An Adaptive Moment estimation method for online AUC maximization.

Area of Science:

• Machine Learning
• Data Mining
• Classification Performance Metrics

Background:

• Area Under the ROC Curve (AUC) is a key metric for classification performance, crucial for developing AUC maximization algorithms.
• Traditional batch learning methods for AUC maximization are inefficient and not scalable for large datasets.
• Existing online AUC maximization algorithms may fail to converge optimally due to learning rate issues.

Purpose of the Study:

• To propose a novel online algorithm for AUC maximization that addresses the convergence and scalability limitations of existing methods.
• To enhance the efficiency and effectiveness of online AUC maximization through adaptive learning rate strategies.

Main Methods:

• Introduced AdmOAM (Adaptive Moment estimation for Online AUC Maximization).
• Employed gradient moment estimation to accelerate convergence.
• Mitigated rapid learning rate decay for improved stability.
• Established the regret bound for the proposed algorithm.

Main Results:

• AdmOAM demonstrates accelerated convergence compared to existing online methods.
• The algorithm effectively mitigates issues related to fixed or rapidly decaying learning rates.
• Extensive experiments validate the effectiveness and efficiency of AdmOAM.

Conclusions:

• AdmOAM offers an effective and scalable solution for online AUC maximization.
• The proposed method overcomes key limitations of prior online learning approaches for AUC optimization.
• AdmOAM shows significant promise for large-scale classification performance enhancement.