공간 데이터와 기계학습 기법의 결합을 위한 공간적 입력변수 유용성 탐색

When applying ML (Machine Learning) to spatial data, neglecting to account for spatial autocorrelation leads to an increase in prediction errors. Incorporating spatial features is a common approach to consider spatial autocorrelation in ML. This study explores the impacts of incorporating spatial features into ML on the improvement of prediction accuracy. Additionally, this study conducts comparative analyses of different types of spatial features and their effects on ML models. The spatial features include spatial coordinates, spatial lags, and Moran eigenvector spatial filters. The ML techniques employed consist of LM (Linear Regression Models), GAM (Generalized Additive Models), RF (Random Forests), and SVM (Support Vector Machines). Apartment sale prices in Seoul having strong positive spatial autocorrelation are utilized as the response variable. The results indicate LM, GAM, and RF show the highest increases in prediction accuracy with spatial lags, while SVM exhibits the highest increase with spatial coordinates. In the model comparison, the model incorporating RF and spatial lags reports the highest prediction accuracy, and generally spatial features can contribute to increase the prediction accuracies. This study is expected to provide a research base for the development of spatial machine learning by summarizing spatial features types and highlighting their utilization with the corresponding feature selection methods.