Machine learning-based variable selection: An evaluation of Bagging and Boosting

Abstract

Variable selection is a necessary step to build a useful regression. In this paper, an evaluation of different methods (variable selections) including Bagging and Boosting were performed. Large datasets from 1924 observations were taken and the second interaction data which contains 435 variables were employed. In big data, there is no single variable selection technique that is robust towards different families of regression algorithm. The existing variables techniques produce different results with different predictive models. Variable selections only provide the rank of important variables which means that the techniques did not have rules in selecting the suitable range of variable importance.  Each of 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, and 120 highest variable important were selected. Several validations such as Sum Square of Error (SSE), R-square, and Mean Square Error (MSE) were used to compare its performances. As the result, bagging for the 90 highest variable important was better than others SSE (31077.8295), R-square (0.9210), and MSE (17.8344), respectively. Hence, the variable selection using bagging has been considered as the best model.