大規模商業施設におけるデータ駆動型買い回り行動モデルの開発

抄録

The use of ICT is becoming common in architectural design. It is expected that the increasing use of simulations and various AI technologies will lead to further improvements in the evaluation of spatial performance during the design stage. This study develops a data-driven agent model to simulate consumer shopping behavior in a large-scale commercial facility, with the ultimate aim of identifying the optimal store arrangement and passage shape. The subject commercial facility is a shopping center with 232 stores and 68,640 square meters of floor space. Four stories of the facility are above ground and one story is below. The data obtained from the facility and used in the model included (1) the number of entering/exiting visitors, (2) the purchase histories of membership cardholders, and (3) the number of visitors passing various points in the facility’s passageways. The membership cards of shoppers gave the authors access to 300,000 to 400,000 purchase histories of approximately 100,000 cardholders per month. To model consumer shopping behavior, the transition probabilities of moving between stores were generated from the available purchase histories. The results are as follows.

Two types of stores were identified by analyzing the 215×215 matrix of store-to-store transition probabilities generated from the available purchase histories. The first type includes stores targeted by visitors whose have a clear purpose upon entering the SC. The second type includes stores where visitors tend to make purchases as a result of shopping around rather than having a clear purchase objective in mind immediately upon entry to the facility. The between-store transition probabilities involving stores on the same floor tended to be relatively high, especially in the food court area and the area where similar clothing stores was clustered. On the other hand, the transition probabilities involving stores on different floors tended to be relatively low, particularly the probabilities of going from the first basement floor, the third floor, and the fourth floor to the second floor.

The simulation was generally successful in reproducing macroscopic phenomena such as the number of visitors staying and leaving the SC by using the number of entering visitors and the between-store transition probabilities as input data. The exit-selection probabilities showed a high degree of reproducibility under the assumption that the exit used by leaving visitors would be the same as the entrance they had used. The ability of the simulation to reproduce passage use was rather low when it was assumed that the visitor would select only the shortest route to move from one store to another. However, the simulation results more closely approximated the measured values when the route selection parameters for some of the passages were adjusted.