공간데이터 기반 기계학습 모형의 예측 정확도 향상을 위한 연접성 제약을 반영한 공간 군집의 활용

Improving prediction accuracy is important in machine learning. One method to increase prediction accuracy is building machine learning models for each cluster which is generally created based on only attribute similarities using cluster analyses. However, spatial data requires clustering to additionally reflect spatial similarities, as considering spatial autocorrelation in a model leads to an increase in prediction accuracy. Therefore, this study explores the impact of spatial clustering with spatial similarities on machine learning prediction accuracy. Specifically, it compares the prediction accuracies of machine learning models generated based on clusters that consider only attribute similarities and those that consider both spatial and attribute similarities. The machine learning techniques employed consist of linear regression model, random forest, and gradient boosting for predicting average daily ridership. The independent variables consist of 11 variables explaining land, population, facility characteristics. The analysis results show that considering both spatial and attribute similarities yields significantly higher prediction accuracy across all models. This study can contribute to the literature on spatial data analyses by demonstrating that considering spatial similarity can improve prediction accuracy when applying spatial data to machine learning.