Generalizing Few-Shot Graph Streams: A Semi-Supervised Community Streams Generation Framework for Community Detection

The goal of community detection is to identify and partition closely related groups within a graph. Previous community detection methods primarily identified communities by learning the topological structure and similar features of the graph through extensive training data. However, they overlook the dynamic interaction patterns between information, making it difficult to adapt to changes in community scenarios and the incorporation of new data streams. To address these issues, a Semi-supervised Community Stream Generative framework, SemiCSG, has been proposed. Our idea is to dynamically expand community information by analyzing the consistency of graph data stream states without requiring a large amount of labeled data. Specifically, we introduce an interactive controller that utilizes generative models to identify the interaction states of subgraph information within graph data streams, considering the steady-state information as belonging to the same community. The subgraph encoding module is used to fill in the semantic gaps in the community structure of the generative model. Additionally, we propose a community streams generation method that employs a domain constraint strategy to guide the interactive controller in explicitly delineating community boundaries, thereby enhancing the model’s adaptability. Extensive experiments demonstrate that our method achieves outstanding performance on real-world datasets in varying scenarios.