Forecasting the number of visitors for public events, which we call event crowd forecasting (ECF), has attracted much attention in recent years due to its practical importance. In particular, forecasting the event crowd can be used to predict the demand for purchases by visitors. Furthermore, such forecasting can be used to avoid confusion caused by congestion.

Then, what are typical factors and conditions that attract many visitors to events? In general, more people will go to an event if its contents are attractive, if the venue is easily accessible, and if the event occurs on a weekend or sunny day.

Existing ECF methods have adopted a strategy of designing features based on a list of information extracted from such event conditions. Although existing methods are successful by taking the list of features as input, such a listing strategy limits the model’s scalability to other types of events, as different events have different contents that attract visitors. Listing the event contents and engineering the features for all the different types of events requires expert knowledge, which is impractical in terms of costs.

To address this issue, we focus on event announcements published on the web by organizers. Event announcements are considered to be useful for forecasting event crowds because they include a detailed description of the events. On the other hand, the format of event announcements varies from event to event, so they cannot be handled uniformly among various events. Therefore, we proposed unified and effective features that can describe the event conditions by converting the announcement information into a unified text-embedded vector, utilizing large language models.

We conducted performance evaluation experiments on 305 events at 24 venues around Tokyo, using actual event announcement information and GPS-based mobility log data. The results showed that the proposed framework outperforms state-of-the-art ECF methods for various types of events.

Soto Anno, Dario Tenore, Kota Tsubouchi, and Masamichi Shimosaka.
Are Crowded Events Forecastable from Promotional Announcements with Large Language Models?.
SIGSPATIAL’24: Proceedings of the 32nd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, Atlanta, GA, USA, Oct. 29 – Nov. 1. 2024. (To appear)

安納 爽響, 坪内 孝太, 下坂 正倫.
イベント告知情報と大規模言語モデルに基づく イベント会場周辺の早期群衆混雑予報
情報処理学会研究報告 第82回UBI合同研究発表会, 鹿児島県熊毛郡, 5 2024.