Tackling Class-imbalanced Learning Issues Based on Local Neighborhood Information and Generative Adversarial Networks

Sensors are extensively used to collect data from systems.For example, in intelligent manufacturing, accelerometers are employed to gather process inputs and outputs in real time.However, abnormal events represent a small portion of the data, posing challenges for machine learning algorithms.Most algorithms lack the ability to account for equivalent sample representations.When addressing class imbalance, the widely used synthetic minority oversampling technique (SMOTE) has limitations.SMOTE does not consider the relative distributions between minority and majority class samples, potentially creating minority samples within the majority distribution.Additionally, its linear approach may miss nonlinear relationships among sample attributes.To overcome these issues, we propose a novel data augmentation method based on local neighborhood information and generative adversarial networks (GANs).Our approach first leverages density-based spatial clustering of applications with noise to identify minority class noises and then computes neighborhood types for minority samples using the k-nearest neighbors algorithm.On the basis of these neighborhood types (safe or dangerous), we create synthetic samples using GANs and bootstrapping.Evaluation on ten publicly available imbalanced datasets shows that our proposed method surpasses all other approaches for the majority of the datasets.