Online Learning from Heterogeneous Data Streams
Learning under irregular, drifting, and heterogeneous streaming conditions
Real-world data streams rarely arrive clean. Observations may be incomplete, the underlying distribution may shift as time passes, and different sources may not share a common feature space. This project develops an online learning framework that treats these three conditions together rather than one at a time.
Key activities:
- Addressed irregular streams with incomplete observations through granular computing over a latent buffer representation.
- Handled concept drift with granular-ball construction and weighted attenuation, so outdated structure decays as the stream advances.
- Integrated heterogeneous sources through an ensemble that adaptively weights the contribution of each representation.
Manuscript: Online Learning from Irregular, Drifting, and Heterogeneous Streaming Data — Di Wu, Zhili Wang, Xin Luo, Yi He, Guoyin Wang, Shuyin Xia. Under review at IEEE/CAA Journal of Automatica Sinica.