Conceptual

Kernel-Induced Self-Representation Learning for Concept Drift in Co-evolving Time Series

A representation-learning method for co-evolving time series that models each series as a nonlinear self-representation of the others in a reproducing-kernel Hilbert space, producing a coefficient matrix regularized toward a block-diagonal structure whose blocks are latent concepts. By recomputing this representation over sliding windows and tracking how series migrate between blocks, it identifies concepts, tracks concept drift, and forecasts emerging concepts, and integrates as a module into deep-learning forecasting backbones.