Joint Learning of Feature Extraction and Clustering for Large-Scale Temporal Networks

Area of Science:

• Network Science
• Data Mining
• Computational Social Science

Background:

• Temporal networks are crucial for understanding dynamic systems.
• Existing dynamic community detection algorithms struggle with vertex-level dynamics, feature independence, and computational complexity.

Purpose of the Study:

• To propose a novel joint learning model for dynamic community detection in temporal networks (jLMDC).
• To address limitations of current algorithms by integrating feature extraction and clustering.
• To improve accuracy and efficiency in identifying dynamic and overlapping communities.

Main Methods:

• Formulated jLMDC as a constrained optimization problem.
• Classified vertices into dynamic and static groups based on temporal network topology.
• Updated dynamic vertex features while preserving static vertex features during optimization.
• Extended jLMDC for overlapping dynamic community detection.

Main Results:

• jLMDC integrates feature extraction and clustering for improved performance and reduced running time.
• Experimental results on 11 temporal networks show accuracy improvements up to 8.23%.
• The model achieved an average reduction of 24.89% in running time compared to state-of-the-art methods.

Conclusions:

• jLMDC effectively characterizes vertex-level dynamics in temporal networks.
• The joint learning approach enhances accuracy and efficiency in dynamic community detection.
• jLMDC offers a robust solution for both non-overlapping and overlapping dynamic community detection.