A meta-framework for multi-label active learning based on deep reinforcement learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-label Active Learning (MLAL) enhances classifier performance with reduced annotation effort by selecting valuable examples.
• Current MLAL methods rely on manually designed evaluation strategies, which can be dataset-specific and suboptimal.
• The effectiveness of MLAL is often hindered by label correlation and data imbalance issues.

Purpose of the Study:

• To propose a generalizable Deep Reinforcement Learning (DRL) based evaluation method for Multi-label Active Learning (MLAL).
• To develop a meta-framework enabling the application of learned DRL strategies to unseen datasets.
• To address label correlation and data imbalance challenges within the MLAL framework.

Main Methods:

• A Deep Reinforcement Learning (DRL) model was developed to automatically learn data evaluation strategies for MLAL.
• A self-attention mechanism and a tailored reward function were integrated into the DRL model.
• The proposed DRL method was evaluated on multiple datasets within a meta-learning framework.

Main Results:

• The DRL-based MLAL method demonstrated comparable performance against existing state-of-the-art approaches.
• The model showed effectiveness in generalizing learned strategies to unseen datasets.
• The integrated self-attention and reward function successfully addressed label correlation and data imbalance.

Conclusions:

• The proposed DRL-based MLAL framework offers a generalizable and effective approach to data selection.
• This method reduces the need for manual feature engineering in MLAL evaluation.
• The DRL approach shows promise for improving annotation efficiency and classifier performance in multi-label learning scenarios.