Improving performance of mass real estate valuation through application of the dataset optimization and Spatially Constrained Multivariate Clustering Analysis

Mass real estate valuation is a multidimensional and complex matter because it depends on many constant and time-varying factors. It is desirable to have high level of model performance in the development of mass real estate valuation models for the development of sustainable real estate management strategies. For this reason, this study aims to develop a comprehensive methodology that increases the performance of mass real estate valuation models by using optimized datasets and clustering geographical value in Geographic Information Systems (GIS) modeling environment. A case study was carried out in Istanbul and Kocaeli provinces covering neighborhoods with different levels of socio-economic-development. This study was carried out using the big data, which was prepared for 121 criteria incorporating approximately 200.000 real estate values. Firstly, datasets were optimized by using the Boxplot technique concerning dataset-based outliers and Cluster and Outlier Analysis techniques were used regarding the location-based outliers. Next, 22 of the criteria affecting the value was determined with Pearson Correlation technique through analyzing the local relationship between real estate value and the criteria. Based on the result of the Spatially Constrained Multivariate Clustering Analysis (SCMCA) analysis, five different geographical value clusters with similar socio-development characteristics were detected. Mass valuation performances were tested covering all study area and these five clustered areas assessed with the use of Multiple Regression Analysis (MRA) model were used commonly in developing mass real estate valuation models. The model accuracies were evaluated through performance measurement metrics used in machine learning (MAE, MSE, RMSE) and mass real estate valuation (WtR, COD, PRD) technique that was recommended by IAAO. Considering the performances of the models, value prediction models based on geographical value clusters were more successful than the entire of study area model.