Probability Sparse Attention Based House Price Prediction

House price prediction is a hot topic in the estate field. It is of great significance for market analysis, policy formulation, and investment planning. Traditional prediction methods often do not consider the interactions between different features and learn the complex relationships between them. Therefore, it is hard for them to effectively deal with the complex correlations of high-dimensional and multivariate features. To address these challenges, we proposed a novel house price prediction model HPP-Informer based on Informer. HPP-Informer can efficiently capture the nonlinear relationship between input features and housing prices with Probability Sparse Attention (ProbSparse Attention). Firstly, we add a learnable embedding vector to each feature and construct initial feature representation. Consequently, the learnable feature embeddings are fed into the Informer based feature extraction encoder, which is mainly composed of multiple ProbSparse Attention blocks, to model important feature associations. We conducted comparative experiments on the Boston Price Dataset with classic prediction methods such as linear regression, XGBoost, and multilayer perceptron to validate the performance of HPP-Informer. The experimental results show that the proposed method exhibits significant advantages in both prediction accuracy and generalization ability. This paper provides new solutions for price prediction problems with high-dimensional features and has important practical value for the analysis of the real estate market.