GT-EBRB: An Extended Belief Rule Base System Optimized via Clustering Graph Structure and Graph Transformer

The Extended Belief Rule Base (EBRB) serves as a rule-based decision-making system and has been widely applied in classification tasks due to its interpretability and transparency. However, it faces issues of low reasoning efficiency and limited modeling capability when dealing with highdimensional nonlinear data. Previous approaches have improved EBRB by constructing rule indexes and introducing neural network algorithms. Nevertheless, these methods still suffer from overly local aggregation of rule information and instability of the constructed graph structure. To address these issues, we propose an EBRB system optimized based on clustering graph structure and Graph Transformer (GT-EBRB). Specifically, GTEBRB first performs rule clustering on the rule base, employing the Silhouette Coefficient to filter out noisy rule nodes. Subsequently, a probability-based edge connection strategy is designed to capture and represent a more global rule community structure, modeling a more stable clustering graph structure for the rule base. Given that Graph Transformer possesses self-attention mechanisms and global information aggregation capabilities, we thus propose integrating the Graph Transformer algorithm with rule positional encoding for rule representation learning and generating activation weights for rules. Moreover, sparse evidential reasoning is adopted in the inference phase to further enhance inference efficiency and the consistency of activated rules. We conducted comprehensive experiments on 11 UCI classification datasets, and the results demonstrate the superior classification performance of GT-EBRB.