New AI Method Adapts to Dynamic Data Shifts and Imbalance

Machine learning models often struggle when deployed in real-world environments where data distributions continuously shift, a problem known as domain shift. Test-Time Adaptation (TTA) aims to enable models to adapt online to unlabeled test data. While recent TTA methods have begun to tackle continual domain shifts, they frequently overlook another critical real-world challenge: class imbalance in dynamic data streams.This research introduces a new method called Balanced and Prototype-Guided Test-Time Adaptation (BP-TTA). BP-TTA is designed to simultaneously handle both continual domain shifts and class imbalance. It achieves this through two main mechanisms. First, it constructs balanced adaptation batches by combining current samples with high-confidence historical instances, which helps to mitigate bias towards dominant classes and stabilize online updates.Second, BP-TTA maintains and evolves class prototypes during inference. It then uses the similarity to these prototypes as a constraint during model adaptation. This prototype-guided approach enhances the reliability of pseudo-labels and further improves the stability of online updates, even under persistent domain shifts. Extensive experiments confirm that BP-TTA consistently outperforms existing state-of-the-art TTA methods in dynamic test-time streaming settings.