The success of a neural network depends on, among others, an architecture that is appropriate for the task at hand. This study attempts to identify an optimal architecture of a neural network in the context of property valuation, and aims to test the ability of connecting related neural networks to reduce the property valuation error.
This study explores efficient network architectures to estimate land and house prices in Seoul, South Korea. The input is structured data, and the embedding technique is used to process high-cardinality categorical variables.
The shared architecture of a network for simultaneous estimation of both land and houses was revealed to be the best performing network. Through weight sharing between relevant layers in networks, the root-mean-square error (RMSE) for land price estimation was reduced significantly, from 0.55–0.68 using the baseline architecture, to 0.44–0.47 using the shared architecture.
The study results are expected to encourage active investigation of efficient architectures by using domain knowledge, and to promote interest in using structured data, which is still the dominant type in most industries.
