Predicting Households' Residential Mobility Trajectories: A Geographically Localized Interpretable Model-agnostic Explanation

Keywords: Explainable GeoAI, Model-Agnostic Explanation, Recurrent Neural Networks (RNNs), Trajectory prediction, Residential mobility
Abstract Type: Paper Abstract

Abstract

Human mobility analytics through artificial intelligence (AI) is receiving more attention thanks to increased computational power and the proliferation of high-resolution spatial data. However, few studies in the social sciences and human geography have applied deep learning to investigate underlying processes of socioeconomic phenomena, partially due to their insufficient explainability. This research utilizes an explainable GeoAI approach, Geographically Localized Interpretable Model-agnostic Explanation (GLIME), to understand human mobility at larger spatial and longer temporal scales. We develop a two-layered long short-term memory (LSTM) model to predict individual-level residential relocation patterns across the United States from 2012 to 2019 and to interpret the deep neural networks from geographical perspectives by applying the GLIME at different spatial scales. The result demonstrates that the localized model agnostic index enables spatially explicit interpretations of local variances in the impacts of individual and neighborhood factors. This study also provides insights into the importance of considering path dependency in understanding residential mobility. The prediction result is not as accurate as other deep learning applications. While there is room for improvement in predicting complex human spatial decision-making, our findings provide fundamental knowledge highlighting the importance of spatial heterogeneities in drivers of residential mobility.