| Penulis/Author |
NOVIANA PRATIWI (1) ; Prof. Dr.rer.nat. Dedi Rosadi, S.Si.,, M.Sc. (2); Prof. Dr. Abdurakhman, S.Si., M.Si. (3); Drs. Danardono, MPH., Ph.D. (4) |
| Abstrak/Abstract |
Orthogonal Projection to Latent Structures Discriminant Analysis (OPLS-DA) is a multivariate classification method that
effectively addresses multicollinearity and provides stable, interpretable models by separating relevant class discrimination information
from correlative data. However, classic OPLS-DA is sensitive to outliers, which can distort model estimates and reduce predictive
performance. Addressing this limitation, we propose a robust form of OPLS-DA by incorporating Huber and Tukey weighting schemes to
improve resistance against outliers. The Huber method adaptively combines squared error and absolute deviation, reducing the influence
of moderate outliers while preserving efficiency for normally distributed data. Meanwhile, the Tukey method further limits the impact
of extreme outliers by assigning negligible weights to highly deviating points, making the model more resilient in highly contaminated
datasets. By integrating these robust weighting strategies, our approach enhances the stability, reliability, and accuracy of OPLS-DA,
particularly when handling noisy and high-dimensional datasets. To evaluate the effectiveness of our robust OPLS-DA, we conducted
extensive simulations across various contamination scenarios and applied the method to real-world datasets. The results indicate that
our approach significantly improves classification performance, model generalizability, and robustness to outliers, outperforming the
classical OPLS-DA in challenging conditions. These findings suggest that robust OPLS-DA is a valuable enhancement for applications
requiring reliable discrimination analysis, such as biomedical research, chemometrics, and metabolomics. |
