Second-hand car transaction information management method

Abstract

The invention discloses a second-hand car transaction information management method, and relates to the field of data: obtaining second-hand car historical transaction data, and performing hierarchical processing on the obtained data to obtain a hierarchical data set; the hierarchical data set adopts a brand-vehicle series-vehicle type three-level data structure; establishing a hierarchical graph neural network model through transfer learning according to the hierarchical data set; according to the second-hand car historical transaction data, establishing an evaluation model based on a residual value rate; mapping to-be-evaluated vehicle data to corresponding nodes in the hierarchical data set, and calculating a data sufficiency coefficient of the to-be-evaluated vehicle; according to the data sufficiency coefficient, selecting a hierarchical graph neural network model or an evaluation model to carry out second-hand car price prediction; and performing second-hand car transaction information management according to the predicted second-hand car price. As the sparse data vehicle type price prediction precision is low, the second-hand vehicle price prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image data processing, and in particular to a used car transaction information management method. BACKGROUND

[0002] With the continuous maturity of the automobile consumption market and the acceleration of the replacement of new cars, the used car transaction market is showing a rapid growth trend. As the core link of the transaction, the price evaluation of the used car directly affects the market transaction efficiency and the interest protection of the participants. Accurate price prediction not only can provide reasonable pricing reference for both buyers and sellers, but also can reduce transaction risks and improve market transparency and liquidity. Traditional methods of evaluating the price of a used car mainly rely on the experience of professional evaluators or linear regression analysis based on simple statistical models. However, these methods have obvious limitations: on the one hand, manual evaluation is affected by the personal experience and subjective factors of the evaluator, making it difficult to ensure the consistency and objectivity of the evaluation results; on the other hand, simple statistical models cannot effectively handle the complex nonlinear relationships and multidimensional influencing factors in the used car market, especially when facing complex factors such as brand effect, vehicle model difference, and configuration change, the prediction accuracy is often insufficient.

[0003] In the practical application of used car price prediction, the problem of data sparsity is particularly prominent. The automobile market has typical long-tail distribution characteristics, and popular car models are traded frequently with sufficient data, while a large number of unpopular car models or specific configuration versions have relatively few transaction records. This data imbalance phenomenon leads to serious technical challenges: for popular car models with sufficient transaction records, existing machine learning methods can achieve good prediction results; but for data-sparse car models, the model often overfits or has low prediction accuracy due to insufficient training samples.

[0004] Existing technologies attempt to alleviate the sparsity problem by increasing similar car model data or using collaborative filtering methods, but these methods have the following shortcomings: first, the definition of similar car models is often too rough, only based on brand or car series, ignoring the differences between car models in configuration, performance, and other key dimensions; second, there is a lack of effective hierarchical information utilization mechanism, which cannot fully exploit the inherent rules contained in the automobile classification system; finally, the information aggregation method is single, and it is difficult to balance the weight distribution of similar car model information and cross-class similar car model information.

[0005] Therefore, there is an urgent need for a technical solution to improve the price prediction accuracy of sparse data car models. SUMMARY

[0006] In view of the low price prediction accuracy of sparse data car models, the present application provides a used car transaction information management method, which improves the price prediction accuracy of used cars through hierarchical graph neural networks and multi-source information aggregation.

[0007] The present application provides a method for managing used car transaction information, including: S1, obtaining historical used car transaction data, and performing hierarchical processing on the obtained data to obtain a hierarchical data set; the hierarchical data set adopts a three-level data structure of brand-car series-car model, wherein the brand node is the root node, the car series node is the intermediate node, and the car model node is the leaf node, and each node stores corresponding transaction records and price data; S2, establishing a hierarchical graph neural network model through transfer learning based on the hierarchical data set; S3, establishing an evaluation model based on residual value rate based on the historical used car transaction data; S4, mapping the vehicle data to be evaluated to the corresponding node in the hierarchical data set, and calculating the data adequacy coefficient of the vehicle to be evaluated; selecting a hierarchical graph neural network model or an evaluation model to predict used car prices based on the data adequacy coefficient; S5, managing used car transaction information based on the predicted used car prices.

[0008] Furthermore, a hierarchical data set is obtained, including: collecting historical transaction data of used cars, which includes vehicle basic information, transaction price, transaction time, vehicle configuration parameters, mileage and vehicle condition grade data fields; performing a primary classification of historical transaction data according to brand affiliation, and grouping all transaction records of the same brand under the corresponding brand node; under each brand node, performing a secondary classification of historical transaction data according to vehicle series affiliation, establishing a vehicle series intermediate node, and grouping transaction records of the same vehicle series under the corresponding vehicle series node; under each vehicle series node, performing a tertiary classification of transaction records according to vehicle model, establishing a vehicle model leaf node, and storing transaction records of the same vehicle model in the corresponding vehicle model node; calculating the statistical parameters of the brand node, vehicle series node and vehicle model node respectively, and the statistical parameters include the number of transaction records, average price, price variance and transaction time; among which, the statistical parameters of the brand node are calculated by summarizing the data of all subordinate vehicle series nodes; the statistical parameters of the vehicle series node are calculated by summarizing the data of all subordinate vehicle model nodes, and generating a three-level data structure of brand-vehicle series-vehicle model.

[0009] Further, the hierarchical graph neural network model is established through migration learning, comprising: according to the three-level data structure of brand-vehicle series-vehicle model, a hybrid graph structure containing hierarchical edges and similarity edges is constructed; according to the hybrid graph structure, node features are extracted and multi-dimensional feature vectorization processing is performed to obtain a node feature matrix containing statistical features, configuration features and time features; the brand layer data in the node feature matrix is subjected to graph convolutional neural network training to obtain brand layer graph neural network weight parameters W1; according to the brand layer graph neural network weight parameters W1, migration learning training from the brand layer to the vehicle series layer is performed to obtain vehicle series layer graph neural network weight parameters W2; according to the vehicle series layer graph neural network weight parameters W2, migration learning training from the vehicle series layer to the vehicle model layer is performed to obtain vehicle model layer graph neural network weight parameters W3; a sparse vehicle model node with a transaction record quantity less than a preset threshold in the statistical features is obtained, and multi-source information aggregation processing based on hierarchical edges and similarity edges is performed on the sparse vehicle model node to obtain a sparse node enhanced feature matrix; and according to the weight parameters W1 to W3 and the sparse node enhanced feature matrix, a hierarchical graph neural network model is constructed.

[0010] In particular, the sparse vehicle model can only obtain information from the same vehicle series, but the vehicle models in the same vehicle series may have large configuration differences (such as the 1.4T and 3.0T versions of Audi A4); there is a lack of connection between similar vehicle models across vehicle series and brands (such as the competitive relationship between BMW 3 series and Audi A4).

[0011] In this application, on the one hand, a similarity calculation mechanism based on vehicle configuration parameters is added, for example, the similarity between Audi A4 1.4T and BMW 3 series 1.5T can be identified, instead of being forced to be aggregated with Audi A4 3.0T, thereby improving the accuracy of information aggregation.

[0012] On the other hand, the original single information aggregation of the same vehicle series is expanded to multi-source information aggregation; the sparse vehicle model can not only obtain information from the same vehicle series, but also obtain information from similar configuration across brands, thereby significantly improving the sparse processing capability.

[0013] Further, the hybrid graph structure containing hierarchical edges and similarity edges is constructed, comprising: extracting the node relationship in the three-level data structure, and marking the brand node, vehicle series node and vehicle model node as different types of graph vertices; establishing a hierarchical edge connection relationship, establishing a directed edge between the brand node and its subordinate vehicle series node, and establishing a directed edge between the vehicle series node and its subordinate vehicle model node to form a hierarchical graph structure; extracting vehicle configuration parameters from the transaction records stored in the vehicle model node, including engine displacement, maximum power, vehicle body length, vehicle body width, vehicle body height, official guide price and other numerical configuration data; and performing Z-score standardization processing on the configuration data to eliminate the influence of different parameter dimensions and numerical ranges.

[0014] The weighted Euclidean distance algorithm is used to calculate the configuration similarity between vehicle model nodes. The calculation formula is:

[0015] Among them, w k is the weight coefficient of the kth configuration parameter; set the similarity threshold θ. When the similarity sim(i,j) between vehicle model nodes is greater than θ, establish an undirected similarity edge between the corresponding vehicle model nodes; merge the hierarchical edges and similarity edges to construct a hybrid graph topology containing hierarchical and similarity relationships, forming an adjacency matrix representation of the graph.

[0016] Furthermore, the weight parameter W1 of the brand layer graph neural network is obtained, which includes: extracting brand layer node data by node type from the node feature matrix, obtaining brand node statistical feature data for statistical parameter calculation; performing data aggregation processing on the vehicle model node configuration features under each brand node, calculating the mean and standard deviation of the configuration parameters, and generating brand-level configuration feature data; splicing the statistical feature data and configuration feature data by column to form the brand layer feature matrix X brand ;

[0017] From the brand layer feature matrix X brand Extract statistical feature column data from the dataset and calculate the average price range overlap, price variance similarity, and transaction activity similarity between brands respectively; perform weighted summation on the three similarity indicators to calculate the comprehensive similarity value between brands; binarize the comprehensive similarity value according to the preset similarity threshold to generate inter-brand connection relationship data; convert the connection relationship data into an adjacency matrix format and add self-connection markers on the diagonal position to form the brand layer adjacency matrix A brand ;

[0018] Statistical brand layer adjacency matrix A brand The number of non-zero elements in each row of , generate the degree vector; convert the degree vector into the angle matrix D brand ; According to the normalization formula Perform matrix operations to obtain normalized adjacency matrix data;

[0019] Initialize the first layer weight matrix is a random value matrix of [feature dimension × 64], initializes the first layer bias vector b1 to a 64-dimensional zero vector; initializes the second layer weight matrix is a random value matrix of [64×32], and the second layer bias vector b2 is initialized to a 32-dimensional zero vector;

[0020] Perform the first layer of graph convolution matrix operation: Obtain the first-layer linear output data; perform ReLU activation function processing on the first-layer linear output data to generate the first-layer hidden feature matrix H1;

[0021] performing a second layer graph convolution matrix operation: obtaining brand layer prediction output data H2;

[0022] extracting average price data from the statistical parameters as a real label vector y true ; calculating the mean square error of the prediction output data H2and the real label vector y true ; using the Adam optimizer to calculate the gradient update amount of the weight matrix and the bias vector based on the loss value; applying the gradient update amount to complete the iterative update of the parameter data;

[0023] calculating the loss value of the validation set data after each round of training, and comparing it with the historical minimum loss value;

[0024] when the validation set loss value is less than the historical minimum value, the current parameter data is saved as the optimal parameter; the optimal parameter data is organized as a weight parameter set W1, which is used as the initialization parameter data for the vehicle series layer transfer learning.

[0025] In particular, the brand nodes as root nodes are indeed independent of each other, lacking a natural hierarchical attribution relationship. If direct extraction is performed for graph convolution, an adjacency matrix of almost zero will be obtained, resulting in interruption of information propagation. In the present application, by "binarizing the comprehensive similarity value according to a preset similarity threshold", the system converts the originally isolated brand nodes into a graph structure with stable connection relationship. The similarity threshold mechanism ensures that only truly similar brands (such as Mercedes-Benz and BMW, Toyota and Honda) are connected, avoiding unreasonable brand association (such as Rolls-Royce and Geely). The reconstructed adjacency matrix A brand has moderate sparsity, which not only maintains the difference between brands, but also ensures effective information propagation in the graph convolution process. Through the convolution calculation, each brand node can learn the pricing rules and market characteristics of similar brands, thereby improving the generalization ability of the weight parameter W1. This "isolated node → similarity connection → effective graph structure" technical path fundamentally solves the feasibility problem of brand layer graph neural network training.

[0026] Further, the vehicle series layer graph neural network weight parameter W2 is obtained, including: filtering and extracting vehicle series layer node data from the node feature matrix according to node types, obtaining vehicle series statistical feature data calculated by vehicle series node statistical parameters; performing data aggregation processing on the configuration features of each vehicle type node under the vehicle series node, calculating the mean and standard deviation of the configuration parameters, and generating vehicle series level configuration feature data; concatenating the vehicle series statistical feature data and the vehicle series level configuration feature data by column to form a vehicle series layer feature matrix X series ; ​

[0027] From the vehicle series layer feature matrix X series , the average price interval overlap degree, price variance similarity, and transaction activity similarity between vehicle series are calculated respectively; the similarity between vehicle series of the same brand is enhanced in weight, and the similarity between vehicle series across brands is adjusted in weight; the adjusted similarity indexes are weighted and summed to calculate the comprehensive similarity value between vehicle series; the comprehensive similarity value is binarized according to the preset vehicle series similarity threshold to generate connection relationship data between vehicle series; the connection relationship data is converted into an adjacency matrix format, and a self-connection identifier is added to the diagonal position to form a vehicle series layer adjacency matrix A series ;

[0028] The number of non-zero elements in each row of the vehicle series layer adjacency matrix A series is counted to generate a degree vector of the vehicle series layer;

[0029] The degree vector of the vehicle series layer is converted into a diagonal matrix D series ;

[0030] The matrix operation is performed according to the normalization formula to obtain the normalized adjacency matrix data of the vehicle series layer;

[0031] The first layer weight matrix and the bias vector b1 are extracted from the brand layer graph neural network weight parameters W1 as the initial weight and the initial bias of the first layer of the vehicle series layer; The second layer weight matrix and the bias vector b2 are extracted from the brand layer graph neural network weight parameters W1 as the initial weight and the initial bias of the second layer of the vehicle series layer; According to the dimensions of the vehicle series layer feature matrix X series , the dimensions of the initial weight are adapted to ensure the compatibility of matrix operation;

[0032] The first layer graph convolution matrix operation is performed: to obtain the first layer linear output data of the vehicle series layer; the first layer linear output data of the vehicle series layer is processed by the ReLU activation function to generate the first layer hidden feature matrix The second layer graph convolution matrix operation is performed: to obtain the prediction output data of the vehicle series layer

[0033] The average price data is extracted from the vehicle series node statistical parameters as the real label vector y true,series of the vehicle series layer;

[0034] The prediction output data of the vehicle series layer is calculated with the real label vector ytrue,series Mean Squared Error of: Set a small learning rate for fine-tuning training, use the Adam optimizer to calculate the gradient update amount of the weight matrix and bias vector based on the vehicle series layer loss value;

[0035] Apply the gradient update amount to Update the fine-tuning of the vehicle series layer parameters;

[0036] Calculate the loss value of the vehicle series layer validation set data after each round of fine-tuning training, and compare it with the historical minimum loss value;

[0037] When the vehicle series layer validation set loss value is less than the historical minimum value, the current Parameter data is saved as the optimal vehicle series layer parameter; after completing the preset number of rounds of fine-tuning training, the optimal vehicle series layer parameter data is organized as a weight parameter set W2, which is used as the initialization parameter data for vehicle model layer transfer learning.

[0038] In particular, the vehicle series layer has higher data granularity and more complex correlation than the brand layer. The number of vehicle series nodes is much larger than the number of brand nodes, and the similarity judgment between vehicle series is more complex (such as the competitor relationship between BMW 3 series and Mercedes-Benz C-class vs. the same brand relationship between BMW 3 series and BMW X3). The present application "combines the brand ownership relationship between vehicle series, and adjusts the weight of the similarity between vehicle series in the same brand and the weight of the similarity between vehicle series across brands", so that the system integrates the brand gene factor in the similarity calculation. The vehicle series in the same brand naturally have similar design concepts, manufacturing processes and market positioning, so they need to be given higher connection weights; while the similarity across brands is more reflected in the functional level, and needs to be adjusted moderately to avoid excessive connection. This design ensures that the vehicle series layer adjacency matrix A series can reflect the market competition relationship and maintain the product system logic within the brand, forming a more reasonable graph topology.

[0039] Further, obtaining the vehicle model layer graph neural network weight parameter W3 includes: filtering and extracting vehicle model layer node data from the node feature matrix according to node types to obtain vehicle model statistical feature data calculated based on vehicle model node statistical parameters; directly extracting vehicle configuration parameters from the transaction records stored in the vehicle model nodes, including engine displacement, maximum power, vehicle size, official guide price and other configuration data, and generating vehicle configuration feature data after standardization processing; concatenating the vehicle model statistical feature data and the vehicle configuration feature data by column to form a vehicle model layer feature matrix X model ;

[0040] Extracting statistical feature column data from the vehicle model layer feature matrix X model , respectively calculating the average price interval overlap, price variance similarity and transaction activity similarity between vehicle models; extracting configuration feature column data from the vehicle model layer feature matrix Xmodel Extract configuration feature column data from the dataset and use the weighted Euclidean distance algorithm to calculate the configuration similarity between models. Combine the series and brand affiliation between models to enhance the weight of the similarity between models within the same series, adjust the medium weight of the similarity between models across series of the same brand, and reduce the weight of the similarity between models across brands. Perform weighted fusion of statistical similarity and configuration similarity to calculate the comprehensive similarity value between models. Binarize the comprehensive similarity value according to the preset model similarity threshold to generate the connection relationship data between models. Convert the connection relationship data into the adjacency matrix format and add self-connection marks on the diagonal position to form the model layer adjacency matrix A. model ;

[0041] Statistical model layer adjacency matrix A model The number of non-zero elements in each row of , generates the vehicle model layer degree vector; converts the vehicle model layer degree vector into the diagonal matrix D model ; According to the normalization formula: Perform matrix operations to obtain the normalized adjacency matrix data of the vehicle model layer;

[0042] Extract the first layer weight matrix from the vehicle layer graph neural network weight parameter W2 and the bias vector As the initial weight of the first layer of the vehicle model layer and initial bias Extract the second layer weight matrix from the vehicle layer graph neural network weight parameter W2 and the bias vector As the initial weight of the second layer of the vehicle model layer and initial bias According to the vehicle model layer feature matrix X model The initial weights are adapted to the dimensions to ensure the compatibility of matrix operations.

[0043] Perform the first layer of graph convolution matrix operation: Get the first layer linear output data of the vehicle model layer; perform ReLU activation function on the first layer linear output data of the vehicle model layer to generate the first layer hidden feature matrix of the vehicle model layer Perform the second layer of graph convolution matrix operation: Get the model layer prediction output data

[0044] Extract the average price data from the vehicle model node statistical parameters as the true label vector y of the vehicle model layer true,model ; Calculate the model layer prediction output data and the true label vector y of the vehicle model layer true,model The mean square error is:

[0045] A smaller learning rate than the vehicle series layer is set for fine-tuning training, and the Adam optimizer is used to calculate the gradient update amount of the weight matrix and bias vector based on the vehicle type layer loss value;

[0046] The gradient update amount is applied to the vehicle type layer parameters in a smaller range The fine-tuning update of the vehicle type layer parameters is completed.

[0047] The loss value of the vehicle type layer validation set data is calculated after each round of fine-tuning training, and compared with the historical minimum loss value.

[0048] When the vehicle type layer validation set loss value is less than the historical minimum value, the current The parameter data is saved as the optimal vehicle type layer parameter; after the preset number of rounds of fine-tuning training, the optimal vehicle type layer parameter data is organized as a weight parameter set W3, which is used as the final vehicle type layer prediction parameter of the hierarchical graph neural network model.

[0049] Further, obtaining the sparse node enhanced feature matrix comprises: extracting the transaction record quantity data of each vehicle type node from the vehicle type node statistical parameters; calculating the distribution statistics of the transaction record quantities of all vehicle type nodes, including the mean, median and quartile; according to a preset threshold value, marking the vehicle type nodes with transaction record quantities less than the threshold value as sparse vehicle type nodes; generating a sparse vehicle type node index list to record the position information of the sparse nodes in the vehicle type layer feature matrix; based on the hierarchical edge relationship data of the mixed graph structure, finding the vehicle series affiliation information of each sparse vehicle type node; under the same vehicle series node, extracting other vehicle type nodes except the sparse vehicle type nodes to construct a hierarchical neighbor node set; extracting the feature vector data of the hierarchical neighbor nodes from the vehicle type layer feature matrix X model ; arranging the feature vectors of the hierarchical neighbor nodes according to the node numbers to form a hierarchical neighbor feature matrix H hierarchy .

[0050] Based on the similarity edge relationship data of the mixed graph structure, finding the similar vehicle type nodes directly connected to each sparse vehicle type node; selecting the nodes across vehicle series and brands from the similar vehicle type nodes to construct a similarity neighbor node set; extracting the feature vector data of the similarity neighbor nodes from the vehicle type layer feature matrix X model ; arranging the feature vectors of the similarity neighbor nodes according to the similarity degrees to form a similarity neighbor feature matrix H similarity .

[0051] Calculating the similarity scores of the sparse vehicle type node feature vector and each neighbor feature vector in the hierarchical neighbor feature matrix H hierarchy ; using the softmax function to normalize the similarity scores to obtain the hierarchical attention weight vector a hierarchy ; calculating the similarity scores of the sparse vehicle type node feature vector and each neighbor feature vector in the hierarchical neighbor feature matrix Hsimilarity The similarity score of each neighbor feature vector; the similarity score is normalized using a softmax function to obtain a similarity attention weight vector α similarity ;

[0052] Perform hierarchical neighbor feature weighted aggregation operation: Get hierarchical aggregation feature vector;

[0053] Perform similarity neighbor feature weighted aggregation operation: Get similarity aggregation feature vector; dynamically calculate fusion weight β according to the number of hierarchical neighbor nodes and the number of similarity neighbor nodes similarity and β similarity ;

[0054] Perform double feature fusion operation: V enhanced = β hierarchy × V hierarchy + β similarity × V similarity , generate enhanced feature vector of sparse vehicle model node.

[0055] According to the sparse vehicle model node index list, locate the row position of each sparse node in the vehicle model layer feature matrix X model ; replace the original sparse node feature vector data with the corresponding enhanced feature vector V enhanced ;

[0056] In particular, the present application, hierarchical neighbor (other vehicle models in the same vehicle series): ensures the consistency of brand genes and vehicle series positioning, but may have the problem of too large configuration difference (such as 1.4T and 3.0T versions of Audi A4); Similar neighbor (similar vehicles across vehicle series and brands): based on the functional similarity of configuration parameters, it can provide more accurate reference information, but may lack brand premium and other factors; By "selecting nodes across vehicle series and brands from similar vehicle model nodes", the system can break through the traditional brand-vehicle series boundary restrictions and find truly valuable data sources for sparse vehicle models.

[0057] Further, the construction of hierarchical graph neural network model includes: based on brand layer graph neural network weight parameter W1, vehicle series layer graph neural network weight parameter W2, vehicle model layer graph neural network weight parameter W3, constructing three-layer cascaded graph neural network architecture; connect the three-layer cascaded architecture in the order of brand layer→vehicle series layer→vehicle model layer to form a top-down hierarchical prediction pipeline; set a feature dimension conversion module between each level to ensure that the dimension of the upper layer output feature matches the dimension of the lower layer input feature; construct a hierarchical feature transmission mechanism to transmit the abstract features learned by the upper layer to the lower layer for refinement processing level by level;

[0058] The brand layer weight parameter W1 is decomposed into a brand layer first layer weight matrix a second layer weight matrix and a corresponding bias vector The vehicle series layer weight parameter W2 is decomposed into a vehicle series layer first layer weight matrix a second layer weight matrix and a corresponding bias vector The vehicle model layer weight parameter W3 is decomposed into a vehicle model layer first layer weight matrix a second layer weight matrix and a corresponding bias vector A weight parameter management dictionary is constructed, and all weight matrices and bias vectors are organized according to levels and network levels, so as to facilitate model calling and updating.

[0059] A sparse node enhanced feature matrix X is obtained model,enhanced , and the original vehicle model layer feature matrix X is replaced model ; data integrity inspection is performed on the enhanced feature matrix, to ensure that all sparse vehicle model nodes have completed feature enhancement processing; statistical differences between the feature matrices before and after enhancement are calculated, including mean, variance and distribution changes of the feature vectors, to verify the effectiveness of the enhancement effect.

[0060] The enhanced feature matrix is subjected to dimension compatibility inspection with the vehicle model layer weight parameter W3, to ensure the correctness of subsequent graph convolution calculation.

[0061] A brand layer prediction submodule is constructed, graph convolution calculation is performed on the brand layer feature matrix X brand , the brand layer adjacency matrix A brand and the weight parameter W1, to obtain a brand layer prediction result P brand ; a vehicle series layer prediction submodule is constructed, graph convolution calculation is performed on the vehicle series layer feature matrix X series , the vehicle series layer adjacency matrix A series and the weight parameter W2, to obtain a vehicle series layer prediction result P series ; a vehicle model layer prediction submodule is constructed, graph convolution calculation is performed on the enhanced feature matrix X model,enhanced , the vehicle model layer adjacency matrix A model and the weight parameter W3, to obtain a vehicle model layer prediction result P model ; a hierarchical weight fusion strategy is designed, and the fusion weights of the three level prediction results are dynamically adjusted according to the data sufficiency of the vehicle to be predicted.

[0062] The data sufficiency coefficients α brand , α series and α model of the vehicle to be predicted in the brand layer, the vehicle series layer and the vehicle model layer are calculated; normalization processing is performed according to the data sufficiency coefficients, to obtain a fusion weight vector:

[0063] [w brand ,w series ,w model ]=softmax([α brand ,α series ,α model ]); perform weighted fusion operation: P final =w brand ×P brand +w series ×P series +w model ×P model , get the final predicted price; post-process the fusion prediction results, including price range rationality check and outlier filtering;

[0064] Furthermore, establishing a residual value-based valuation model involves: obtaining historical used car transaction data from a hierarchical dataset, extracting key data fields such as vehicle basic information, transaction price, transaction time, mileage, and vehicle condition grade; querying the corresponding official guide price of new cars based on the vehicle basic information, and establishing a corresponding relationship dataset between new car prices and used car transaction prices; calculating the vehicle age for each transaction record (vehicle age = transaction time - vehicle factory time), and generating vehicle age characteristic data; cleaning and standardizing the mileage data, eliminating abnormal mileage records, and generating standardized mileage characteristic data;

[0065] The residual value rate for each transaction record is calculated based on the ratio of the used car transaction price to the corresponding official suggested retail price of a new car: residual value rate = used car transaction price / official suggested retail price of a new car. The calculated residual value rate is then tested for rationality, with upper and lower thresholds set to filter out abnormal data records. The residual value rate data is grouped by vehicle age, including age ranges of less than 1 year, 1-3 years, 3-5 years, 5-8 years, and over 8 years. The distribution characteristics of the residual value rate within each age range are analyzed, including the mean, median, quantile, and standard deviation.

[0066] Construct a feature vector of factors influencing residual value, including characteristic dimensions such as vehicle age, mileage, vehicle condition, brand influence factor, and vehicle model influence factor. The brand and vehicle model influence factors are calculated based on the statistical parameters of the brand and vehicle model nodes. A multivariate linear regression algorithm is used to establish a residual value prediction model: residual value = β0 + β1 × vehicle age + β2 × mileage + β3 × vehicle condition + β4 × brand factor + β5 × vehicle model factor + ε. The least squares method is used to estimate the parameters of the regression coefficients β0 to β5, and significance tests and model fit assessments are performed.

[0067] According to the three-level data structure of brand-vehicle series-vehicle model, a brand-level, vehicle series-level and vehicle model-level residual value rate model is respectively constructed; the brand-level residual value rate model is trained using average residual value rate data at the brand level and is suitable for value evaluation at the brand level; the vehicle series-level residual value rate model is trained using residual value rate data at the vehicle series level and increases vehicle series characteristic variables based on the brand-level model; and the vehicle model-level residual value rate model is trained using residual value rate data of a specific vehicle model and contains complete vehicle configuration characteristic information;

[0068] The historical transaction data is divided into a training set and a test set according to a ratio of 8:2, and the training set is used for model parameter estimation; the mean absolute error MAE, the root mean square error RMSE and the determination coefficient R 2 of residual value rate prediction are calculated on the test set; the cross-validation method is used to evaluate the generalization ability of the model to prevent overfitting; and the model parameters are optimized according to the verification results, including feature selection, regularization coefficient setting and abnormal value processing strategy;

[0069] The optimized hierarchical residual value rate model is organized into a unified evaluation model system, including three levels of sub-models of brand level, vehicle series level and vehicle model level; a model selection strategy is constructed to automatically select a residual value rate model at a suitable level according to the data sufficiency of a vehicle to be evaluated; a prediction interface of the residual value rate model is established to input vehicle basic information and characteristic data and output predicted residual value rate and confidence interval; and the residual value rate evaluation model is used as a candidate prediction model to form a complement with a hierarchical graph neural network model.

[0070] Further, calculating the data sufficiency coefficient α of the vehicle to be evaluated includes: obtaining basic information of the vehicle to be evaluated, including brand, vehicle series, vehicle model, vehicle age, mileage, vehicle condition grade and configuration parameters; mapping the vehicle to be evaluated to corresponding brand nodes, vehicle series nodes and vehicle model nodes of the hierarchical data set according to the brand, vehicle series and vehicle model information of the vehicle; and extracting transaction record quantity, latest transaction time and price statistical parameters of the vehicle to be evaluated at each level node;

[0071] The data sufficiency coefficient α of the vehicle to be evaluated at the brand level, the vehicle series level and the vehicle model level is respectively calculated, and the calculation formula is: A data sufficiency threshold α threshold is set. threshold When α≥α threshold , the hierarchical graph neural network model is selected for prediction; when α<α threshold , the residual value rate-based evaluation model is selected for prediction; and the level with the highest data sufficiency and meeting the threshold requirement is selected for price prediction in the priority order of vehicle model level→vehicle series level→brand level.

[0072]

[0073] Alpha (Data Adequacy Coefficient): Alpha = 0.5 indicates that the data adequacy reaches the critical state; Alpha >= 0.8 is generally considered adequate data, suitable for using graph neural network model; Alpha < 0.5 is generally considered sparse data, and it is recommended to use residual rate evaluation model.

[0074] N (Number of Valid Transaction Records): Statistical time window: transaction records in the past 24 months; hierarchical calculation: brand layer N: the total number of transaction records of all models under the brand; vehicle series layer N: the total number of transaction records of all models under the vehicle series; vehicle model layer N: the total number of transaction records of the specific vehicle model.

[0075] N0 (Adequacy Reference Value): Preferably: brand layer N0 = 200 (large sample reference); vehicle series layer N0 = 100 (medium sample reference); vehicle model layer N0 = 50 (small sample reference).

[0076] σ N (Sensitivity Adjustment Parameter): Preferably: brand layer σ N = 80; vehicle series layer σ N = 40; vehicle model layer σ N = 20; setting principle: σ N ≈ N0 / 2.5, to ensure a reasonable transition interval.

[0077] Compared with the prior art, the application has the following advantages:

[0078] In one aspect of the application, the inherent hierarchical structure characteristics of the automobile market are utilized through the transfer learning mechanism of brand layer -> vehicle series layer -> vehicle model layer. The brand layer has the most abundant transaction data, and the weight parameter W1 obtained through the backpropagation training using the average price of the brand node statistical parameter as the supervision label can learn the brand-level price rules and market characteristics. The knowledge learned by the upper layer is passed down level by level, so that the data sparse vehicle model layer can inherit the rich experience of the brand layer and the vehicle series layer. This "knowledge distillation" mechanism fundamentally changes the dilemma of "isolated learning" of sparse vehicle models. Even if the transaction records of a specific vehicle model are few, it can also obtain a stable and reliable price prediction basis through hierarchical weight inheritance, thereby significantly improving the prediction accuracy and model robustness in the sparse data scenario.

[0079] On the other hand, this application breaks the limitations of traditional single-dimensional information aggregation through the hybrid graph structure design of "hierarchical edges + similarity edges". Hierarchical edges ensure the inheritance of brand genes (such as the luxury positioning of the Audi brand), while similarity edges establish cross-brand functional associations based on configuration parameters (such as the performance characteristics of the 1.4T engine). "The attention mechanism is used to calculate the weight coefficients of the hierarchical neighbor feature set and the similarity neighbor feature set respectively", which can dynamically adjust the contribution of different information sources according to the characteristics of specific models. When a certain model is very different from other models in the same series, the system will automatically increase the weight of the similarity neighbors; otherwise, it will rely more on hierarchical neighbor information. This adaptive multi-dimensional information fusion mechanism ensures that sparse models can obtain information from the most relevant data sources, thereby achieving more accurate feature representation and price prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The present application will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein:

[0081] Figure 1 is an exemplary flow chart of a method for managing used car transaction information according to some embodiments of the present application;

[0082] Figure 2 This is a schematic diagram of brand data distribution in this embodiment;

[0083] Figure 3 is an exemplary flow chart of a hierarchical graph neural network model according to some embodiments of the present application;

[0084] Figure 4 Schematic diagram of sparse vehicle type distribution in this embodiment. DETAILED DESCRIPTION

[0085] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0086] like Figure 1As shown, the historical transaction data of the used car is acquired, and the acquired data is hierarchically processed to obtain a hierarchical data set; the hierarchical data set adopts a three-level data structure of brand-vehicle series-vehicle model, wherein the brand node is the root node, the vehicle series node is the intermediate node, and the vehicle model node is the leaf node, each node stores the corresponding transaction record and price data; according to the hierarchical data set, a hierarchical graph neural network model is established through transfer learning; according to the historical transaction data of the used car, an evaluation model based on the residual value rate is established; the data of the vehicle to be evaluated is mapped to the corresponding node in the hierarchical data set, and the data sufficiency coefficient of the vehicle to be evaluated is calculated; the hierarchical graph neural network model or the evaluation model is selected according to the data sufficiency coefficient to predict the price of the used car; and the used car transaction information is managed according to the predicted price of the used car.

[0087] The transaction records containing complete configuration information are extracted from the transaction platform database, and the data field coverage rate reaches more than 95%. The main data fields include: vehicle basic information (brand, vehicle series, vehicle model), transaction price, transaction time, vehicle configuration parameters (engine displacement, maximum power, vehicle size, official guide price), mileage, vehicle condition grade and other key information.

[0088] The Audi brand contains 23456 transaction records, including 3892 A4 series, 4123 A6 series, 2567 Q5 series, etc. Each record records the engine displacement (range 1.4L-6.0L), maximum power (range 90kW-450kW), vehicle length (range 4200mm-5300mm), vehicle width (range 1750mm-2100mm), vehicle height (range 1400mm-1900mm), official guide price (range 180-150 million yuan) and other configuration parameters.

[0089] According to the brand affiliation relationship, 458623 transaction records are distributed to 68 brand nodes, such as Figure 2 As shown, the data distribution of the three German luxury brands is: Audi 23456, BMW 21789, and Mercedes-Benz 25234, accounting for 15.3% of the total data amount. Japanese brands Toyota 18967, Honda 16234, and Nissan 14123, accounting for 10.8% of the total data amount. See Table 1 for main brand statistical parameters.

[0090] Table 1 Main brand statistical parameters

[0091] brand Number of transaction records Average price (10,000 yuan) Price variance Latest trading time Audi 23456 24.8 8.7*10^10 2025-06-15 BMW 21789 27.2 9.4*10^10 2025-06-16 Benz 25234 29.6 1.1*10^11 2025-06-14 Toyota 18967 16.4 4.2*10^10 2025-06-16 Honda 16234 15.8 3.9*10^10 2025-06-15 Nissan 14123 14.2 3.1*10^10 2025-06-13

[0092] Specifically, the Audi brand node statistical parameters are: transaction record quantity 23456, average transaction price 248,000 yuan, price variance 8.7x10^10, and the latest transaction time is June 15, 2025. The BMW brand node statistical parameters are: transaction record quantity 21789, average transaction price 272,000 yuan, price variance 9.4x10^10, and the latest transaction time is June 16, 2025.

[0093] Under the Audi brand node, it is further subdivided into 21 vehicle series: A3 series 1234 records, A4 series 3892 records, A6 series 4123 records, Q3 series 967 records, Q5 series 2567 records, Q7 series 1845 records, etc. The data density of each vehicle series is significantly different, with mainstream vehicle series having sufficient data and niche vehicle series having relatively sparse data.

[0094] Audi A4 series node: transaction record quantity 3892, average transaction price 235,000 yuan, price variance 4.2x10^10, covering various power configurations such as 1.4T, 2.0T, and 3.0T. BMW 3 series node: transaction record quantity 4156, average transaction price 248,000 yuan, price variance 4.6x10^10, forming a direct competitive relationship with Audi A4.

[0095] Under the Audi A4 series, it is further subdivided into 32 specific vehicle configurations. Data-rich vehicle models: A4L 2.0T luxury type 658 records, A4L 2.0T sports type 543 records. Data-sparse vehicle models: A4L 1.4T fashion type only 23 records, A4L 3.0T quattro flagship type only 8 records.

[0096] Set the transaction record quantity threshold to 50, identify 2847 sparse vehicle models, accounting for 22.1% of the total number of vehicle models. These sparse vehicle models are mainly distributed in high-end configurations, entry-level configurations, discontinued models, etc., making it difficult for traditional prediction methods to obtain reliable results.

[0097] As shown in Figure 3 , six core configuration parameters are extracted from the 12847 vehicle model nodes: engine displacement, maximum power, vehicle body length, vehicle body width, vehicle body height, and official guide price. The configuration parameter distribution of the entire sample is: engine displacement mean 2.1L, standard deviation 0.8L; maximum power mean 156kW, standard deviation 78kW; vehicle body length mean 4567mm, standard deviation 245mm.

[0098] Z-score standardization is performed on all configuration parameters to eliminate dimensional effects. Take Audi A4L 1.4T Fashion as an example: the original configuration [1.4L, 110kW, 4818mm, 1843mm, 1432mm, 269800 yuan] is standardized to [-0.875, -0.590, 1.024, 0.356, -0.234, -0.412].

[0099] Based on the analysis of factors affecting the value of used cars, the weights of configuration parameters are set: engine displacement 0.30, maximum power 0.25, body length 0.15, body width 0.10, body height 0.10, official guide price 0.10. The weight distribution reflects the dominant role of power performance in the value of used cars.

[0100] The weighted Euclidean distance algorithm is used to calculate the configuration similarity between car models. Key findings: the similarity between Audi A4L 1.4T Fashion and BMW 320Li Fashion is 0.847, and the similarity between Audi A4L 3.0T quattro of the same car series is only 0.623. This verifies the rationality of cross-brand configuration matching and breaks the traditional brand boundary restrictions.

[0101] Set the similarity threshold to 0.75, and establish undirected similarity edges between car models with configuration similarity exceeding the threshold. Statistical results: a total of 8967 cross-brand similarity edges and 15234 cross-car series similarity edges are established, providing rich information sources for sparse car models.

[0102] The final mixed graph contains 25694 hierarchical edges (12847 brand-car series, 12847 car series-car model) and 24201 similarity edges. The average degree of the graph is 3.9, with good connectivity, laying a solid foundation for subsequent graph neural network training.

[0103] Brand layer feature matrix X brand The dimension is [68x128], including 64-dimensional statistical features (transaction quantity, average price, price variance, time characteristics, etc.) and 64-dimensional configuration features (statistical aggregation of configuration parameters of subordinate car models). The Audi brand feature vector integrates the statistical information of 23456 transaction records and the configuration distribution characteristics of subordinate car models.

[0104] Calculate the comprehensive similarity between brands. The similarity between the three German luxury brands is relatively high: Audi-BMW 0.834, Audi-Mercedes-Benz 0.812, and BMW-Mercedes-Benz 0.856. The similarity within Japanese brands: Toyota-Honda 0.789, Honda-Nissan 0.756. The similarity between cross-country brands is generally low, which is consistent with the market positioning differences, as shown in Table 2.

[0105] Table 2 Brand Similarity

[0106] brand Audi BMW Benz Audi 1 0.834 0.812 BMW 0.834 1 0.856 Benz 0.812 0.856 1

[0107] A two-layer graph convolutional neural network was constructed, with the first layer scaling from 128 to 64 dimensions and the second layer scaling from 64 to 32 dimensions. Reluctant Linear Unit (ReLU) activation function was used, and the Adam optimizer with a learning rate of 0.001 was used. After 200 epochs of training, the validation set loss converged to 0.0023, and the brand-level prediction accuracy reached 92.3%.

[0108] The weight parameter W1 obtained from brand layer training is used as the initial parameter of vehicle series layer, effectively utilizing the abstract features learned from brand layer. Vehicle series layer feature matrix X series The dimension is [356×128], and the feature density is finer than that of the brand layer.

[0109] At the vehicle series level, brand affiliation enhancement was introduced, increasing the weight of similarity within a brand by 1.2 times and reducing the weight of similarity across brands by 0.8 times. Core competitive relationships were identified: the Audi A4 and BMW 3 Series had a similarity of 0.891, and the Audi Q5 and BMW X3 had a similarity of 0.876.

[0110] Using a small learning rate of 0.0005 for fine-tuning training, the validation set loss dropped to 0.0019 after 150 epochs, and the vehicle-level prediction accuracy increased to 94.1%. Transfer learning reduced training time by 60% compared to random initialization.

[0111] Model layer feature matrix X model The dimension is [12847×128] and directly contains the original configuration parameters and statistical features. The feature granularity is the finest, including detailed configuration information such as engine displacement, power, and size.

[0112] The vehicle model layer considers both statistical and configuration similarity, with a weighted fusion ratio of 0.4:0.6. The higher weighting of configuration similarity reflects the emphasis on technical parameters at the vehicle model level. The weighting of models within the same series is increased by 1.5 times, and the weighting of cross-model series within the same brand is increased by 1.2 times. The weighting of cross-brand models remains unchanged.

[0113] Fine-tuning with a smaller learning rate of 0.0002 reduced the validation set loss to 0.0015 after 100 epochs, and the vehicle model prediction accuracy reached 95.7%. This layer-by-layer transfer learning strategy significantly improved model performance.

[0114] like Figure 4 As shown, 2,847 sparsely populated models (<50 transaction records) were identified among 12,847 models, accounting for 22.1%. These sparsely populated models are mainly distributed as follows: 1,234 high-end models (43.4%), 856 entry-level models (30.1%), 542 discontinued models (19.0%), and 215 niche models (7.5%).

[0115] Audi A4L 1.4T Fashion only 23 transactions, BMW 320Li Fashion only 27 records, Mercedes-Benz C180L Fashion only 19 records. These entry-level configuration vehicles have limited market acceptance, sparse historical transaction data, and large errors in traditional prediction methods.

[0116] Taking Audi A4L 1.4T Fashion as an example, 31 neighbor vehicles of the same vehicle line are identified. Data-rich neighbors: A4L 2.0T Luxury 658 records, A4L 2.0T Sport 543 records, A4L 2.5T quattro 412 records. On average, each sparse vehicle can obtain 28.6 features of the same vehicle line neighbors.

[0117] Through the aggregation of the same vehicle line information, the feature vector of Audi A4L 1.4T Fashion is enhanced from the original sparse state (based on 23 records) to the aggregated features of 1613 same vehicle line records, with a 70-fold expansion of the data base.

[0118] Based on the configuration similarity, cross-brand similar neighbors are identified for Audi A4L 1.4T Fashion: BMW 320Li Fashion (similarity 0.847), Mercedes-Benz C180L Fashion (similarity 0.823), Volkswagen Passat 1.4T Comfort (similarity 0.798), Cadillac ATS-L 1.5T Fashion (similarity 0.772).

[0119] Through cross-brand similarity aggregation, 145 similar vehicle transaction records are further integrated. Compared with the traditional method of using only same vehicle line data, multi-source aggregation expands the effective data base to 1758 records, with a significant data enhancement effect.

[0120] Calculate the attention weight of the sparse vehicle and the same vehicle line neighbor. The hierarchical attention distribution of Audi A4L 1.4T Fashion: A4L 2.0T Luxury weight 0.342, A4L 2.0T Sport weight 0.298, A4L 1.8T Comfort weight 0.256, other vehicles weight 0.104. The weight distribution reflects the higher relevance of the configuration similar vehicles.

[0121] Similarity attention weight distribution: BMW 320Li Fashion weight 0.386, Mercedes-Benz C180L Fashion weight 0.294, Volkswagen Passat 1.4T weight 0.213, Cadillac ATS-L weight 0.107. The cross-brand attention mechanism effectively identifies the most relevant reference vehicles.

[0122] According to the number of neighbors, the fusion weight is dynamically calculated. Audi A4L 1.4T Fashion has 31 hierarchical neighbors, 4 similarity neighbors, and a total of 35 neighbors. The hierarchical weight β hierarchy= 31 / 35 = 0.886, similarity weight β similarity = 4 / 35 = 0.114. Feature vector norm before enhancement 2.134, feature vector norm after enhancement 3.672, feature expression capacity improved by 72.1%. Through multi-source information aggregation, sparse vehicle types obtain more rich and accurate feature representation, laying a solid foundation for subsequent price prediction.

[0123] Vehicle to be evaluated: 2021 Audi A4L 1.4T Fashion, mileage 32,000 km, good condition. Data mapping results: brand layer (Audi) 23456 records, vehicle series layer (A4) 3892 records, vehicle type layer (A4L 1.4T Fashion) 23 records.

[0124] Use the sigmoid function to calculate the data sufficiency of each level: brand layer: Vehicle series layer: Vehicle type layer: Set the data sufficiency threshold to 0.8 as the selection boundary for the graph neural network model and the residual rate model. The data sufficiency of the vehicle type layer is 0.272, which is much lower than the threshold, and the data sufficiency of the vehicle series layer is 1.000, which is much higher than the threshold. According to the priority order vehicle type layer → vehicle series layer → brand layer, since the vehicle type layer data is insufficient, the vehicle series layer graph neural network model is automatically selected for prediction. This strategy not only ensures the prediction accuracy, but also ensures the reliability of the data foundation.

[0125] The traditional same vehicle series aggregation method can only use the data of 31 vehicle types of the same vehicle series when predicting the price of Audi A4L 1.4T Fashion, but these vehicle types include high-end configurations such as 3.0T quattro, which have a huge difference with the 1.4T Fashion configuration, resulting in large prediction bias. The multi-source aggregation method of this embodiment can identify cross-brand similar vehicle types such as BMW 320Li and Mercedes-Benz C180L, with higher configuration matching degree, and significantly improve the prediction accuracy. The average absolute error of sparse vehicle type prediction is reduced from 8500 yuan to 5200 yuan, with an improvement rate of 38.8%.

[0126] The system successfully identified 89.2% of the market-recognized competitive relationships, such as Audi A4 and BMW 3 series (similarity 0.891), Mercedes-Benz C-Class and BMW 3 series (similarity 0.878), and Lexus ES and Audi A6 (similarity 0.845). The recognition accuracy far exceeds traditional brand-based classification methods. Through the configuration similarity analysis, potential competitive relationships beyond traditional cognition are found, such as Buick Terran 2.0T and Audi A4L 2.0T (similarity 0.756), providing a new perspective for market analysis. This embodiment successfully breaks through the boundaries of traditional brands and vehicle series, establishes vehicle similarity based on actual configuration parameters, and solves the problem of sparse vehicle information silos. By identifying the high similarity between Audi A4L 1.4T and BMW 320Li, a more suitable reference sample is provided for sparse vehicle prediction. A dual aggregation mechanism of hierarchical neighbor features and similarity neighbor features is constructed, enabling sparse vehicles to obtain information from both the same vehicle series and cross-brand similar vehicles. The data basis has expanded from 23 to 1758. An intelligent model selection mechanism based on data sufficiency is established, which automatically selects the optimal prediction strategy according to the data conditions of different vehicles. It not only ensures the prediction accuracy of data-sufficient vehicles, but also solves the prediction problem of sparse vehicles.

[0127] The above describes the application creation and its embodiments in a schematic manner, which is not restrictive, and the application can be realized in other specific forms without departing from the spirit or essential characteristics of the application. The embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the spirit of the application, similar structural forms and embodiments can be designed without creative design, which should belong to the protection scope of the application. In addition, the word "comprising" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The words "first", "second", etc. are used to represent the name, and do not represent any specific order.

Claims

1. A method for managing second-hand car transaction information, characterized in that: include: Obtain historical used car transaction data and perform hierarchical processing on the acquired data to obtain a hierarchical data set; the hierarchical data set adopts a three-level data structure of brand-car series-model, where the brand node is the root node, the car series node is the intermediate node, and the model node is the leaf node. Each node stores the corresponding transaction record and price data; Based on the hierarchical dataset, a hierarchical graph neural network model is established through transfer learning; Establish a valuation model based on residual value rate based on historical transaction data of used cars; Map the vehicle data to be evaluated to the corresponding nodes in the hierarchical data set and calculate the data sufficiency coefficient of the vehicle to be evaluated; Select a hierarchical graph neural network model or evaluation model based on the data sufficiency coefficient to predict used car prices; Manage used car transaction information based on predicted used car prices.

2. The method for managing used car transaction information according to claim 1, characterized in that: Get a hierarchical data set, including: Collect historical transaction data of used cars, including vehicle basic information, transaction price, transaction time, vehicle configuration parameters, mileage and vehicle condition grade data fields; Classify historical transaction data into a primary category based on brand ownership, and group all transaction records of the same brand into the corresponding brand node. Under each brand node, historical transaction data is classified into two levels according to the car series affiliation, and a car series intermediate node is established to aggregate transaction records of the same car series under the corresponding car series node; Under each vehicle series node, transaction records are classified into three levels according to vehicle models, and vehicle model leaf nodes are established. Transaction records of the same vehicle model are stored in the corresponding vehicle model node. Calculate the statistical parameters of the brand node, car series node, and model node respectively. The statistical parameters include the number of transaction records, average price, price variance, and transaction time. Among them, the brand node statistical parameters are calculated by summarizing the node data of all subordinate car series; the car series node statistical parameters are calculated by summarizing the node data of all subordinate vehicle models, generating a three-level data structure of brand-car series-model.

3. The method for managing used car transaction information according to claim 2, characterized in that: Build a hierarchical graph neural network model through transfer learning, including: Based on the three-level data structure of brand-series-model, a hybrid graph structure containing hierarchical edges and similarity edges is constructed; According to the hybrid graph structure, node features are extracted and multi-dimensional feature vectorization is performed to obtain a node feature matrix containing statistical features, configuration features, and time features. Perform graph convolutional neural network training on the brand layer data in the node feature matrix to obtain the brand layer graph neural network weight parameter W1; Based on the brand layer graph neural network weight parameter W1, transfer learning training is performed from the brand layer to the vehicle series layer to obtain the vehicle series layer graph neural network weight parameter W2; Based on the weight parameter W2 of the neural network at the vehicle series level, transfer learning training is performed from the vehicle series level to the vehicle model level to obtain the weight parameter W3 of the neural network at the vehicle model level. Obtain sparse vehicle model nodes whose transaction records in the statistical features are less than a preset threshold, perform multi-source information aggregation processing on the sparse vehicle model nodes based on hierarchical edges and similarity edges, and obtain a sparse node enhanced feature matrix; A hierarchical graph neural network model is constructed based on the weight parameters W1 to W3 and the sparse node enhanced feature matrix.

4. The method for managing used car transaction information according to claim 3, wherein: Obtain the brand layer graph neural network weight parameter W1, including: Extract brand-level node data from the node feature matrix, as well as statistical features generated by brand node statistical parameters, and aggregate configuration features from subordinate vehicle model nodes to form a brand-level feature matrix; Based on the statistical parameters of the brand nodes in the brand layer feature matrix, the similarity between brands is calculated, the connection relationship between brands is established according to the preset threshold, and the brand layer adjacency matrix including self-connection is generated; Normalize the brand layer adjacency matrix to obtain a normalized adjacency matrix; Construct a multi-layer graph convolutional neural network and use the brand layer adjacency matrix and the normalized adjacency matrix to perform graph convolution calculations to obtain the brand layer prediction output; The average price of the brand node statistical parameters is used as the supervision label, the prediction error is calculated and back-propagation training is performed to obtain the brand layer graph neural network weight parameter W1.

5. The method for managing used car transaction information according to claim 3, characterized in that: Obtain the vehicle layer graph neural network weight parameter W2, including: Extract vehicle series node data from the node feature matrix, and combine the vehicle series metering features of the vehicle series node statistical parameters and the configuration features aggregated from the subordinate vehicle model nodes to form a vehicle series feature matrix. Calculate the similarity between vehicle series based on the statistical features in the vehicle series feature matrix, establish the connection relationship between vehicle series nodes based on the preset threshold, and generate a vehicle series adjacency matrix including self-connections; Normalize the vehicle-level adjacency matrix to obtain the vehicle-level normalized adjacency matrix; The brand layer graph neural network weight parameter W1 is used as the initialization parameter of the vehicle series layer graph neural network; The vehicle-level feature matrix and the vehicle-level normalized adjacency matrix are used to perform graph convolution calculations to obtain the vehicle-level prediction output. The average price of the vehicle series node statistical parameters is used as the supervision label, the prediction error is calculated and back-propagation training is performed to obtain the vehicle series layer graph neural network weight parameter W2.

6. The method for managing used car transaction information according to claim 3, characterized in that: Obtain the vehicle model layer graph neural network weight parameter W3, including: Extract vehicle-level node data from the node feature matrix, and combine the vehicle-level statistical features of the vehicle-level node statistical parameters and the configuration features aggregated from the subordinate vehicle-level nodes to form a vehicle-level feature matrix. The similarity between vehicle models is calculated based on the statistical features and configuration features in the vehicle model layer feature matrix. The connection relationship between vehicle model nodes is established based on the preset threshold, and the vehicle model layer adjacency matrix including self-connection is generated. Normalize the vehicle model layer adjacency matrix to obtain the vehicle model layer normalized adjacency matrix; The weight parameter W2 of the vehicle series layer graph neural network is used as the initialization parameter of the vehicle model layer graph neural network; The vehicle model layer feature matrix and the vehicle model layer normalized adjacency matrix are used to perform graph convolution calculations to obtain the vehicle model layer prediction output; The average price of the vehicle model node statistical parameters is used as the supervision label, the prediction error is calculated and back-propagation training is performed to obtain the vehicle model layer graph neural network weight parameter W3.

7. The method for managing used car transaction information according to claim 6, characterized in that: Get the sparse node enhanced feature matrix, including: Identify vehicle model nodes with a number of transaction records less than a threshold from the vehicle model node statistical parameters and mark them as sparse vehicle model nodes; Based on the hierarchical edge relationship of the hybrid graph structure, find other model nodes under the car series to which each sparse model node belongs, extract the features of neighboring models in the same car series, and form a hierarchical neighbor feature set; Based on the similarity edge relationship of the hybrid graph structure, similar model nodes connected to the sparse model nodes are found, and similar neighbor model features across car series and brands are extracted to form a similar neighbor feature set; The attention mechanism is used to calculate the weight coefficients of the hierarchical neighbor feature set and the similarity neighbor feature set, respectively, to obtain the hierarchical aggregation feature vector and the similarity aggregation feature vector; Perform weighted fusion of hierarchical aggregation feature vectors and similarity aggregation feature vectors to obtain enhanced feature vectors of sparse vehicle model nodes; The enhanced feature vector is updated to the corresponding sparse node position of the vehicle model layer feature matrix to obtain the sparse node enhanced feature matrix.

8. The method for managing used car transaction information according to claim 3, characterized in that: Build a hierarchical graph neural network model, including: Based on the weight parameters W1 to W3, a three-layer cascaded graph neural network structure is constructed, including the brand layer, the car series layer, and the model layer; Integrate the sparse node enhanced feature matrix into the vehicle model layer, replacing the original vehicle model layer feature matrix; Graph convolution calculations are performed on the brand layer, car series layer, and model layer respectively to obtain prediction results at each level.

9. The method for managing used car transaction information according to any one of claims 2 to 8, characterized in that: Establish a valuation model based on residual value rate, including: Obtain historical transaction data from a hierarchical dataset, extract basic vehicle information, transaction price, vehicle age, mileage, and vehicle condition rating, and obtain the corresponding official guide price for new vehicles; Calculate the residual value rate based on the transaction price of the used car and the corresponding official guide price of the new car; Construct a multi-factor residual value regression model that includes vehicle age, mileage, vehicle condition, brand influencing factors, and vehicle model influencing factors. The brand influencing factors and vehicle model influencing factors are calculated based on the statistical parameters of the corresponding nodes. Based on the three-level data structure of brand-series-model, residual value rate sub-models at the brand level, series level and model level are constructed respectively, and combined to form an evaluation model based on residual value rate.

10. The method for managing used car transaction information according to claim 9, characterized in that: Calculate the data adequacy coefficient α of the vehicle to be evaluated, including: Among them, α is the data sufficiency coefficient; N is the number of valid transaction records of the corresponding node within the preset time window; N0 is the sufficiency benchmark value; σ N is the sensitivity adjustment parameter.

Images