Performance Evaluation of Some Machine Learning Regression Models with Application

Abstract

      Currently, Machine learning is an advanced algorithm that yielding accurate classifications or predictions for huge samples sizes. The fat index data is asymmetric and has a right skew so this creates problem when using statistical techniques. For such data, there is still a lot of performance that needs to be improved when comparing statistical techniques and machine learning algorithms. This paper aimed to compare the traditional statistical methods represented by Linear, penalized linear such as Ridge and Lasso regression and machine learning models represented by convolutional neural network for their prediction performance through simulation experiments and real data of fat index. A total of 252 records were used. The prediction performance of fat index by Linear, Ridge, Lasso and Convolutional Network were compared using mean square error and mean absolute error. We concluded that Ridge and linear regression had the worst performance with the biggest mean and absolute prediction errors, while the Convolutional Neural Network technique achieved the lowest mean and absolute prediction errors, providing the best predictive performance for the fat index data and simulation experiments.