Mitigating data center bias in cancer classification: Transfer bias unlearning and feature size reduction via...

Bias in the decision-making processes of trained deep models poses a significant threat to their reliability. Such bias can lead to overoptimistic results on observed data while compromising generalization to unseen datasets. Training data may contain hidden patterns related to task-irrelevant attributes, such as data centers, causing models to exploit these unintended correlations rather than learning the main task. This results in biased predictions that favor certain attributes. To address this issue, we propose an unlearning approach based on Conflict-of-Interest-Free Multi-Objective Optimization, designed to train an unlearning layer that explicitly reduces reliance on irrelevant patterns. Our method aims to minimize the gap between internal accuracy (evaluated on data centers seen during training) and external accuracy (evaluated on entirely unseen data centers) caused by biased model behavior. As a case study, we investigate how data center-specific signatures embedded in cancerous features can lead to misleadingly high internal performance and a significant drop in performance on test samples from external data centers. By evaluating various methods and objective functions, our proposed approach achieves strong generalizability on external validation data by jointly reducing feature dimensionality and excluding conflict-of-interest samples during the k-Nearest Neighbor (KNN) searching process. We compare our method against multi-task and adversarial learning approaches for bias mitigation. Results show that our method outperforms others in narrowing the internal-external performance gap while also improving external validation accuracy. To ensure robustness, we conducted experiments using k-fold cross-validation across k different data centers, further validating the generalizability of our approach. Although this study focuses on cancer-related features and data center biases, the proposed method is model-agnostic and can be applied to any biased feature set extracted by a deep learning model.