AUTHOR=Pena-Campos Johan , Patino Diego , Ocampo-Martinez Carlos , Ramos-Fernández Julio C. , Salas-Brown Margot , Caicedo Alexander 
  
TITLE=Retrieving interpretability to support vector machine regression models in dynamic system identification
  
JOURNAL=Frontiers in Artificial Intelligence
  
VOLUME=Volume 8 - 2025
  
YEAR=2025
  
URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1706566
  
DOI=10.3389/frai.2025.1706566
  
ISSN=2624-8212
  
ABSTRACT=Black-box models, particularly Support Vector Machines (SVM), are widely employed for identifying dynamic systems due to their high predictive accuracy; however, their inherent lack of transparency hinders the understanding of how individual input variables contribute to the system output. Consequently, retrieving interpretability from these complex models has become a critical challenge in the control and identification community. This paper proposes a post-hoc functional decomposition algorithm based on Non-linear Oblique Subspace Projections (NObSP). The method decomposes the output of an already identified SVM regression model into a sum of partial (non)linear dynamic contributions associated with each input regressor. By operating in the non-linear feature space, NObSP utilizes oblique projections to mitigate cross-contributions from correlated regressors. Furthermore, an efficient out-of-sample extension is introduced to improve scalability. Numerical simulations performed on benchmark Wiener and Hammerstein structures demonstrate that the proposed method effectively retrieves the underlying partial nonlinear dynamics of each sub-system. Additionally, the computational analysis confirms that the proposed extension reduces the arithmetic complexity from 𝒪(N3) to 𝒪(Nd2), where d is the number of support vectors. These findings indicate that NObSP is a robust geometric framework for interpreting non-linear dynamic models, offering a scalable solution that successfully decouples blended dynamics without sacrificing the predictive power of the black-box model.