A-SFS: Semi-supervised feature selection based on multi-task self-supervision

Feature selection is an important process in machine learning. It builds interpretable and robust model by selecting the features that contribute most to prediction target. However, mature feature algorithms, including supervised semi-supervised, fail fully exploit complex potential structure between features. We believe these structures are very for process, especially when labels lacking data noisy. To this end, we innovatively introduces a deep learning-based self-supervised mechanism into problems, namely batch-Attention-based Self-supervision Selection(A-SFS). Firstly, multi-task autoencoder designed uncover hidden structural among with support of two pretext tasks. Guided integrated information from multi-self-supervised learning model, batch-attention generate weights according batch-based patterns alleviate impacts introduced handful noisy data. This method compared 14 major strong benchmarks, LightGBM XGBoost. Experimental results show A-SFS achieves highest accuracy datasets. Furthermore, design significantly reduces reliance on labels, only 1/10 labeled needed achieve same performance as those state art baselines. Results also missing