تحليل تفسيرپذيري و پيشبيني پذيري با رويكرد رگرسيون خطي قطعه اي خودكار

Balancing Interpretability an‎d Predictiveness in Learning: An Automatic Piecewise Linear Regression Approach

رياضي كاربردي و علوم كامپيوتر

With the increasing adoption of machine learning algorithms in sensitive application domains, achieving a balance between interpretability an‎d predictive accuracy has become a major chal-lenge. Complex models often provide high predictive performance; however, their decision-making processes are difficult to interpret. In contrast, simpler models are generally more interpretable but may suffer from lower predictive accuracy. In this study, the Automatic Piecewise Linear Regression algorithm is investigated as a method for simultaneously achieving interpretability an‎d predictive performance. First, the theoretical foundations an‎d underlying mechanism of the algorithm are examined. Subsequently, its perfor-mance is compared with several established methods, including Multivariate Adaptive Regression Splines (MARS), Ran‎dom Forest, Light Gradient Boosting Machine, Generalized Additive Model Boosting, an‎d Generalized Linear Model Boosting. The eva‎luation is conducted on both simulated datasets an‎d several real-world datasets. The re-sults demonstrate that APLR achieves competitive performance in most scenarios an‎d outperforms the other methods in many cases. By employing piecewise linear basis functions, the algorithm maintains an interpretable model structure while effectively capturing nonlinear relationships an‎d interactions among variables. The findings suggest that APLR is a promising approach for regres-sion problems in which both interpretability an‎d predictive accuracy are of primary importance.

5

167933