Stacking Ensemble Learning for Housing Price Prediction: a Case Study in Thailand
Gan Srirutchataboon, Saranpat Prasertthum, Ekapol Chuangsuwanich, Ploy N. Pratanwanich, Chotirat Ann Ratanamahatana
Abstract
In this paper, we analyze the housing price data obtained from a leading Thai real estate website and Open Street Maps (OSM) to identify the features that affect the housing price in Thailand from 2015 to 2019. Moreover, we propose a model based on a stacking ensemble learning framework, where the predictions are generated by stacking three base learning models consisting of a convolutional neural network (CNN), an ensemble model (such as random forests (RF), extreme gradient boosting (XGBoost) and adaptive boosting (AdaBoost)) and a simple linear regression technique. The CNN is used to extract features from house images which are then combined with traditional features to estimate the initial price. The prediction is then calibrated using linear regression. Compared to individual models, the proposed model achieves a Mean Absolute Percentage Error (MAPE) of 17.83%, significantly outperforming other baselines.