Deep Learning Static and Dynamic Movie Attributes for Box Office Prediction

The daily audience data and static movie attributes are both important factors that influence the movie's succeeding box offices. This paper proposes the first framework that utilizes both daily audience data and static movie attributes to accurately the performance of movies’ box-office prediction. To use the daily audience data dynamics, we utilized the recent proposed rank pooling strategy to encode multi-scale audience data dynamics. Meanwhile, we also consider 15 static movie attributes. Both static and dynamic features are combined in a multi-stream residual network for box-office prediction. The experiments conducted on the dataset that contains 120 movies’ daily audience data show that the proposed multi-scale dynamic encoding achieved promising results in prediction the next one- or two-days’ box office while the static-dynamic fusion model achieved the best performance under all conditions