Automated Class Imbalance Learning via Few-shot Multi-Objective Bayesian Optimization with Deep Kernel Gaussian Processes

Automated Class Imbalance Learning (AutoCIL) is an emerging paradigm that leverages Combined Algorithm Selection and Hyperparameter Optimization (CASH) to automate the configuration of resampling strategies and classifiers for imbalanced classification tasks. Existing AutoCIL methods focus solely on single-objective optimization. However, real-world applications often involve multiple, conflicting objectives—such as predictive performance and computational cost—that must be jointly optimized. Ignoring such trade-offs limits the adaptability and practicality of current methods. In this work, we propose a novel approach called AutoCIL-FMOBO (AutoCIL via Few-shot Multi-Objective Bayesian Optimization). Specifically, we design meta-learned deep kernel Gaussian process surrogates trained on a meta-dataset constructed from pre-evaluated results obtained by running configurations in the search space on class-imbalanced datasets. Then, these surrogate models with prior optimization knowledge are combined with the Expected Hypervolume Improvement (EHVI) acquisition function in a Bayesian optimization framework to efficiently discover Pareto-optimal configurations for the target task, which enables AutoCIL-FMOBO to jointly optimize key components, such as resampling methods, classifiers, and their hyperparameters, under a multi-objective setting. Experimental results on 15 real-world class-imbalanced datasets demonstrate that our approach outperforms baselines in both effectiveness and sample efficiency, while maintaining generalization across tasks and achieving competitive performance under a multi-objective setting.