Geographical Support Vector Machines (GSVM) to Analyze Spatially Correlated Data

• Activity Number: 340 - SPEED: SPAAC SESSION III
• Type: Topic-Contributed
• Date/Time: Thursday, August 12, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Shachi Patel* and Stephen Pittman and Ajay Nangia and Jeffrey Thompson
• Companies: University of Kansas Medical Center, Department of Biostatistics & Data Science and University of Kansas Medical Center, Department of Urology and University of Kansas Medical Center, Department of Urology and Department of Biostatistics & Data Science, University of Kansas Medical Center
• Keywords: Support Vector Machines; Geographically Weighted Regression; Binary Classification; Geographically Correlated Data; Prediction; Model Performance

Abstract

Geographically Weighted Regression (GWR) has been developed to capture the strong effect of the local variations. For binary classification, GWR assumes linear relationship between log-odds of the response and the independent variables. Support Vector Machines (SVM) does not require a specific relationship between response and independent variables. Further, it can automatically account for the interactions in the data. However, SVM has never been applied to geographically correlated data. Therefore, we developed a method called Geographical Support Vector Machines (GSVM), which combines geographically related data with SVM. This approach works by creating separate SVM for each local context and weighting observations based on their distance to the local context. To test our method, we build a model to predict counties with increase in the urologists’ availability from 2010 to 2018. We used socioeconomic variables of each county as a predictive parameter. In this situation, our GSVM (AUC 0.80) model performs significantly (p< 0.05) better than SVM (0.70). In conclusion, we have developed GSVM, that can be used to classify geographically correlated data.