머신 러닝 방법을 이용한 오피스 임대료 산정 -랜덤 포레스트, 인공 신경망, 서포트 벡터 머신 활용을 중심으로-

Abstract

본 연구는 최근 자동 평가 모형과 관련한 연구에서 보편적으로 활용되고 있는 랜덤 포레스트, 인공 신경망, 서포트 벡터 머신을 사용하여 오피스 임대료 산정 모형을 구축하고 이들의 적용 가능성을 검토하였다. 이를 위해 서울시에 소재한 507동 오피스의 임대 조사 자료를 활용하여 각 방법별 최적 모형을 구축하고 이들의 성능을 비교하였고, 가장 좋은 성능을 보인 모형으로부터 PD plot을 도출하였다. 표본 전체를 대상으로 하여 임대료 산정 모형을 구축한 결과, 모형별 성능의 순위는 서포트 벡터 머신 모형, 인공 신경망 모형, 랜덤 포레스트 모형 순으로 나타났고, 서포트 벡터 머신 모형으로부터 PD plot을 도출하여 해당 모형이 공실률이 높을수록, 전용률이 높을수록, 연면적이 클수록, 층수가 높을수록, 연식이 낮을수록, 승강기수가 많을수록 임대료를 높게 산출하는 것을 확인하였다. 주차대수는 모형의 예측값과 2차 곡선의 형태를 보이는 것으로 나타났다. 추가적으로 표본을 초대형, 대형, 중대형, 중형 등급으로 나누어 각 규모 등급별 임대료 산정 모형을 각각 구축하고 이들의 PD plot을 도출하였다. 본 연구는 다양한 머신 러닝 방법을 사용하여 오피스 임대료 산정 모형을 구축하고자 하였다는 점, PD plot을 통해 모형의 학습 결과에 대한 해석을 시도하였다는 점, 규모 등급별 임대료 산정 모형을 각각 구축하여 그 결과를 제시함으로써 규모 등급별 분석의 필요성을 입증하였다는 점에서 의의를 갖는다.|This study constructed an office rental estimation model using random forests, artificial neural networks, and support vector machines, which are commonly used in recent studies related to automatic evaluation models, and reviewed their applicability. To this end, the optimized models were constructed for each method and their performances were compared using the rental survey data of the 507 offices located in Seoul, and PD Plots were derived from the best performing model. When the models were built for the entire observations, the ranking of performances by the models was shown in the order of support vector machines model, artificial neural networks model and random forests model, and the higher the vacancy rate, the higher the utilization rate, the higher the floor level, the higher the annual rate, the higher the rent was calculated. The number of parking vehicles showed the shape of the quadratic curve about predictions of the model. In addition, the samples were divided into super-large, large, medium-large, and medium size classes, each of which estimated a model for calculating rent for each class of scale and derived their PD plots. This study is meaningful in that it was intended to build an office rental estimation model using various machine running methods, attempted to interpret the learning outcome of the model by drawing PD plots, and demonstrated the need for analysis by scale classes by constructing and presenting the results.; This study constructed an office rental estimation model using random forests, artificial neural networks, and support vector machines, which are commonly used in recent studies related to automatic evaluation models, and reviewed their applicability. To this end, the optimized models were constructed for each method and their performances were compared using the rental survey data of the 507 offices located in Seoul, and PD Plots were derived from the best performing model. When the models were built for the entire observations, the ranking of performances by the models was shown in the order of support vector machines model, artificial neural networks model and random forests model, and the higher the vacancy rate, the higher the utilization rate, the higher the floor level, the higher the annual rate, the higher the rent was calculated. The number of parking vehicles showed the shape of the quadratic curve about predictions of the model. In addition, the samples were divided into super-large, large, medium-large, and medium size classes, each of which estimated a model for calculating rent for each class of scale and derived their PD plots. This study is meaningful in that it was intended to build an office rental estimation model using various machine running methods, attempted to interpret the learning outcome of the model by drawing PD plots, and demonstrated the need for analysis by scale classes by constructing and presenting the results.