Linear-exponential loss incorporated deep learning for imbalanced classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Class imbalance is a persistent challenge in machine learning, often leading to poor performance on minority classes.
• Traditional methods struggle with imbalanced data, misclassifying minority samples as majority, with significant real-world implications.

Purpose of the Study:

• To introduce a novel deep learning loss function, DLINEX, to effectively handle class imbalance problems.
• To extend the linear-exponential (LINEX) loss function for multi-class imbalanced classification tasks.

Main Methods:

• The study proposes DLINEX, a multi-class extension of the LINEX loss function, adapted for deep learning.
• DLINEX features an asymmetric geometry, allowing adaptive focus on minority and hard-to-classify samples via a single parameter.
• It simultaneously optimizes between- and within-class diversity by considering instance properties.

Main Results:

• DLINEX achieved 42.08% G-means on CIFAR-10 (imbalance ratio 200), 79.06% G-means on HAM10000, and high F1 scores on DRIVE, CHASEDB1, and STARE datasets.
• Experiments demonstrated DLINEX's superior performance in both image-level and pixel-level imbalanced classification tasks.

Conclusions:

• DLINEX offers an effective solution for class imbalance in deep learning.
• The proposed loss function shows significant improvements in identifying minority and challenging samples across various datasets.