Empirical analysis of I-GBDT to improve the accuracy of mass appraisal method

Abstract

The transactions related to real estate in social and economic activities need to appraise the market value of commercial housing. In order to improve the accuracy and efficiency of mass appraisal model (MA), the paper proposed the Improved-GBDT algorithm based on hybrid feature selection, which combines the concept of supervised learning and unsupervised learning, selecting appropriate features based on the impact of features on results and the changing trend of feature data. The feature weights can be calculated by Random Forest and XGBoost. Considering grey correlation analysis, the correlation degree among features can be calculated, and appropriate features can be selected. In the paper, we took the housing sales data from May 2014 to May 2015 in USA into simulation analysis. The simulation results show that IGBDT model has better accuracy and stability than SVR, RF and XGBoost. Compared with RF and XGBoost model under optimal parameters, IGBDT not only reduced the average absolute error of prediction by 98% and 11%, but also make the maximum relative error drop by 97% and 12%. The results show that IGBDT can provide a powerful reference in evaluating the property market value.