Research On The Evaluation Of Used Car Prices Based On GWO Optimized Hybrid Model

With the continuous development of China’s automobile economy,the used car market has become an important part of automobile transactions.Vehicle valuation is one of the important factors affecting the transaction,reasonable valuation can effectively maintain the order of market transactions and promote the transaction of vehicles.Therefore,building an accurate and robust evaluation method through machine learning algorithm is a very worth reseach.The main content of the thesis is:（1）Several algorithms which are used widely in the field of evaluation are reseached,and the Grey Wolf algorithm（GWO）in swarm intelligence algorithm is used to optimize the machine learning algorithm with more parameters.The MAE of GWO-LightGBM model and GWO-CatBoost model obtained by GWO algorithm optimization is 2.58% and 1.47% lower than that of the original model under the five-fold cross-validation.（2）To solve the problem that the price-related feature mining is not deep enough in the valuation research,the feature crossover method is used to mine the vehicle features,and a double filter method based on binary GWO algorithm and LightGBM feature importance is proposed to optimize the features.Tested with LightGBM algorithm and CatBoost algorithm,the predicted MAE on the new dataset is reduced by2.08% and 2.01%.（3）Because of the prediction preferences of each algorithm,the prediction of a single evaluation algorithm on some price intervals will produce large deviations.This thesis presents a hybrid model after considering the convergence and price-related characteristics of each model.The GWO-LightGBM model and GWO-CatBoost model with different convergence characteristics are generated by introducing various Loss models.The prediction results of the better models are combined as new features with a subset of important features selected by GWO as the model base,and the LR model is used as the fitting layer to build a mixed model framework.According to the above,the prediction performance of different valuation models were compared on real used car data,and the results showed that the CatBoost model and XGBoost model had better predictions in the low and middle price ranges,the LightGBM model had better predictions in the high price range,while the Random Forest model and the Neural Network model both had deviations in the overall price range.The hybrid model constructed in this thesis shows better stability on the overall data set,with an overall MAPE of 10.9%,and the percentage of samples with predictions below 10% is 60.8%,and the overall predicted MAE is 2.52% and 2.22%lower than that of the LightGBM and CatBoost models,respectively.The experimental results show that the hybrid model based on GWO optimization can further improve the current prediction accuracy on used car valuation and has strong robustness.