Multimodal deep learning for tourism demand forecasting

This study explores the application of multimodal data (online review text, search engine data, holiday calendars, weather, and historical arrivals) to forecast tourism demand across five destinations (Jiuzhaigou, Macau, and Mount Siguniang in China; Hawaii in the United States; and Singapore) during stable and turbulent periods. We develop three multimodal fusion strategies (early, intermediate, and late fusion) and compare their forecasting performance. Early fusion yields the best performance with higher accuracy, and it also achieves better performance compared with traditional unimodal feature extraction methods. Additionally, we extend the Mean Impact Value method to improve the interpretability of multimodal models. This interpretability allows us to understand how different types of information impact prediction results. Beyond providing model transparency, it also offers valuable references for tourism management regarding which types of information are more significant when the environment is uncertain.