Search

CN-121504521-B - Training method and device for secondary handcart price evaluation model based on deep learning

CN121504521BCN 121504521 BCN121504521 BCN 121504521BCN-121504521-B

Abstract

The application discloses a training method and device for a secondary handcart price evaluation model based on deep learning, which are characterized in that secondary handcart historical transaction data are obtained, data preprocessing is carried out to obtain training samples, sample pairs of samples to be evaluated and real samples are selected from the training samples, the real samples and the samples to be evaluated are respectively input into a first sub-network and a second sub-network of a double-tower model to obtain a first sample vector and a second sample vector, the first sub-network and the second sub-network are both in a deep cross network architecture, element level products are carried out on the first sample vector and the second sample vector to obtain interaction feature vectors, the interaction feature vectors are mapped into scalar quantities through an output layer, all parameters of the double-tower model are updated through back propagation according to loss functions of scalar quantities and sample price differences, and the trained secondary handcart price evaluation model is obtained, so that the problems of weak feature cross and insufficient utilization of sample relevance of the secondary handcart price prediction method are solved.

Inventors

  • XU YAFENG
  • Hao Qianqiong
  • HAN XIAOMING
  • LIN QUAN

Assignees

  • 北京淘车科技有限公司

Dates

Publication Date
20260508
Application Date
20260112

Claims (10)

  1. 1. The training method of the second-hand car price evaluation model based on deep learning is characterized by comprising the following steps of: step 1, acquiring historical second-hand vehicle transaction data, and extracting basic attribute features, use attribute features and market quotation features from the historical second-hand vehicle transaction data; Step 2, carrying out data preprocessing on the basic attribute characteristics, the using attribute characteristics and the market quotation characteristics to obtain training samples, wherein the data preprocessing comprises missing value filling, outlier rejection and feature coding; Step 3, selecting a sample from the training samples as a sample to be evaluated, and selecting samples with the same brand, same train and same year as the sample to be evaluated from the training samples as real samples to form a sample pair, wherein the sample price difference of the sample pair is the difference between the residual value of the sample to be evaluated and the residual value of the real sample, and the residual value is the ratio of the price of a transaction to the price of official guidance; Step 4, inputting the real sample and the sample to be evaluated in the sample pair into a first sub-network and a second sub-network of a pre-built double-tower model respectively to obtain a first sample vector and a second sample vector, wherein the first sub-network and the second sub-network adopt a deep cross network architecture; step 5, performing element level product on the first sample vector and the second sample vector to obtain an interaction feature vector, and mapping the interaction feature vector into a scalar through an output layer; Step 6, updating all parameters of the double-tower model through back propagation by taking Huber loss between the scalar and the sample price difference of the sample pair as a loss function; and 7, repeatedly executing the steps 3 to 6 until the preset training termination condition is met, and obtaining the trained second-hand vehicle price evaluation model.
  2. 2. The training method of the second-hand car price evaluation model based on deep learning according to claim 1 is characterized in that in the step 2, the feature coding specifically comprises the steps of performing embedded coding on discrete features in basic attribute features, using attribute features and market quotation features, performing 0-1 coding on the basic attribute features, using the two classification features in the attribute features and the market quotation features, and performing normalization processing on continuous features in the basic attribute features, using the attribute features and the market quotation features.
  3. 3. The training method of the second-hand car price evaluation model based on deep learning according to claim 1, wherein in the step 2, the outlier is identified and removed by a quartile range method.
  4. 4. The training method of the deep learning-based second-hand-car price estimation model according to claim 1, wherein in the step 4, the structures of the first sub-network and the second sub-network each comprise an input layer, a Embedding layer, a feature stitching layer, a cross layer, a depth layer and a feature fusion layer.
  5. 5. The training method of the second-hand car price evaluation model based on deep learning according to claim 4, wherein the cross layer is provided with three layers, and the deep layer is a three-layer full-connection layer.
  6. 6. The training method of the deep learning based second hand car price estimation model according to claim 1, wherein in step 6, the loss function is: ; Wherein Z i is a scalar, label i is a sample-to-residual difference, Is a super parameter.
  7. 7. Second-hand car price evaluation model training device based on deep learning, characterized by comprising: The historical transaction data acquisition module is used for acquiring the historical transaction data of the second hand vehicle and extracting basic attribute characteristics, use attribute characteristics and market quotation characteristics from the historical transaction data of the second hand vehicle; The historical transaction data preprocessing module is used for preprocessing the basic attribute characteristics, the using attribute characteristics and the market quotation characteristics to obtain training samples, wherein the data preprocessing comprises missing value filling, outlier rejection and feature coding; the sample pair construction module is used for selecting a sample from the training samples as a sample to be evaluated, and selecting samples with the same brand, same train and same year as the sample to be evaluated from the training samples as real samples to form a sample pair, wherein the sample price difference of the sample pair is the difference between the residual value of the sample to be evaluated and the residual value of the real sample, and the residual value is the ratio of the price to be submitted to the official instruction price; The training module is used for respectively inputting the real sample and the sample to be evaluated in the sample pair into a first sub-network and a second sub-network of a pre-built double-tower model to obtain a first sample vector and a second sample vector, wherein the first sub-network and the second sub-network adopt a deep cross network architecture; The scalar output module is used for carrying out element level product on the first sample vector and the second sample vector to obtain an interaction feature vector, and mapping the interaction feature vector into a scalar through an output layer; a parameter updating module for updating all parameters of the dual-tower model by back propagation with Huber loss between the scalar and the sample price difference of the sample pair as a loss function; and the circulation module is used for repeating training until a preset training termination condition is met, so as to obtain a training-completed second-hand vehicle price evaluation model.
  8. 8. The second-hand vehicle price evaluation method based on deep learning is characterized by comprising the following steps of: step 1, obtaining second-hand vehicle data to be predicted, and extracting basic attribute characteristics, use attribute characteristics and market quotation characteristics from the second-hand vehicle data to be predicted; Step 2, preprocessing the data of the basic attribute features, the using attribute features and the market quotation features to obtain preprocessed second-hand vehicle data to be predicted; step 3, screening a plurality of data from the real transaction data according to conditions to obtain a reference set; Step 4, inputting each data in the reference set into a first sub-network of a second-hand car price estimation model, and inputting the preprocessed second-hand car data to be predicted into a second sub-network of the second-hand car price estimation model to obtain a plurality of prediction scalars, wherein the second-hand car price estimation model is trained according to the training method of the second-hand car price estimation model based on deep learning as set forth in any one of claims 1 to 6; Step 5, obtaining a plurality of residual value differences according to a plurality of prediction scalar quantities, and calculating the average value of the residual value differences to obtain a residual value difference average value; step 6, calculating a real residual value mean value of the reference set, and calculating the sum of the real residual value mean value and the residual value difference mean value to obtain a secondary handcart residual value to be predicted; And 7, obtaining the trading price of the secondary handcart to be predicted according to the secondary handcart residual value to be predicted and the official guide price.
  9. 9. The method for evaluating the price of the second hand truck based on deep learning according to claim 8, wherein in step 7, when the price of the second hand truck to be predicted is obtained according to the second hand truck residual value to be predicted and the official guide price, the calculation formula is as follows: ; Wherein, P X represents a price exchange, R X represents a secondary car residual value to be predicted, and P guide represents an official guide price.
  10. 10. A secondary car price assessment device based on deep learning, characterized by comprising: The data to be predicted acquisition module is used for acquiring the second-hand vehicle data to be predicted and extracting basic attribute characteristics, use attribute characteristics and market quotation characteristics from the second-hand vehicle data to be predicted; The data preprocessing module is used for preprocessing the data of the basic attribute characteristics, the using attribute characteristics and the market quotation characteristics to obtain preprocessed data of the second hand vehicle to be predicted; The reference set construction module is used for screening a plurality of data from the real transaction data according to conditions to obtain a reference set; The scalar prediction module is used for inputting each data in the reference set into a first sub-network of a second-hand car price estimation model, and inputting the preprocessed second-hand car data to be predicted into a second sub-network of the second-hand car price estimation model to obtain a plurality of prediction scalars, wherein the second-hand car price estimation model is trained according to the training method of the second-hand car price estimation model based on deep learning as set forth in any one of claims 1 to 6; The residual value difference average value calculation module is used for obtaining a plurality of residual value differences according to a plurality of prediction scalar quantities, calculating the average value of the residual value differences and obtaining a residual value difference average value; The second-hand vehicle residual value calculation module is used for calculating a real residual value average value of the reference set and calculating the sum of the real residual value average value and the residual value difference average value to obtain a second-hand vehicle residual value to be predicted; and the second-hand vehicle price prediction module is used for obtaining the price of the second-hand vehicle to be predicted according to the second-hand vehicle residual value to be predicted and the official guide price.

Description

Training method and device for secondary handcart price evaluation model based on deep learning Technical Field The application relates to the technical field of price evaluation of a second-hand vehicle, in particular to a training method and device for a price evaluation model of the second-hand vehicle based on deep learning. Background With the continuous increase of the amount of automobile maintenance and the continuous shortening of the period of changing cars of consumers, the second-hand car trade market has shown a rapidly developing situation in recent years. Under the background, scientific, accurate and efficient second-hand car price evaluation becomes a key link for connecting buyers and sellers, guaranteeing fair transaction and improving market transparency. At present, the price evaluation of the second hand truck mainly depends on the following three evaluation methods, which are respectively applied to different scenes: 1. The manual evaluation method is that an evaluator gives a price interval according to visual information such as brands, age, mileage, vehicle conditions and the like of the second hand vehicle and market experience. This method is still widely used in the secondary off-line market, but it requires the specialized ability of the evaluator to be relied upon and has no unified standard. 2. The traditional machine learning method is based on models such as linear regression, random forests, gradient lifting trees (XGBoost, lightGBM) and the like, takes structural features such as mileage, age, color, province, number of passes and the like of a second hand vehicle as input, and directly predicts the price. The prediction precision is improved through feature screening and super-parameter optimization, but the method has limited processing capacity on high-order feature intersections such as 'train-year-color-province', and the like, and is difficult to capture market segment scenes. 3. The single-tower deep learning method adopts a neural network with a single path to process the characteristics of a second-hand vehicle, and learns the nonlinear relation of the characteristics through a plurality of full-connection layers, but the model only pays attention to the 'independent characteristic-price' mapping of a single sample, and does not utilize the relevance among different samples, so that the suitability of the model for market similar vehicle types is insufficient. However, these three evaluation methods have the following drawbacks: 1. The subjectivity is strong, the efficiency is low, the manual evaluation method depends on personal experience, the evaluation result difference of the same vehicle can reach 10% -15%, the evaluation of a single vehicle generally takes several minutes to tens of minutes, and the efficiency requirement of online large-scale second-hand vehicle transaction cannot be met. 2. The feature crossing capability is weak, a traditional machine learning method (such as Xgboost and Extreme Gradient Boosting, an extreme gradient lifting tree model is a classical traditional machine learning algorithm, prediction precision is improved by constructing a plurality of decision trees and iteratively optimizing a loss function based on a gradient lifting frame) can only realize low-order crossing (such as ' vehicle age multiplied by mileage ') through ' feature combination, and can not process three-order or more feature crossing (such as ' vehicle system-province-7-day residual value quantile 50% ') and the like, so that fitting precision to market conditions is low. 3. The sample relevance is not utilized, namely the existing single-tower Deep learning method and a twin Network scheme either do not consider price differences among samples or measure similarity only through simple vector distance, the sample price difference is not used as a training sample price difference, a Deep-learning architecture for processing high-dimensional sparse features is not introduced, a high-order interaction relation among features is explicitly learned through a 'Cross layer', complex nonlinear features are learned through a 'Deep layer', and the model is suitable for scenes needing accurate feature intersection such as price evaluation) architecture optimization feature intersection, so that the model cannot accurately learn the mapping relation of the 'sample relevance-price difference'. Disclosure of Invention Therefore, the application provides a training method and device for a secondary handcart price evaluation model based on deep learning, which are used for solving the problems of weak feature intersection and insufficient utilization of sample relevance in a secondary handcart price prediction method in the prior art. In order to achieve the above object, the present application provides the following technical solutions: in a first aspect, a training method for a second-hand vehicle price evaluation model based on deep learning includes: step 1, acquiring