Mobile Phone Price Prediction with Feature Reduction

Abstract

Feature reduction can reduce data dimensionality and streamline model size, which focuses on the high relevance data and inferences the output faster. This paper aims to explore the performance and effectiveness of feature reduction methods that accompany the Multilayer Perceptron classifier in predicting the mobile phone price range. Pearson’s Correlation and Principal Components Analysis are chosen as the feature reduction techniques in the research. The experiment sorts the features in significant order with two distinct methods. The three experimental groups reduce 5 features each time and the control group has no feature selection. Then all the groups use the open dataset to train and test the accuracy and loss through MLP. The result indicates that the feature selected by the correlation coefficient facilitates the accuracy of the classification model. When PCA is implemented and only a few features get reduced, the performance improves a little bit, but when more features are eliminated there are huge negative influences. Pearson’s correlation has a better performance than PCA in this experiment, which achieves 95.8% accuracy and validate the effectiveness of the feature reduction method.