Community Detection via Autoencoder-Like Nonnegative Tensor Decomposition.

Area of Science:

• Network Science
• Data Mining
• Machine Learning

Background:

• Community detection partitions networks into dense subgraphs.
• Nonnegative matrix factorization (NMF) is common but overlooks multihop topology and data sparsity.
• Existing NMF methods struggle with sparse and complex network structures.

Purpose of the Study:

• To develop a novel community detection method addressing limitations of NMF.
• To incorporate multihop network topology and handle sparse adjacency matrices.
• To enhance community detection accuracy and robustness in complex networks.

Main Methods:

• Proposed an adjacency tensor to capture multihop network topology.
• Developed autoencoder-like nonnegative tensor decomposition (ANTD) using Tucker decomposition.
• Introduced an encoder component for improved community detection quality and an efficient optimization algorithm.

Main Results:

• ANTD demonstrates superior performance compared to 27 state-of-the-art methods on benchmark networks.
• The proposed method shows effectiveness, efficiency, and robustness in community detection.
• A graph regularized variant of ANTD was also studied, showing promising results.

Conclusions:

• ANTD offers a significant advancement in network community detection.
• The method effectively handles multihop topology and sparse data.
• ANTD provides a robust and efficient solution for complex network analysis.