Heterogeneous graph transformer with slot-based message passing for node classification

As a widely utilized data structure, heterogeneous graphs can accurately model complex structures and multi-typed relationships, finding extensive applications in social network analysis, recommendation systems, knowledge graphs, and biomedical analysis. Node classification supports these applications by identifying entity categories or functional roles. However, during the message-passing process, mapping different types of nodes into a unified representation space often leads to cross-type feature confusion, resulting in the loss of type-specific semantic information. To address this issue, this paper proposes the HGT-Slot model. First, the model employs a type-aware Slot initialization to construct type-specific representation spaces for different node types, preserving their unique semantic attributes. Subsequently, a Slot attention mechanism is employed during the message-passing phase to preserve heterogeneous semantic information through different Slot channels. Meanwhile, the Transformer architecture is leveraged to enhance interactions among node representations, thereby improving the representation capability of the model. Finally, a Slot aggregation mechanism integrates the Slot representations for local prediction, which is fused with the global prediction from the Transformer branch for classification. Experiments on three datasets show that HGT-Slot achieves the best mean Macro-F1 and Micro-F1 scores among the compared methods for node classification.