Predicting tourism demand using data based on a two-stage feature selection: A hybrid deep learning approach incorporating Time2Vec

Abstract

Accurate tourism demand forecasting is important for regional tourism planning, management, and industry development. However, existing models often struggle with the complexity of external variables or fail to capture essential temporal patterns and multi-scale temporal correlations, directly limiting their accuracy and robustness. Therefore, we propose a predictor with Two-Stage Feature Selection and Time2Vec-enhanced Extraction Mechanisms (TFS-T2VEM). The model employs a two-stage feature selection strategy to refine predictive variables and integrates a Time2Vec-driven temporal pattern extraction module to effectively capture key temporal patterns across multiple scales. By leveraging multi-scale features from intermediate layers of Convolutional Neural Networks (CNN), it captures both mid-short-term fluctuations and long-term trends. Time2Vec further serves as an implicit temporal decomposition module, replacing traditional methods by embedding temporal information directly into the network. This enables dynamic attention adjustment based on intrinsic periodicity and external disturbances, enhancing the temporal attention mechanism by focusing on critical time points and reducing noise from irrelevant features. These improvements ultimately contribute to higher predictive accuracy and robustness. Extensive experiments on three datasets show that our model consistently outperforms baseline methods, confirming its effectiveness in tourism demand forecasting.