Curiosity-driven cooperation for long-tailed multi-label learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Long-tail imbalance is a significant challenge in multi-label learning, where common labels overshadow rare ones.
• This imbalance leads to poor performance on under-represented classes crucial for practical applications.

Purpose of the Study:

• To address the long-tail imbalance problem in multi-label learning.
• To develop a novel framework that improves prediction accuracy for rare labels.

Main Methods:

• Introduced the Curiosity-Driven Game-Theoretic Multi-Label Learning (CD-GTMLL) framework.
• Cast the multi-label learning task as a cooperative potential game with 'players' sharing global accuracy and earning curiosity rewards.
• Curiosity rewards are based on label rarity and inter-player disagreement to boost gradients for under-represented tags.

Main Results:

• Demonstrated state-of-the-art performance gains on conventional and extreme-scale datasets.
• Achieved up to +4.3% Rare-F1 and +1.6% P@3 improvements over strong baselines.
• Ablation studies revealed emergent division of labor and faster consensus on rare classes.

Conclusions:

• CD-GTMLL offers a principled and scalable approach to enhance long-tail robustness in multi-label prediction.
• The curiosity-driven game-theoretic method effectively injects gradient on rare labels without requiring hand-tuned class weights.