Cross-Modal Metric Learning for AUC Optimization.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Cross-modal metric learning (CML) aims to learn distance functions for matching data across different modalities.
• Existing CML methods often struggle with imbalanced intraclass and interclass sample pairs, leading to suboptimal performance.
• The area under the receiver operating characteristic curve (AUC) is a more robust performance metric for imbalanced datasets.

Purpose of the Study:

• To develop a novel CML method that directly optimizes AUC for improved cross-modal data matching.
• To extend the method for partial AUC (pAUC) optimization, focusing on performance within specific false positive rate (FPR) ranges.
• To enhance the comparability of samples across different modalities.

Main Methods:

• A new CML approach is proposed that directly maximizes AUC.
• The method is extended to optimize partial AUC (pAUC) for targeted performance evaluation.
• The problem is formulated as a log-determinant regularized semidefinite optimization problem.
• A minibatch proximal point algorithm is developed for efficient optimization and experimentally verified for stability.

Main Results:

• The proposed CML method demonstrates effectiveness and significant improvements over existing approaches on various datasets, including face recognition.
• Direct AUC maximization effectively handles imbalanced sample pairs inherent in cross-modal learning.
• Partial AUC (pAUC) optimization shows competitive results for metrics like Rank-1 and verification rate at low FPRs (e.g., 0.1%).

Conclusions:

• The novel CML approach effectively addresses data imbalance by directly optimizing AUC.
• Partial AUC optimization offers a valuable strategy for applications requiring high performance within specific FPR thresholds.
• The developed algorithm is stable and efficient, paving the way for practical applications in cross-modal retrieval and recognition.