High-dimensional population inflow time series forecasting via an interpretable hierarchical transformer

• Published in: Transportation research. Part C, Emerging technologies Emerging technologies, 2023-01, Vol.146, p.103962, Article 103962
• Main Authors: Hu, Songhua, Xiong, Chenfeng
• Format: Article
• Language: English
• Subjects: Crowd flow; Human mobility; Interpretability; Mobile device location data; Time series forecasting; Transformer
• ISSN: 0968-090X; 1879-2359
• DOI: 10.1016/j.trc.2022.103962

•A deep learning framework is introduced for nationwide population inflow forecasting.•Temporal dynamics are captured via Transformer with local enhancements.•External effects from various factors with different dimensions are fused and selected.•Models are interpreted from static and temporal aspects via built-in parameters. Mobile device location data (MDLD) are emerging data sources in the transportation domain that contain large-scale, fine-grained information on population inflow. However, limited studies have built forecasting models based on large-scale MDLD-based population inflow time series. This task is challenging due to complex nonlinear temporal dynamics, high-dimensional time series structure (i.e. multiple time series with multi-shape inputs and outputs), and non-negligible impacts from various external factors. To address these challenges, this study introduces a deep learning framework, the Interpretable Hierarchical Transformer (IHTF), for nationwide county-level population inflow time series forecasting and interpretation. A variety of cutting-edge deep learning techniques are fused, including the variable selection network to incorporate external effects, the gated residual network to handle nonlinearity, and the transformer architecture to learn temporal dynamics. Different interior parameters, such as variable selection weight and temporal attention weight, are extracted to explain patterns learned by the framework. Numerical experiments show that IHTF outperforms extensive baseline models in forecasting accuracy. In addition, feature importance generated by IHTF is similar to the tree-based model, LightGBM, but exhibits a more even distribution, among which point-of-interests (POIs) count, county location, median household income, and percentage of accommodation and food services are the most important static variables. Moreover, attention weight demonstrates that IHTF can automatically learn the seasonality from time series. Taken together, this framework can serve as a reliable travel demand forecasting component in the transportation planning process that allows modeling the travel demand continuously instead of by snapshot.