Imbalance classification based on supervised contrastive optimization

Accurate classification under class imbalance remains a significant challenge in machine learning. While deep learning offers potent solutions for classification tasks, purely data-driven models often function as black boxes. The lack of interpretability can obscure the reasoning behind predictions, potentially leading to adverse outcomes. Integrating domain knowledge with contrastive learning presents a promising avenue to address these issues. In this work, we propose a novel prediction model for imbalanced classification based on a Supervised Contrastive Feature Optimization (SCFO) framework. The SCFO framework introduces a knowledge association module to quantify and evaluate identified influence factors, ensuring alignment between the selected features and established domain principles. Subsequently, a supervised contrastive feature module extracts discriminative, statistical representations from the selected feature samples. Our approach is particularly advantageous for imbalanced datasets, as it improves the model's ability to discern minority classes. The effectiveness of our method is validated using real-world datasets. Experimental results demonstrate that SCFO successfully extracts meaningful statistical contrastive representations and delivers significant performance improvements for the target imbalanced classification task.