Explicable Machine Learning Models Using Rich Geospatial Data

Machine learning includes a variety of analysis tools with improved predictive performance compared to traditional statistical approaches. However, that performance comes at a cost of transparency and explainability because much of the models' details are not accessible or interpretable by humans. This problem is compounded by the ability of AI to make use of any correlated data to improve predictions, and practitioners' willingness to sacrifice insight for accuracy. However, there is growing push for “explainable AI”, especially in applications where people's lives, safety, or well-being are affected. Here we explore an intermediate solution suited to policy decision making: leverage the power of tree-based machine learning methods (such as LightGBM) but restrict oneself to variables with plausible causal influence on the target feature. We explore this “explicable AI” approach with an application to housing price prediction and show that using rich geospatial data can replace implicitly spatial variables (such as coordinates and neighborhood names) and achieve similar or better accuracy. And because the explanatory features are naturally interpretable, feature importance analyses provide genuine insights into contributing factors in a way that is similar to hedonic pricing models.