Task-Feature Collaborative Learning with Application to Personalized Attribute Prediction.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-task learning (MTL) enhances performance with limited data by sharing knowledge across tasks.
• Negative transfer, where knowledge sharing with dissimilar tasks degrades performance, is a key challenge in MTL.
• Existing methods often overlook feature-level transfer and focus solely on task correlations.

Purpose of the Study:

• To propose a novel multi-task learning method, task-feature collaborative learning (TFCL), to address negative transfer.
• To alleviate negative transfer by considering collaborative knowledge sharing across both features and tasks.
• To improve overall model performance by mitigating knowledge sharing issues.

Main Methods:

• Developed a base model incorporating a heterogeneous block-diagonal structure regularizer for collaborative feature and task grouping.
• Designed an optimization method for the proposed TFCL model.
• Extended the base model to handle overlapping features and differentiate hard tasks, applying it to personalized attribute prediction.

Main Results:

• Theoretical analysis guarantees global convergence and block-diagonal structure recovery.
• Experimental results on simulated and real-world datasets validate the effectiveness of TFCL.
• The method successfully demonstrated improved performance in personalized attribute prediction.

Conclusions:

• TFCL effectively mitigates negative transfer in multi-task learning by jointly considering features and tasks.
• The proposed method offers theoretical guarantees and practical applicability.
• TFCL represents a significant advancement in multi-task learning, particularly for complex problems like personalized attribute prediction.