Box-Office Prediction Based on Essential Features Extracted from Agent-Based Modeling

Abstract

In prediction utilizing machine learning techniques, feature selection is significant for increasing the prediction accuracy. Various techniques for feature selection are available, but it is still challenging to provide a set of component features. To overcome this difficulty, we propose to extract essential features through agent-based modeling. Usually, an agent-based model is characterized by a set of parameters. If an agent-based model can mimic the behaviors of the target well, such parameters can be viewed as essential features of capturing the property of the prediction target. To verify the effectiveness of such feature extraction, we focus on opening/gross box-office revenues based on Twitter data and build a regression model that incorporates the information diffusion model as an agent-based model. The experimental results show that our model can extract essential features from Twitter data and predict the gross box-office revenue for 106 movies more accurately than the baseline model.