Learning With Multiclass AUC: Theory and Algorithms.

Area of Science:

• Machine Learning
• Data Science
• Statistical Learning

Background:

• Area Under the ROC Curve (AUC) is a key metric for imbalanced learning and recommender systems.
• Existing AUC optimization methods primarily focus on binary-class problems, neglecting multiclass scenarios.
• Multiclass AUC metrics are crucial for evaluating performance in complex classification tasks.

Purpose of the Study:

• To develop a framework for learning multiclass scoring functions by optimizing multiclass AUC metrics.
• To address the limitations of existing AUC optimization techniques in handling multiclass problems.
• To propose an imbalance-aware generalization error bound for multiclass AUC optimization.

Main Methods:

• Revisiting the M metric, a multiclass extension of AUC, to understand its properties in handling imbalance.
• Proposing an empirical surrogate risk minimization framework for approximate optimization of the M metric.
• Developing acceleration methods for exponential, squared, and hinge loss functions to improve scalability.

Main Results:

• Theoretical analysis shows that optimizing popular surrogate losses leads to the Bayes optimal scoring function asymptotically.
• The proposed framework provides an imbalance-aware generalization error bound, prioritizing minority class performance.
• Experimental results on 11 datasets validate the effectiveness of the proposed multiclass AUC optimization framework.

Conclusions:

• The developed framework effectively optimizes multiclass AUC metrics, offering significant improvements over existing methods.
• The imbalance-aware generalization bound highlights the framework's ability to handle challenging minority classes.
• Acceleration methods enhance the practical applicability and scalability of the proposed approach for real-world datasets.