| 英文摘要 |
In recent years, a variety of statistical methods have been used to estimate house prices, including machine learning techniques such as linear regression, random forest, etc. Among these methods, deep learning has demonstrated exceptional performance. Numerous researchers have utilized artificial neural networks to predict house prices with outstanding accuracy. However, the majority of these studies rely on tabular data, focusing solely on the characteristics of house transaction records while neglecting the possibility that the house price may be influenced by the surrounding environment, hence raising the issue of spatial nonstationarity. Therefore, the purpose of this study was to connect the convolutional neural network with multilayer perception, proposing a hybrid neural network structure that also extracts the distribution of surrounding variables, with the intention to enhance the generalizability of the model by reducing the impact of spatial heterogeneity. This study selected Taoyuan District of Taoyuan City as the primary research area, with a research period from January 1, 2015, to December, 2018. As the source of real estate feature data, we relied on the actual price registration data published by the Ministry of the Interior, R.O.C. Regarding the environment component, we utilize Sentinel-2 satellite images, land use inventory data provided by the Ministry, and distances to public transportation.
According to the research findings, the multilayer perceptron model had an adjusted R-squared value that was roughly 0.103 percentage points greater than the geographically weighted regression. In comparison to the multilayer perceptron, the adjusted R-squared of the hybrid neural network model was increased by around 0.102. This indicated that building an additional spatial feature extraction pipeline did improve the model's performance. Therefore, this practice was more applicable in extremely heterogeneous environments due to the structure of the proposed model. |
本站僅提供期刊文獻檢索。
