Predicting the price of used cars with deep neural networks

OData support
Supervisor:
Dr. Bolgár Bence Márton
Department of Measurement and Information Systems

With the new generation GPUs (Graphical Processing Units) and countless amount of data uploaded to the Internet, artifical intelligence based solutions live their golden age. Artificial intelligence and deep learning systems contribute greatly to improving the quality of our lives. Without these solutions, there would be no search result, self-driving car or face recognition.

Recently car useage has become an essential part of our everyday life. However in Hungary the newly introduced vehicles only make up a fraction of the purchases and the transfer of second-hand cars that are imported or are in circulation in our country continues to determine the vast majority of purchases.

This is why I have been working on estimating the value of used vehicles during my thesis. From simple MLP (Multi-Layer Perceptron) to deep learning applications I have tried several approaches to predict the value of the second-hand vehicles as accurately as possible. The price prediction was done using two completely different methodologies. First of all I investigated the accuracy of the various neural networks with regression on the continuous set of numbers. Then I tried a completely different perspective to predict the price of the second-hand cars: classification on the set of discrete numbers. In the latter I did not use the value of the vehicles directly, but they were categorized into different price categories and I used classification instead of regression.

Furthermore during my work I have determined the properties which have the greatest influence on the value of the second-hand vehicles. First I used just the internal features of the vehicles (year, cylinder capacity, mileage, etc.) for the price prediction. Then I tried to predict the value of the vehicles based on only a photograph of them. And lastly the combination of this methods. I managed to prove that using those data together with a more complex network works better than using them separately. I used MLP to predict the price based on only the internal features and CNN (Convolutional Neural Network) to predict the value based on only the images. The model which use the internal features and images together is a mixture of MLP and CNN.

Downloads

Please sign in to download the files of this thesis.