Core Concepts
HACD is a novel attributed community detection model that leverages heterogeneous graph attention networks to capture semantic similarity between node attributes and exploit mesoscopic community structure for improved performance in community detection tasks.
Stats
HACD achieves the highest improvement of 23.49%, 24.26%, 17.19%, 21.45%, and 4.58% in five evaluation indicators compared to the best-performing baseline methods.
HACD shows significant performance gains on the Pubmed and DBLP datasets, demonstrating its capability to handle large-scale data.
Using the updated graph structure, where node attributes are treated as a distinct node type, leads to improved performance compared to models using the original graph structure.
On the DBLP dataset, as the range of noise distribution expands, the modularity only decreases by about 5%, indicating the robustness of HACD.
The running time of HACD is competitive and scales well with the dataset size.
Quotes
"Existing methods usually disregard the semantic similarity between node attributes within communities, leading to the omission of crucial nodes in the detected communities."
"Inherent community structure, serving as a crucial mesoscopic description of network topology, imposes constraints on node representation at a higher structural level."
"By embedding with A2M, the representation learns the importance of different attributes and captures the semantic similarity between node attributes."
"By encoding initial community membership information and introducing a new modularity function to formulate CMF as a modularity optimization problem, we guide network embedding to explore mesoscopic community structures, ensuring the structural cohesiveness of detected communities."