How can we use neural network with entity embedding for product valuations? A case study for the car industry

Abstract

Deep learning has demonstrated outstanding success in unstructured data areas, such as computer vision. However, it has not exhibited noteworthy performance in exploiting structured data (tabular data). A neural network is inefficient in treating the structured data type, particularly when categorical variables are dominant. This explains the substandard achievements of deep learning in structured data. This study applies a dense (fully connected) neural network to a structured dataset (three-million car registration records) to estimate car prices, and attempts to improve the network performance using entity embedding layer, when the prevalent data types are categorical variables. It is demonstrated that the information can be captured efficiently via entity embedding technique. The network designed for this study is expected to be reused in valuation tasks where data are highly cardinal. Additionally, the empirical findings identified in embedding matrix could provide insights for the asset valuation industry in the form of transfer learning.