Toplu taşınmaz değerlemede makine öğrenme algoritmalarının kullanımı ve konumsal/konumsal olmayan özniteliklerin tahmin doğruluğuna etkilerinin karşılaştırılması

Abstract

The objective determination of real estate values with current technological approaches has an important role in effective and sustainable real estate management plans. Mass appraisal is the process of valuing a large number of real estate simultaneously instead of evaluating the real estate individually for reducing the loss in terms of time and cost. Machine learning methods, known as advanced estimation approaches, are used to obtain more objective, accurate and fast results in mass valuation processes. In addition, besides sufficient objectivity and accuracy in value determination, these methods can evaluate the relations between the value and the criteria affecting the value holistically. In this context, model successes in mass valuation were examined using Multiple Linear Regression (MLR), Generalized Linear Model (GLM), Support Vector Machines (SVM), Decision Trees (DT) and Random Forest (RF) algorithms. The datasets were divided into 3 groups as Geographic (G), Non-Geographic (NG) and Geographic + Non-Geographic (GNG) and applied separately for modeling with different methods. Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Squared Error (MSE) and R2 were calculated for determining the model measures. Pendik district of Istanbul province was chosen as the application area. By applying different methods with 1475 sampling points representing the real estate sales values for the application area, the performances of the models established were examined with 3 different data sets (G, NG, GNG). Accordingly, RF is the method that gives the highest accuracy while DT and GLM were found as the methods with the lowest accuracy. When the effects of different datasets on the model accuracy were examined, big difference was not observed between the use of the G and the GNG datasets.