랜덤 포레스트 알고리즘을 통한 주택 대량평가모형 연구

Abstract

This paper discusses how to apply the random forest algorithm in building a mass appraisal system for residential property and analyzes issues related to modelling process. The paper investigates the relationship between the random forest model and the complexities of housing market. Based on the findings, various qualitative analyses has been attempted for effective model design. The findings are summarized as followed;-. First, the random forest model is performative in capturing the non-linearity from sub-market and locational effects. The random forest model has significantly low average percentage error (approx. 4%) compared to with a linear Hedonic model (approx. 11%). Second, the random forest model can efficiently capture the locational effects only with locational information (coordinates), without proxy variables. Third, using dummy variables may reduce explanatory power of the model, compared to label indexes because the number of variables included in an operation increases. Fourth, the advantage from model complexity seems overwhelm the disadvantage from overfitting. Fifth, modellers are not required to ensure consistency in the time periods contained in a dataset.