This paper addressed the issue of accurate house price forecasting by considering only the trend of price samples or the functional relationship between correlation attributes and prices. To overcome this challenge, this paper proposed a spatiotemporal attention graph convolution long short-term memory model, abbreviated as AG-LSTM, including local feature extraction module, regional feature extraction module, and composite prediction module. The local feature extraction module used isomorphic graph and heterogeneous graph neural network respectively to extract local feature information on the communities and their multiple correlation attributes, as well as the correlation between the communities and their neighboring supporting facilities. The region feature extraction module firstly clustered the adjacent communities, explored the importance of each community to its corresponding region by using graph attention network, and then established a mapping matrix between the communities and their regions. The module extracted the regional features based on the information of these communities and the mapping matrix. The composite prediction module used a long short-term memory model to perform modeling on the composite features time series composed of local and regional features. The paper conducted experiments using the Beijing housing price data from Lianjia website, and the results show that the AG-LSTM model outperformed the existing baseline models. This model can simultaneously explore the influence of location relationship between the communities, location relationship between the communities and their supporting facilities, multiple correlation attributes, and price trend on the time series, to achieve good performance in house price forecasting.

[2023-06-07] Accepted Paper /doc/2023.03.0104/v1

[2023-11-05] Printed Article /doc/2023.03.0104/v2