Abstract—Forecasting opening box-office earnings has become an emerging demand, affecting filmmakers’ financial decisions and promotional efforts by advertising studios that create trailers. Decision-makers have a complex and challenging task due to a large amount of data and several complex considerations. Based on deep multimodal visual features derived from trailer content and a cross-input neighborhood feature fusion, an innovative Deep Multimodal Predictive Cross-Input Neural Network model (DMPCNN) is proposed for predicting opening movie box-office revenue. DMPCNN is a fully-connected recurrent neural network with two architectures: A Visual Feature Extraction Model (ResNet+LSTM) block for extracting and learning mid-level temporal visual content and Cross-Input Neural Network fusion for uncovering and fusing high-level spatial features in trailers to predict movie revenue. The ResNet+LSTM block focuses on learning various trailer segments, while the Cross-Input Neural Network simultaneously learns and combines features from movie trailers and metadata and corresponding similarity metrics. DMPCNN aided in developing a decision support system that incorporates useful revenue-related trailer features. We evaluated DMPCNN’s performance on the Internet Movie Dataset by obtaining metadata for 50,186 movies from the 1990s to 2022 and comparing it with different state-of-the-art frameworks. The erudite features in trailers and the predicted results outperformed baseline models, achieving 81% feature precision and 84.40% accuracy.
 
Keywords—box-office, recurrent neural networks, long short-term memory, cross-input neural network, multimodal features, movie trailers


Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.