New one-shot clustering method uses eigenspaces for personalized system ID

Researchers Abdulmoneam Ali, Dipankar Maity, and Ahmed Arafa have introduced a novel approach to clustered system identification that eliminates the need for iterative training. Their method, detailed in a recent arXiv preprint, leverages the structure of locally observed data to group systems with shared dynamics. Instead of relying on training-based cluster assignment, which can be sensitive to learning uncertainty and model initialization, the proposed technique uses a one-shot, training-free clustering method. Each system estimates a local state covariance matrix, and cluster identities are inferred by measuring the alignment between the leading covariance eigenspaces of different systems.

The paper provides a mathematical interpretation of the similarity score and develops a finite-sample analysis that characterizes how covariance estimation error induces eigenspace perturbations. The authors derive a probability bound for pairwise false merges and a global clustering success guarantee. Numerical experiments demonstrate that this eigenspace-based clustering method effectively identifies systems with shared dynamics, leading to lower personalized model-estimation error compared with training-based clustering and non-clustered baselines. The approach has potential applications in control systems, signal processing, and machine learning where system heterogeneity is common.