DMHG-CD

Firstly, text and image unimodal features are extracted using RoBERTa and CLIP ViT models, and combined with bidirectional cross modal attention mechanism to achieve semantic collaborative fusion and user level feature aggregation; Secondly, construct a dynamic heterogeneous graph snapshot covering 5 types of social relationships, and use relationship specific attention and gated loop units to dynamically update heterogeneous social features; Finally, by capturing long-range evolutionary dependencies through temporal Transformer and combining temporal attention pooling to focus on critical stages of behavior mutations, the classification results are output.