In recent years, entity matching methods based on graph neural networks have significantly improved the ability of entity structure representation through multilayer neighborhood aggregation. However, such methods still suffer from oversmoothing and noise diffusion in local structures, as well as insufficient global topological consistency and limited geometric expression. Existing methods usually rely on local aggregation to obtain structural representations, making it difficult to explicitly model the global topological patterns and hierarchical structures between entities. Especially in heterogeneous or cross‐lingual graphs, traditional Euclidean embedding spaces cannot fully represent complex semantic hierarchies and multiscale geometric relationships. To this end, this paper proposes a Triview Dual‐space Contrastive Perception Matching (TriDCPM) method, which unifies the modeling of local correlations, global topologies, and cross‐graph consistency under a multiview representation and geometric collaborative learning framework. Specifically, TriDCPM constructs a triview framework consisting of local, global, and cross‐graph views and simultaneously learns multiscale entity representations in both Euclidean and hyperbolic geometric spaces. A global structure enhancement module based on singular value decomposition (SVD) is adopted to extract key topological patterns, and a gated residual unit (GRU) is introduced to alleviate noise propagation and oversmoothing. In the local encoding stage, multilayer attention aggregation and a degree‐aware relation fusion mechanism are employed to further enhance heterogeneous neighborhood and relational semantic features. Finally, a dual‐level contrastive consistency learning mechanism is adopted to jointly optimize feature consistency between the local–global levels and the Euclidean–hyperbolic spaces, achieving collaborative perception and discriminative unification of multiview representations. Experimental results on four public benchmark datasets demonstrate that the proposed method significantly outperforms existing structure‐driven entity alignment approaches in terms of Hit@1, MRR, and other metrics, with particularly outstanding performance in relation‐heterogeneous and cross‐lingual scenarios. Further ablation experiments and visualization analysis verify the effectiveness and stability of the proposed method in global structure modeling, noise suppression, and multispace contrastive optimization.
Tan et al. (Thu,) studied this question.