Revisiting gravity model through explainable GeoAI to predict survivability of restaurants

Keywords: Explainable GeoAI, Recurrent Neural Network (RNN), Survival analysis, Location theory, Urban dynamics
Session Type: Virtual Paper Abstract
Day: Tuesday
Session Start / End Time: 3/1/2022 08:00 AM (Eastern Time (US & Canada)) - 3/1/2022 09:20 AM (Eastern Time (US & Canada))
Room: Virtual 35

Abstract

The gravity model of competing location for businesses has been widely used to explain spatial distributions of businesses, estimate potential revenue, and predict entrepreneurship of businesses with a simple concept, which attractiveness of competitors is discounted by distance. A practical limitation of gravity models is sensitiveness to distance-decay parameter and its arbitrariness. Although many studies have attempted to advance methods to obtain better parameters for better predictions, it is still challenging to predict accurately in real-world cases through theory-based approaches. The recent data-rich environment has also improved performance of Artificial Intelligence (AI) techniques in geospatial research, particularly deep learning models. GeoAI has been applied to a variety of practical research, but it remains in its early stages with many technical and theoretical challenges. In particular, the limited explainability of deep learning models is big obstacles in comprehending complex processes of urban dynamics. In this study, we will investigate the impact of relative location between competitors on survivability of restaurants with explainable recurrent neural networks (RNNs) based on gravity models. As gravity models explain the relative location of a given point is a surrogate for a function of potential profit and the function determines restaurant survival, RNN models will define the unknown non-linear function with distance-decay parameters that provide best predictions. We will compare traditional methods to evaluate performance of the proposed method. In addition, we will test several structures to verify the model’s advanced explainability.