STGCM: Sign-aware triple graph collaborative model for recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommendation systems, as a core technology for mitigating information overload, play a crucial role across many fields. In recent years, graph neural networks (GNNs) have significantly improved recommendation performance by modeling user-item interactions as a graph structure. However, most existing sign-aware methods directly partition the user-item interaction graph into separate positive and negative graphs for modeling. While this local modeling approach can distinguish different feedback types, it disrupts the global topology of the original graph. This impairment exacerbates data sparsity and weakens the propagation of signed information. To address this, we propose STGCM, a S ign-aware T riple G raph C ollaborative M odel for Recommendation. STGCM employs a multi-perspective graph convolution embedding learning framework, which preserves the complete topology of the original graph while thoroughly excavating local preference features from both positive and negative feedback. Building on this, a two-stage hierarchical attention fusion mechanism is designed to achieve a global-information-supplemented, synergistic fusion of multiple perspectives, thereby enhancing the model’s information complementarity capability. The experimental results provide compelling evidence that STGCM outperforms state-of-the-art baselines across three real-world datasets. The source code is publicly accessible at https://github.com/Dongqin-Wang/STGCM .