Multi-Horizon Visitor Forecasting for a Shopping Mall with Limited Historical Data Under Covid -19 Pandemic

Takashi Michikata, Shunya Taniguchi, Toma Ishii, Chisaki Hori, Lifeng Lin, Hangli Ge, Noboru Koshizuka

Abstract: Forecasting visitors for commercial facilities (e.g., shopping mall) plays an important role for their business objectives such as adjusting human resources and curbing greenhouse gas emissions. Among numerous studies on human mobility, few studies were targeted at commercial facilities. Such forecasting becomes very difficult under unprecedented situations such as Covid-19, and it is imperative to adopt machining learning models for forecasting. In a real-world situation, historical data of human mobility for a commercial facility is limited except for big data of cellular phones and that makes it very difficult to build a forecasting model with high accuracy. In this work, we propose a two-layer deep learning model to forecast multi-horizon daily visitors for a shopping mall under difficult circumstances where human mobility is heavily influenced irregular events such as Covid-19 and historical data are only available for a short period of time (e.g., four months). We carried out a case study at a real-world shopping mall for a year with infrared beams sensors we installed, apply our proposed model on the shopping mall, and confirm its effectiveness. Our proposed two-layer model (Transformer-LSTM) demonstrates the significant performance improvement over a conventional single-layer model by twenty percent.