Paper accepted at the SCL Workshop @ ECML PKDD 2026

29 May 2026

We are excited to share our new paper, “Inference-Time Prior Adaptation for Imbalanced Data Streams under Concept Drift,” accepted at the Streaming Continual Learning Workshop, co-located with ECML PKDD 2026.

This work is authored by Tobias Callies, Prof. Eirini Ntoutsi, and Prof. Arthur Zimek from the University of Southern Denmark.

Data-stream classifiers must continuously respond to changes in incoming data while maintaining good performance for underrepresented classes. Existing approaches frequently address these challenges through complex memory-management strategies, such as removing, reorganizing, resampling, or synthetically generating stored instances.

In this paper, we introduce PAI-KNN, a k-nearest-neighbour classifier that instead adapts its decisions at inference time.

Key takeaways:

  • Adaptation without restructuring memory. PAI-KNN retains a simple sliding-window buffer and does not continuously modify its stored instances.

  • Dynamic prior adjustment. The method estimates recent class proportions and uses them to adapt its decision rule at inference time.

  • Designed for two simultaneous challenges. PAI-KNN jointly addresses concept drift and changing class imbalance.

  • Competitive performance. Across a wide range of synthetic data streams, the method performs competitively with state-of-the-art approaches and is particularly effective when class imbalance and concepts change dynamically.

  • A complementary form of adaptation. Inference-time adjustment enables a fast response to changing stream conditions, while memory-based approaches can remain useful when information must be retained over longer periods.

The results show that effective stream adaptation does not always require increasingly complex memory operations. Adjusting predictions at inference time can provide a simple and responsive alternative for highly dynamic environments.