An Adaptive Deep Metric Learning Loss Function for Class-Imbalance Learning via Intraclass Diversity and Interclass...

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep metric learning (DML) effectively extracts discriminant features but struggles with class-imbalance learning (CIL) issues like data scarcity and density.
• Existing DML and CIL losses fail to simultaneously address feature overlapping, data scarcity, and data density, leading to misclassification.

Purpose of the Study:

• To propose a novel loss function, intraclass diversity and interclass distillation (IDID) loss with adaptive weight, capable of mitigating DML and CIL challenges concurrently.
• To enhance feature representation by generating diverse intra-class features and preserving inter-class semantic correlations.

Main Methods:

• Developed the IDID loss function, which promotes intra-class diversity irrespective of sample size to combat data scarcity and density.
• Implemented learnable similarity within IDID loss to maintain semantic correlations between classes while pushing dissimilar classes apart, reducing overlapping.
• Introduced an adaptive weight mechanism to balance the contributions of diversity and distillation components.

Main Results:

• IDID loss successfully mitigates data scarcity, data density, and feature overlapping simultaneously, outperforming traditional DML and CIL losses.
• The proposed method generates more diverse and discriminant feature representations, leading to improved generalization ability.
• Experiments on seven public datasets demonstrated superior performance in G-mean, F1-score, and accuracy, with significant improvements on imbalanced classes.

Conclusions:

• The IDID loss offers a unified solution for deep metric learning in the presence of class imbalance, data scarcity, and density.
• This approach eliminates the need for time-consuming hyperparameter fine-tuning, simplifying practical application.
• IDID loss provides a robust and effective method for enhancing classification performance in challenging real-world scenarios.