PA-FEAT: Fast Feature Selection for Structured Data via Progress-Aware Multi-Task Deep Reinforcement Learning

Feature selection is an effective technique for structured data analytics, aiming to eliminate redundant features and irrelevant features for downstream tasks (e.g., classification). With the deepening of data-driven decision-making applications in various industries, the demand for real-time structured data analysis is constantly increasing. At this time, high requirements are placed on the time cost of feature selection. However, existing feature selection methods may easily fall into the dilemma of efficiency and effectiveness when faced with this situation due to the huge feature space. In this paper, we study a novel fast feature selection scenario, which is to generalize the knowledge of feature selection from historical structured data analytics tasks (seen tasks) and then quickly apply it to the process of feature selection for future structured data analytics tasks (unseen tasks). We propose a novel Progress-Aware multi-task deep reinforcement learning method for Fast fEAture selecTion (PA-FEAT), which makes full use of various progress-related information generated during the knowledge generalization process to achieve efficiency and effectiveness simultaneously. Extensive results on eight real-world datasets show that PA-FEAT consistently outperforms eight baselines in terms of efficiency and effectiveness.