Collinearity-Resilient Sparse Identification of Nonlinear Dynamics via Cluster-Stability and Representative Selection
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Sparse identification of nonlinear dynamics (SINDy) exhibit structural fragilely when the candidate library contains highly correlated functions, as standard feature-wise selection tends to scatter selection frequencies across near-multicollinear terms and can therefore omit truly active variables. To overcome this limitation, in this paper, we propose a collinearity-aware SINDy framework that groups correlated candidates, evaluates cluster-level stability over repeated sparse-regression fits, and disambiguates each retained group by selecting a unique representative through independent-trajectory validation, followed by coefficient re-estimation on the fixed support. Experiments show that the proposed cluster-level approach achieves 85.0% exact recovery at a derivative-perturbation level of 0.30, compared with 67.7% for tuned STRidge and 35.3% for tuned feature-wise stability selection. Its applicability is further validated on a damped nonlinear pendulum, where 100.0% system-level exact recovery was achieved for the highly correlated terms and . Collectively, these results demonstrate that cluster-level stability and independent-trajectory disambiguation can improve the structural recovery in highly collinear candidate libraries.
- Primary area:
- learning on time series and dynamical systems