GeoGTI: Towards a General, Transferable and Interpretable Site Recommendation

Yunfan Gao, Dong Han, Haofen Wang, Maohong Zhang, Fangjie Hou, Dongqing Yu, Yun Xiong

Abstract: Lack of data and weak interpretability are the main problems faced by store site recommendations. This paper presents a unified site recommendation system called GeoGTI (General,Transferable and Interpretable), which applies to different brands in various industries. Different from the existing single-dimensional transfer methods, we adopt multi-layer knowledge transfer, leveraging knowledge from industries, competitive brands, upper administrative districts, and other cities to alleviate the problems of data scarcity and cold-start. Besides, to fill in the gap of weak interpretability, we score the candidate locations into a five-dimension radar chart from population, business, living, working, and transportation aspects, making the recommended result more convincing and instructive. Extensive experiments are conducted on real-world datasets from various industries, demonstrating GeoGTI’s practicability and effectiveness on store site recommendations.