Sequential Label Enhancement.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Label distribution learning (LDL) addresses ambiguous tasks by quantifying label relevance.
• Current LDL methods often overlook the sequential nature of label distribution recovery.
• Obtaining accurate label distributions is challenging due to high costs and quantification difficulties.

Purpose of the Study:

• To propose a novel approach for label distribution recovery that incorporates sequential decision-making.
• To develop a method that better mimics human cognitive processes in annotating label distributions.
• To enhance the performance of LDL by leveraging reinforcement learning for sequential label enhancement.

Main Methods:

• Formulated label distribution recovery as a sequential decision process named sequential label enhancement (Seq_LE).
• Employed a reinforcement learning (RL) agent to serially map discrete labels and their description degrees.
• Designed a joint reward function to optimize the RL agent's decision-making policy.

Main Results:

• Seq_LE demonstrated superior performance across 16 diverse LDL datasets.
• The proposed sequential approach significantly improved label distribution recovery compared to state-of-the-art methods.
• Experimental results validated the effectiveness of the RL-based sequential label enhancement strategy.

Conclusions:

• Sequential label enhancement (Seq_LE) offers a more intuitive and effective paradigm for LDL.
• The RL-driven sequential approach provides a robust solution for ambiguous classification tasks.
• This work opens new avenues for research in sequential modeling for machine learning problems.