Machine Learning for Predicting the Antidiabetic Properties of Novel Schiff Bases through FTIR

Abstract

This research tackles the issue of selecting variables and making predictions in high-dimensional datasets by employing a range of regression techniques, such as Bayesian regression, Lasso, Elastic Net, Orthogonal Matching Pursuit, and RANSAC Regression. The main aim is to determine the most efficient method for forecasting a dependent variable using a large array of independent variables and to identify the key predictors. To assess these techniques, we use synthetic data with one dependent variable and 1627 independent variables. Each model undergoes testing and training 50 times, with performance measured by the average Mean Squared Error (MSE) across various data splits and crossvalidation. The results have crucial implications for domains that require reliable methods for variable selection and prediction. Future research will aim to apply these methods to real-world datasets and further refine them to boost their predictive accuracy.