Publications
カンファレンス (国際) Are Crowded Events Forecastable from Promotional Announcements with Large Language Models?
Soto Anno (Tokyo Institute of Technology), Dario Tenore (Swiss Federal Institute of Technology in Zurich), Kota Tsubouchi, Masamichi Shimosaka (Tokyo Institute of Technology)
The 32nd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL 2024)
2024.11.25
the perspective of understanding purchasing demand and ensuring crowd safety. Existing event forecasting involves specialized feature design based on event-specific domain knowledge, which is useful only in very limited scenarios, such as sports games or the final stages of league play; for most public events, such as fireworks or festivals, it is difficult to define features that are useful for prediction. To address this issue, we propose a novel crowd dynamics forecasting framework that can forecast various types of events. The proposed method uses LLM to summarize the factors that cause crowds from raw event information on the web and model the nonlinear interactions among the extracted event conditions. We evaluate the performance of the proposed framework in predicting crowd dynamics based on a year’s worth of event information obtained from 24 venues of various types of events in Japan. The results show that the proposed framework outperforms existing event forecasting and crowd dynamics forecasting models for a variety of events. We also evaluate the predictability of the proposed method for venues that are not included in the training data, and demonstrate the practicality of the proposed method from the viewpoint of model flexibility.
Paper :
Are Crowded Events Forecastable from Promotional Announcements with Large Language Models?
（外部サイト）