Predicting Video Game Sales Based on Machine Learning and Hybrid Feature Selection Method

Sales prediction plays a critical role in the rapid development of the internet. Unlike the goal of predicting physical items sales, the virtual items of video games needn’t consider the relationship of inventory and demand, and it reflects consumer preferences to a certain extent. Predicting sales of the video game can pre-adjust sales strategies and development plans in advance. In the process of predicting, an excellent feature selection method can remove the features irrelevant to enhance the accuracy of the model. In this paper, we proposed a new hybrid feature selection method Pearson correlation coefficient - Random Forest Feature Selection(PCC-RFFS), and utilized 9 machine learning methods combined with PCC-RFFS to predict the sales of video games. For the filter-based stage, we used the absolute value of Pearson correlation coefficient and feature ranking technique, while the wrapper-based stage is based on Random Forest to measure the importance between features and target. The strategy of combination is in addition to both stages. Experiments on the real-world dataset from February 2006 to November 2016 show that the machine learning method combined PCC-RFFS outperforms the machine learning method combined Pearson correlation coefficient or Random Forest.