MODIFIED ADAPTIVE LASSO FOR CLASSIFICATION OF HIGH DIMENSIONAL DATA

Abstract

High-dimensional classification problems, such as gene expression analysis in medical research, require effective variable selection techniques to improve predictive accuracy and interpretability. Traditional penalized logistic regression methods, such as LASSO and Elastic Net, have been widely applied for simultaneous variable selection and coefficient estimation. However, these methods suffer from limitations, including selection bias and inefficiencies in handling correlated predictors. This study introduces the Modified Adaptive LASSO (MALASSO), a novel approach that enhances high-dimensional classification by incorporating an improved weighting mechanism based on ridge regression estimates. The new weighting scheme mitigates the selection bias observed in LASSO-based methods and improves classification performance in datasets with highly correlated features. To evaluate MALASSO’s effectiveness, extensive simulations and real-world applications were conducted using leukemia and colon cancer gene expression datasets. Results indicate that MALASSO outperforms existing methods, achieving superior classification accuracy (98.45% for leukemia and 100% for colon cancer) while selecting fewer, more relevant variables. Compared to Adaptive LASSO (ALASSO) and Adaptive Elastic Net (AEnet), MALASSO demonstrated improved robustness and model sparsity, highlighting its potential for high-dimensional medical diagnostics and biomarker discovery. This study contributes to the advancement of penalized regression techniques by addressing critical shortcomings in existing methods. Future work will explore MALASSO’s applicability to multiclass classification and other high-dimensional domains.