Improvement of SMOTE Based on Graph Neighborhood Structure and Latent Feature Learning and Its Application

The SMOTE algorithm is a classical method for addressing imbalanced classification problems. In this paper, we improves the SMOTE algorithm by incorporating ideas from transfer learning. Initially, a k-nearest neighbor (kNN) graph of minority class samples is constructed to capture the structural relationships between samples. The graph convolutional network (GCN) is then employed to model the adjacency matrix and sample features jointly, generating node feature representations that include neighborhood information. These are used to generate samples in the first phase within the feature space; Subsequently, the generated samples are input into an Autoencoder-Generative Adversarial Network (AE-GAN) for a second phase of sample regeneration. The AE component extracts latent features to enhance the learning of features in the generated samples. The samples generated through these two phases are made closer to the real sample distribution. Finally, results on five imbalanced datasets indicate that the proposed method achieves the best classification performance, validating the effectiveness of the algorithm.