoothness by penalizing abrupt community reassignments between consecutive snapshots. During implementation, a critical bug in the original DLEC architecture was identified: the use of a sigmoid activation on the decoder output layer causes mean collapse on sparse matrices, stagnating NMI at 0.25. Replacing it with a linear output layer improved NMI to 0.87 (+248%). The proposed CIDFE framework will be evaluated on synthetic benchmarks (SYN-FIX, SYN-VAR) and real-world networks (Cellphone Calls, Enron Email) using NMI and Partition Coefficient as evaluation metrics.
Thesis Committee:
Internal Reader: Dr Hamidreza Koohi
External Reader: Dr. Kristoffer Romero
Advisor: Dr. Ziad Kobti
