Heteroscedastic-embedded Ensemble for Imbalanced Massive Data Classification

Abstract The imbalanced learning methods aim to learn the unbiased models from massive class-imbalanced datasets. However, due to the uncertainty of data distributions affected by noise and borderline samples, the models based on heuristic assumptions or meta learning often suffer from the lack of stability and practicability for real-world classification tasks which often uses the skewed dataset of low-quality. In order to effectively cope with these issues, in this work, we propose a novel heteroscedastic-embedded ensemble (HEE) for imbalanced massive data classification. We first design an effective task sensing strategy with data-dependent heteroscedastic to adaptively guide the sampler’s focus on informative samples, which makes the learning method much more robust for noisy data. Then, similar to the learning style of human beings, i.e., easy to difficult, the HEE framework gradually builds a strong ensemble classifier through interactions between harmonizing data resampling and model-training. The simulation results and analysis on both synthetic and real-world tasks demonstrate the effectiveness, robustness, and transferability of the proposed method.