Multiset Feature Learning for Highly Imbalanced Data Classification

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Increasing data volumes lead to more imbalanced datasets.
• High imbalance ratios severely degrade the performance of existing imbalanced learning methods.
• Accurate classification of highly imbalanced data remains a significant challenge.

Purpose of the Study:

• To systematically investigate the problem of highly imbalanced data classification.
• To propose novel methods, Uncorrelated Cost-Sensitive Multiset Learning (UCML) and Deep Metric-based UCML (DM-UCML), to address this challenge.
• To enhance the robustness and performance of classification models on datasets with high imbalance ratios.

Main Methods:

• The Uncorrelated Cost-Sensitive Multiset Learning (UCML) approach constructs multiple balanced subsets via random partitioning and applies multiset feature learning (MFL).
• The Deep Metric-based UCML (DM-UCML) approach integrates generative adversarial networks for subset construction, ensuring similar distribution to the original dataset.
• DM-UCML combines deep metric learning with MFL to handle non-linearity and improve feature discriminability, incorporating a novel discriminant term.

Main Results:

• Experiments on eight traditional and two large-scale imbalanced datasets were conducted.
• The proposed UCML and DM-UCML approaches demonstrated superior performance compared to state-of-the-art methods.
• Both methods exhibited enhanced robustness when dealing with high imbalance ratios.

Conclusions:

• The proposed UCML and DM-UCML methods offer effective solutions for highly imbalanced data classification.
• DM-UCML's integration of deep metric learning and generative adversarial networks significantly improves performance and robustness.
• These novel approaches represent a significant advancement in handling challenging imbalanced datasets.