A Novel Used Vehicles Price Prediction Model Based on Denoising Autoencoder With Convolution Operation

Abstract

Accurate prediction of the price of used vehicles can effectively reduce the undesirable behavior of intermediary platforms that mark up prices indiscriminately, make the link of vehicle sales more transparent and fair, reduce the losses of buyers and sellers, and have great economic significance. However, the previous used vehicle price prediction model suffers from redundant and noisy explicit features, resulting in its poor learning efficiency and prediction ability. Moreover, the previous models did not consider the interaction information among features, resulting in the model missing the acquisition of this information during the learning process, which determines its poor generalization effect during the prediction stage. Therefore, to overcome the shortcomings of previous models, we propose a novel used vehicle price prediction model with denoising autoencoder based on convolution operations (DAECO). Firstly, the DAECO model extracts the latent features of used vehicles using denoising autoencoder to remove the non-discriminatory redundant features. Then the DAECO model uses convolution operation to obtain the interaction information between the latent features. Finally, the DAECO model adds up the interaction information between the latent features and the linear combination information by assigning a certain weight to each of them to obtain the final prediction value. The values of the DAECO model on two real used vehicle transaction datasets with MSE, RMSE and MAE metrics are 1.523 and 2.278, 1.234 and 1.509, 1.009 and 1.226, respectively. Through the comprehensive experimental results on the real used vehicle price dataset, the DAECO model outperforms the current popular baseline algorithms.