MAMRP: Multi-modal Data Aware Movie Rating Prediction

Mingfu Qin, Qian Zhou, Wei Chen, Lei Zhao

Abstract: Despite the prosperity of the film industry in the past few decades, it is not uncommon to experience the phenomenon that some movies receive high box offices but obtain low ratings.The phenomenon indicates that existing studies which predict the movie-related indicator (i.e., box office) are far from satisfactory. Inspired by this, we formulate a novel task in this work, i.e., multi-modal data aware movie rating prediction (MAMRP), which aims to predict the ratings of emerging movies in time based on movie-related attributes. To tackle the task effectively, we propose a novel model that contains feature extraction, two multi-modal fusion modules, and embedding aggregation. Specifically, the transformer-based pre-trained models are first adopted to perform feature extraction for the attributes of each movie. Then, the extracted features are fed into two fusion modules: a weight-based fusion module considering the different contributions of movie attributes, and a tree-based fusion module considering the hierarchical dependencies and complex correlations between movies. Finally, the movie representations are obtained by embedding aggregation. In experiments, we construct a multi-modal benchmark in accordance with online movie platforms, and the experimental results demonstrate the high performance of our proposed model, which achieves nearly 24\(\%\) relative improvement in classification accuracy compared with baselines.