Accurate forecasting of crowd density around event venues is critical for safety management, resource allocation, and operational planning. Microblog posts mentioning future events offer valuable signals for forecasting but often contain noise and unrelated content. To robustly extract useful information, we propose two embedding methods for microblog post sets. One uses a Set Transformer to generate embeddings that capture inter-post relationships and suppress noise via self-attention. The other employs a prompt-based large language model (LLM) to summarize and extract structured event information, yielding interpretable summaries of forecast drivers. In both frameworks, each embedding of posts is fused with historical congestion statistics to generate forecasts up to 24 hours ahead. Experiments on real-world congestion statistics and Twitter datasets show the Set Transformer approach reduces forecasting error by 18.2% compared to average vector pooling, while the LLM summarization method achieves a 21.9% reduction. Moreover, the LLM's outputs enable clear interpretation of the factors influencing each prediction.
Paper
Full text
Robust Crowd Forecasting at Event Venues Using Microblog Posts
Semantic Scholar · Computer Science · 2025
Abstract
Accurate forecasting of crowd density around event venues is critical for safety management, resource allocation, and operational planning. Microblog posts mentioning future events offer valuable signals for forecasting but often contain noise and unrelated content. To robustly extract useful information, we propose two embedding methods for microblog post sets. One uses a Set Transformer to generate embeddings that capture inter-post relationships and suppress noise via self-attention. The other employs a prompt-based large language model (LLM) to summarize and extract structured event information, yielding interpretable summaries of forecast drivers. In both frameworks, each embedding of posts is fused with historical congestion statistics to generate forecasts up to 24 hours ahead. Experiments on real-world congestion statistics and Twitter datasets show the Set Transformer approach reduces forecasting error by 18.2% compared to average vector pooling, while the LLM summarization method achieves a 21.9% reduction. Moreover, the LLM's outputs enable clear interpretation of the factors influencing each prediction.