A comparative evaluation of the enhanced PLS-Tree algorithm with multiple latent score vectors

Abstract

Multicollinearity and heterogeneity are prevalent challenges in the analysis of process industry datasets, necessitating algorithms capable of addressing both simultaneously. The partial least squares (PLS)-Tree algorithm, which integrates PLS regression with decision tree methodologies, stands out by concurrently addressing data heterogeneity and improving predictive performance. However, the PLS-Tree algorithm remains underexplored compared to other machine learning approaches. This study delves into the intricacies of the PLS-Tree algorithm, utilizing synthetic data that mirrors the complexity of real-world process industry scenarios characterized by high collinearity and clustering. This paper further enhances the original PLS-Tree framework by introducing multiple latent score vectors, with the objective of refining the clustering process and boosting predictive accuracy beyond that of standard PLS and regression tree algorithms. Additionally, a comparative analysis is presented, evaluating the performance of the enhanced PLS-Tree against regular PLS and regression tree, highlighting its potential for sophisticated data analysis in the process industries.