Pré-Publication, Document De Travail Année : 2026

Linear Modeling, Random Forests and Interpretability: SPLIM, Stable Piecewise Linear Interpretable Model

Résumé

Many machine learning algorithms, such as generative AI models, neural-networks or random forests, are widely used for their strong predictive performances. However, their 'black-box' nature limits their interpretability. Yet, in many applications, interpretability or simplicity of a model is at least as important as its predictivity. Recently, the SIRUS algorithm addressed this issue by extracting from a random forest a set of simple, stable and interpretable decision rules. However, trees in random forests lack flexibility and extensions such as RaFFLE improved it by allowing linear models within tree nodes, leading to capture more complex input-output relationships. In this paper, we introduce SPLIM (Stable Piecewise Linear Interpretable Model), a method designed to combine the modeling flexibility of RaFFLE with the interpretability principles of SIRUS within a unified framework. SPLIM constructs a forest of trees allowing for regression models in their nodes and derives from it an interpretable rule-based model that incorporates both partitioning effects and their associated linear components. Through a simulation study, SPLIM's predictive accuracy, variable selection and stability performances are investigated for a variety of structural scenarii. The results show that SPLIM achieves predictive performance comparable to RaFFLE, while providing the same interpretable representation of the relationship between predictors and responses as SIRUS. Experiments on real datasets further illustrate SPLIM's practical relevance.

Fichier principal
Vignette du fichier
paper_hal.pdf (2.7 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05594273 , version 1 (16-04-2026)

Licence

Identifiants

  • HAL Id : hal-05594273 , version 1

Citer

Lucie Brolon, Fabien Crauste, Marie Perrot-Dockès. Linear Modeling, Random Forests and Interpretability: SPLIM, Stable Piecewise Linear Interpretable Model. 2026. ⟨hal-05594273⟩
18 Consultations
9 Téléchargements

Partager

  • More