Automated prediction of shopping behaviours using taxi trajectory data and social media reviews

The Huff model is a well used mathematical abstraction for predicting shopping centre patronage. It considers two factors: shopping centre attractiveness, and customers' travel costs. Here, taxi trajectory data (more than three million journeys) and social media data (more than eight thousand customer reviews) is used to calibrate the Huff model for five primary shopping centres in the rapidly expanding metropolitan city of Shenzhen, China. The Huff model is calibrated in two ways: globally, to find the single pair of best-fit parameters for attractiveness and travel cost; and locally, using Geographical Weighted Regression to find the best-fit parameters at each spatial location. Results demonstrate that customer reviews on social media provide relatively high prediction accuracy for weekend shopping behaviours when the Huff model is calibrated globally. In contrast, customer footfall, calculated directly from number of taxi journeys, provides higher prediction accuracy when the Huff model is calibrated locally. This suggests that, at weekends, sensitivity to footfall has greater spatial variance (i.e., customers living in some areas have greater preference for shopping at popular centres) than sensitivity to customer reviews (i.e., regardless of where customers live, positive reviews on social media are equally likely to affect behaviour). We present this geographical homogeneity in review sensitivity and heterogeneity in footfall sensitivity as a novel discovery with potential applications in urban, retail, and transportation planning.