Optimizing Two-Way Partial AUC With an End-to-End Framework.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Area Under the ROC Curve (AUC) is a standard machine learning performance metric.
• Two-way Partial AUC (TPAUC) offers a more refined evaluation by focusing on specific True Positive Rate (TPR) and False Positive Rate (FPR) regions.
• Existing TPAUC metrics face challenges with optimization and consistency.

Purpose of the Study:

• To address the optimization challenges of TPAUC in machine learning.
• To develop a novel framework for efficient end-to-end stochastic training of TPAUC.
• To provide theoretical guarantees for the proposed optimization approach.

Main Methods:

• Development of a generic framework for constructing surrogate optimization problems for TPAUC.
• Utilizing deep learning for efficient end-to-end training.
• Theoretical analysis to bound the original TPAUC objective and ensure generalization.

Main Results:

• The proposed framework enables gradient-based optimization for TPAUC.
• Theoretical analysis confirms the surrogate objective provides an upper bound for TPAUC.
• Empirical studies demonstrate the framework's effectiveness on benchmark datasets.

Conclusions:

• The novel framework successfully optimizes TPAUC, overcoming previous limitations.
• The approach facilitates high TPAUC performance with good generalization.
• This work presents a significant advancement in optimizing advanced AUC metrics for machine learning.