A Convolutional Approach for Estimating Popularity of New Branch Stores

Thousands of shops are opened and closed down constantly every day. Determining the location of a new branch store is a critical issue for retail business. Making decisions for locations of new stores has been widely studied in many fields considering cost saving and profits. In this study, we propose the "Grid-liked Graph" (GLG), a graph-based approach for spatial data division to model the spatial distribution of urban features. Based on the proposed GLG, four footprint features are extracted from location-based social network data and treated as image features. Convolution Neural Network (CNN)-based models including LeNet and AlexNet are then adopted to infer the popularity of each candidate store for optimizing location selection. According to our experimental results on two kinds of retail stores, the proposed methods can increase usage of data, and can also perform well for datasets with fewer features compared with previous studies.