應用圖卷積網路預測房屋價格

Abstract

過往預測房屋價格的相關研究大都未考慮到房屋交易間的時間與空間相關性，因此，本研究提出利用房屋交易的時間與空間相關性預測房屋價格的架構。我們所提出的架構包含六個階段，我們首先萃取每個房屋的特徵包括實價登錄資訊、公共設施位置以及衛星雲圖；接下來的四個階段，依序建構距離、K-近鄰、集群以及公共設施四個注意力機制，聚合影響房屋價格的相關資訊；最後，使用這四個注意力機制聚合的相關資訊預測房屋價格。實驗結果顯示，我們的研究架構在均方根誤差、平均絕對誤差百分比以及決定係數方面均優於最先進的方法。我們的研究架構可以協助有購買房屋意願的買家事先評估相關物件的價格與買賣時機，也可幫助不動產廠商進行房屋價格鑑定以及給客戶有益且合適的建議，亦可幫助決策者更了解不動產的市場結構與動態。

Most previous studies of house price prediction overlook the temporal and spatial dependencies among house transactions. In this study, we propose a framework to predict house prices by considering the temporal and spatial dependencies. The proposed framework contains six phases. First, we derive a feature vector for each house, including the information of house attributes and public facilities, and features extracted from satellite maps. Second, we employ the distance attention mechanism to aggregate the information of neighboring houses. Third, we exploit the k-nearest neighbors (KNN) attention mechanism to integrate the information of the k-nearest houses with the same building type and close house ages. Fourth, we apply the cluster attention mechanism to capture the correlations between the house and neighboring clusters grouped by the k-means clustering algorithm. Fifth, we utilize the public facility attention mechanism to include information of public facilities. Finally, we develop a model to predict house prices based on the information consolidated from the previous attention mechanisms. The experimental results show that our proposed framework outperforms several state-of-the-art models in terms of root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R-squared). Our proposed framework may offer an effective tool for potential homebuyers to make informed decisions about when and where to buy houses, for real estate agents to better appraise the real estate and provide helpful suggestions for their clients, and for policymakers to better understand the structures and dynamics of the real estate market.