A novel radial basis function neural network classifier based on three-way decisions

The radial basis function neural network (RBFNN) has gained widespread adoption due to its simple network topology and excellent learning capability. However, the performance of empirical formula-based RBFNN tends to deteriorate when faced with sample imbalance issues. In order to fully harness the potential of RBFNN classifier, inspired by three-way decisions (TWDs) conforming to the logic of human behavior, this paper investigates a novel classifier named TWDs-RBFNN that integrates TWDs and RBFNN. This novel artificial intelligence approach efficiently handles imbalanced samples in imbalanced classification tasks while achieving enhanced performance within amore streamlined topology. TWDs-RBFNN incorporates a TWDs module into the RBFNN framework, which dynamically stores and efficiently trains difficult-to-classify samples for reinforcement learning of intricate sample information and network topology optimization through parameter updates. Drawing upon current advancements in neuroscience and granule computing theory, we introduce gradient harmonizing mechanism to balance samples at the loss function level and handle outliers to avoid parameter bias and improve training efficiency. We also introduce information granules and establish a novel conditional probability formula to safeguard against information loss. The results of numerous experiments on imbalanced datasets demonstrate that TWDs effectively enhances the compactness of the network structure and improve its generalization ability in RBFNN. Specifically, we provide theoretical proof for the convergence of TWDs-RBFNN to analyze their effectiveness.