Topic: Heterogeneous Data

We highlight a significant advancement in stream learning with the CAMEL framework, recently accepted as an Oral paper for AAAI 2026 by researchers at the University of Technology Sydney. This innovative framework addresses the complexities of heterogeneous multistream learning, which is prevalent in smart cities and industrial IoT where data distributions shift asynchronously. We integrate a Mixture of Experts architecture that distinguishes between private experts for stream-specific traits and assistance experts for cross-stream collaboration using multi-head attention. To handle unpredictable concept drift, we implement an automated lifecycle management system following a Test-Diagnose-Adapt cycle. By utilizing Maximum Mean Discrepancy for drift detection, our approach enables incremental expansion to capture new concepts while pruning redundant experts to maintain efficiency. This methodology effectively prevents catastrophic forgetting and negative transfer, offering a robust solution for models operating in non-stationary, real-world environments.