An Enhanced Dynamic Network Community Identification by Utilizing Deep Learning
MSc Thesis Defense by: Kimia Shahmiri
Date: Wednesday, September 9th, 2026
Time: 9:15
Location: Essex Hall 122
Abstract:
Dynamic networks evolve over time as nodes, connections, and community structures change. Detecting meaningful communities in such networks requires a method that captures both the structure of the current network and its temporal evolution. This thesis investigates and extends Deep Learning and Evolutionary Clustering (DLEC), a framework that combines deep representation learning, temporal regularization, and K-Means clustering for dynamic community detection.
The proposed work focuses on improving the reconstruction stage of DLEC through two modifications. First, the original sigmoid decoder is replaced with a linear output decoder to provide greater flexibility when reconstructing continuous community-similarity values. Second, a sparsity-aware reconstruction loss is introduced to assign greater weight to informative positive entries in sparse similarity matrices. The proposed modifications are evaluated on synthetic dynamic networks with fixed and varying community structures, as well as the real-world Enron email network. Performance is assessed using Normalized Mutual Information (NMI), reconstruction Mean Squared Error (MSE), modularity, and temporal NMI. Experimental results show that the linear decoder consistently improves community detection performance in the synthetic benchmarks, while moderate sparsity weighting further improves performance in fixed-community settings. Results on Enron demonstrate a trade-off between reconstruction quality, community structure, and temporal stability.
Keywords: Dynamic Community Detection; Deep Learning; Evolutionary Clustering; Autoencoder; Network Representation Learning
Thesis Committee:
Internal Reader: Dr. Hamidreza Koohi
External Reader: Dr. Kristoffer Romero
Advisor: Dr.Ziad Kobti
Chair: Dr. Shaon Bhatta Shuvo