Taming polarized fitting: BLINEX-Pcomp with asymmetric risk penalty for robust Pcomp classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Pcomp classification offers a novel paradigm for learning with inexact supervision.
• Existing methods struggle with polarized fitting due to inadequate handling of empirical risk sign differences.

Purpose of the Study:

• To propose a new model, BLINEX-Pcomp, that addresses limitations in Pcomp classification.
• To improve the balance between overfitting and underfitting risks in pairwise sample learning.

Main Methods:

• Introduced the BLINEX-Pcomp model using a bounded linear-exponential function for differentiated risk penalties.
• Developed a multi-view version (MV-BLINEX-Pcomp) integrating multi-view features.
• Designed a dual-stage solver for training the MV-BLINEX-Pcomp model.

Main Results:

• BLINEX-Pcomp effectively balances pairwise-level risks, shifting focus to challenging samples.
• MV-BLINEX-Pcomp demonstrated enhanced performance by incorporating multi-view features.
• Theoretical verification confirmed MV-BLINEX-Pcomp degrades to BLINEX-Pcomp with single-view features.

Conclusions:

• The proposed BLINEX-Pcomp and MV-BLINEX-Pcomp models offer effective solutions for Pcomp classification.
• These methods successfully tackle challenges related to empirical risk differences in pairwise learning.
• Numerical results validate the superiority of the proposed approaches in comparative experiments.