Handling Class Imbalance in SAGIN Heterogeneous Devices via Location-Slack-Fuzzy Broad Learning System

Abstract

In the space-air-ground integrated network (SAGIN) scenario, due to the integration of remote sensors, unmanned aerial vehicles, and satellite-ground communications, heterogeneous and highly imbalanced data frequently appear, which poses huge challenges to learning algorithms. While broad learning system (BLS) is efficient, its least squares optimization struggles with SAGIN's imbalanced and noisy distributions. To address these problems, we propose an imbalance-aware slack factor fuzzy broad learning system (ISFFBLS). The method introduces position information to guide model training. To enhance the impact of minority class data, we construct a fuzzy weighted least squares classifier that assigns weights to training samples through a fuzzy membership matrix. Then, a dynamic adjustment mechanism evaluates the classification difficulty of each sample and updates its weight accordingly. The position parameter controls the weight distribution of majority class samples. Finally, to ensure the optimality and stability of the classification boundary, we develop an iterative optimization framework to further optimize the slack factor and fuzzy membership until convergence. Experiments on 18 imbalanced datasets show that ISFFBLS performs better than recent imbalanced learning methods, especially in identifying minority class samples.