Housing price prediction using convolutional transformer

Abstract

Since the paper ”Attention is All You Need” came out in 2017, the trans former (TF) model has greatly attracted the interest of many scholars. However, for housing price data sets with multiple features and irregular price changes, the original TF shows the weakness that its self-attention calculation method is insensitive to local information, making the model susceptible to outliers and causing potential optimization problems. To further improve this problem in housing price prediction, this project utilizes convolution embedding to enhance the correlation between adjacent data points. The data set used in this paper are the apartments sold-price in Toronto from 2005 to 2010, which holds nearly 81 features. This study stratifies the dataset chronologically, segregating it into training and validation sets in an 8:2 proportion. The initial 80% of the dataset, spanning from 2005 to 2009, is designated for model training. Subsequently, the study examines future housing prices under the ”SalePrice” item. The final 20% of the validation set data, covering the period from 2009 to 2010, is employed for verification and the computation of house price prediction errors. Based on prediction test results, ConvTrans (convolution + transformer) achieves smaller prediction error (0.1567) than traditional TF (0.2487) and LSTM (0.2755). Simultaneously, in comparison to the prediction outcomes obtained by Y. Chen (2021) utilizing identical datasets and employing non-time series model algo rithms, ConvTrans consistently exhibits superior predictive performance.