A0376
Title: Weighted robust hybrid partial least squares regression forest
Authors: Aylin Alin - Dokuz Eylul University (Turkey) [presenting]
Abstract: Partial least squares regression(PLSR) is a widely utilized technique for modeling data characterized by multicollinearity or instances where the predictor count exceeds the number of observations. In parallel, random forest regression or regression forest (RF) is an ensemble method proficient in managing extensive datasets, addressing missing predictor values, and mitigating multicollinearity issues. The integration of PLSR and RF termed hybrid PLSR-RF amplifies modeling efficacy. Notwithstanding their advantages, PLSR, RF, and the hybrid PLSR-RF remain susceptible to outlier influence. Although robust variants of PLSR and RF exist, the literature lacks robustified iterations of the hybrid approach. The robust hybrid partial least squares regression forest methodology is introduced to address this gap. This novel method leverages the robust iteratively reweighted SIMPLS algorithm (RWSIMPLS) to derive orthogonal components that subsequently inform the construction of a regression forest. Weighted predictions are applied within this forest to diminish outlier impact in individual trees. Moreover, an alternative bagging technique is proposed that mitigates outlier effects and constrains tree complexity.
