TCPD: A Transformer-Based Multi-Scale Framework for Dynamic Network Change Point Detection

Dynamic network change point detection is an essential research topic due to its broad applications in social networks, transportation systems, and biological regulatory networks. However, accurately detecting changes in evolving networks is challenging because change points are rare, labels are costly to obtain, and many existing methods either rely on strong modeling assumptions or have limited capacity to capture multi-scale spatio-temporal dependencies. In this paper, we propose TCPD, a Transformer-based multi-scale framework for change point detection in dynamic networks. TCPD first constructs a multi-scale representation that integrates global and local features of each graph snapshot, enabling a more comprehensive characterization of evolving graph patterns. A two-stage Transformer architecture then models temporal evolution and incorporates a focus score as an additional signal to highlight time slices with stronger changes, improving the reliability of change point detection under noisy dynamics. Finally, hypersphere learning shapes a compact latent region for normal time slices and encourages changed slices to lie outside this region, enabling end-to-end training without change point labels. Extensive experiments on one synthetic and four real-world dynamic networks show that TCPD consistently outperforms representative baselines in terms of precision, recall, and F1 score, demonstrating its effectiveness for dynamic network change point detection.