Contrastive learning for multi-layer network community detection via learnable network augmentation

Community detection in multi-layer networks is crucial for revealing the functions of entities and understanding their connections across dimensions. However, existing semi-supervised methods often rely on manual labels, leading to the significant computational overhead in networks with complex structures. Moreover, unsupervised and self-supervised methods usually struggle to integrate the intra-layer and inter-layer features, as well as the local and global features of networks, resulting in the limited accuracy. To address these challenges, this paper proposes a self-supervised Network Augmentation Contrastive Constraint (NACC) method for multi-layer network community detection. Leveraging the ideas of network augmentation and contrastive learning, NACC detects the community structure based on the rich features contained in datasets. Specifically, NACC first integrates the intra-layer and inter-layer features of the multi-layer network to generate a learnable feature-augmented network. Then, it encodes the node and topology features, capturing both the local and global features, and generating the low-dimensional node representations for multi-layer and augmented networks. Moreover, the contrastive learning among different layers is proposed to train the above node representations, further enhancing the fusion of features. Finally, consense communities are detected based on the trained node representation. Extensive experiments demonstrate the performance of NACC in handling networks with numerous layers and complex structures, showcasing its reliability in real-world applications.