Combining Structured and Unstructured Data using Co-Attention for Car Price Prediction

Abstract

Automatic price appraisal is a relatively new feature in e-commerce sites that aims to help sellers and buyers with price suggestions. To this end, we explore the problem of predicting prices from public online car listings. A listing usually contains the car specification along with a description in the form of unstructured text. We leverage these structured and unstructured information for predicting prices by combining them in a neural network model. Our model has a fully-connected neural branch for handling structured data that learn patterns from the tabular attributes; and another branch that models interactions between the tabular attributes and textual description by employing a co-attention mechanism. The combined outputs of these branches are fed into a regression layer that predicts the car price. The latent representation of the car listings, as mapped by the intermediate layers of the learned model, can be used as a feature set by any regression algorithm to perform car price prediction. This learned feature encodes to a vector the relevant aspects of the car listing, as demonstrated by our experiments. We evaluate our solution in a dataset that was built by scraping multiple car classifieds websites. Our findings show that: (1) our network achieved lower prediction errors than the evaluated regressors using the same input features; (2) by using our learned representation, the regressors surpassed their performance using the original feature mapping; and (3) more competitive algorithms, such as LightGBM and Random Forest, using the raw features, are outperformed by simple linear models using our embeddings.