ABSTRACT
Recent developments of deep learning have demonstrated promising results for challenging tasks in computer vision, natural language processing and so on. With the rapid revolution of spatial data (e.g., scale, variety), there have been broad interests in incorporating these deep networks into the analysis and knowledge discovery process, and unlocking new opportunities in major sectors including smart cities, agriculture, transportation, climate, etc. However, direct applications of deep learning methods often fall short for large-scale spatial data due to spatial heterogeneity, a fundamental quality of spatial data, depicting the phenomenon that data distributions are non-stationary over space (i.e., data distribution shifts from region to region). This property undermines the common independent and identical distribution (i.i.d.) assumption underlying most machine learning methods, including deep learning. Ignorance of heterogeneity in space not only decreases the prediction performance of the models, but also has an impact on the fairness of results – which has become a major consideration for the responsible use of machine learning. We summarize recent heterogeneity-aware and fairness-aware frameworks that target on addressing the heterogeneity challenge for spatial data.
