A Multimodal Machine Learning Framework for Box Office Prediction Using CNN-Based Trailer Analysis and Random Forest Models

Abstract

This study presents a comprehensive multimodal machine learning framework for box office prediction in the Chinese film market. We develop an integrated approach combining multiple computational techniques: time series analysis for market trend visualization, matrix heatmap analysis for actor-genre correlation mapping, convolutional neural network (CNN) for quantitative sentiment analysis of movie trailers through image processing, and random forest ensemble learning for final prediction. Our methodology employs web crawling technology to collect comprehensive datasets from major film platforms. The CNN model achieves 98.21% accuracy in classifying emotional tones from trailer frames, while the random forest model demonstrates robust performance with MAPE of 10.7% on validation data. We validate our framework by predicting box office performance for three films scheduled for 2025 release, achieving relative errors below 20%. This research contributes to the intersection of computer vision, machine learning, and entertainment industry analytics, providing a scalable solution for revenue forecasting in digital media markets.