A multi-source behavioral data framework for interpretable urban tourism forecasting

Abstract

Abstract Urban tourism demand prediction remains challenging due to its volatility and complex behavioral patterns. To address these issues, a hybrid forecasting framework integrating Long Short-Term Memory (LSTM) networks with Graph Neural Networks (GNNs) was developed. The framework utilizes multi-source behavioral data to improve forecasting accuracy and robustness. The dataset includes social media sentiment, online travel agency (OTA) activity, meteorological information, and mobile signaling records from eight representative Chinese cities collected between 2022 and 2024. The proposed model achieves a Mean Absolute Percentage Error (MAPE) of 6.31% and a trend accuracy of 83.7%, both outperforming single-model benchmarks. Feature interpretability and structural equation modeling (SEM) analyses indicate that sentiment indices, user engagement, and holiday effects are the main determinants of forecasting accuracy. The study offers a scalable and interpretable intelligent forecasting paradigm for smart tourism management, clarifying how data heterogeneity and model adaptability jointly enhance predictive performance in urban tourism contexts.