A used car price evaluation model training method and device based on deep learning

Abstract

The application discloses a kind of based on deep learning second-hand car price evaluation model training method and device, by obtaining second-hand car historical transaction data, and data preprocessing is obtained training sample;Sample is selected from training sample and constructs the sample pair of real sample and sample to be evaluated;Real sample and sample to be evaluated are input into the first subnetwork and the second subnetwork of double-tower model respectively, and first sample vector and second sample vector are obtained;Wherein, first subnetwork and second subnetwork all adopt deep cross network architecture;First sample vector and second sample vector are carried out element level product, and interactive feature vector is obtained, and interactive feature vector is mapped as scalar by output layer;According to the loss function of scalar and sample price difference, all parameters of double-tower model are updated by back propagation, and the second-hand car price evaluation model of training completion is obtained, solves the problem that second-hand car price prediction method feature cross is weak and sample correlation is insufficient.

Description

Technical Field

[0001] This application relates to the field of used car price assessment technology, specifically to a method and apparatus for training a used car price assessment model based on deep learning. Background Technology

[0002] With the continuous growth of car ownership and the shortening of consumers' car replacement cycle, the used car market has experienced rapid development in recent years. Against this backdrop, scientific, accurate, and efficient used car price assessment has become a crucial link connecting buyers and sellers, ensuring fair transactions, and improving market transparency. Currently, used car price assessment mainly relies on the following three assessment methods, each applied in different scenarios:

[0003] 1. Manual Appraisal Method: Appraisers determine a price range based on readily available information such as the brand, age, mileage, and condition of the used car, combined with market experience. This method is still widely used in offline used car markets, but it relies heavily on the appraiser's expertise and lacks a unified standard.

[0004] 2. Traditional Machine Learning Methods: Based on models such as linear regression, random forests, and gradient boosting trees (XGBoost, LightGBM), these methods directly predict transaction prices using structured features of used cars, such as mileage, age, color, province, and number of ownership transfers. While feature selection and hyperparameter optimization improve prediction accuracy, these methods have limited ability to handle high-order feature interactions such as "car series-year-color-province," making it difficult to capture specific market segments.

[0005] 3. Single-tower deep learning method: This method uses a single-path neural network to process the features of used cars. It learns the nonlinear relationship of features through multiple fully connected layers. However, this type of model only focuses on the "independent feature-price" mapping of a single sample and does not utilize the correlation between different samples, resulting in insufficient adaptability to the market conditions of similar models.

[0006] However, these three evaluation methods have the following drawbacks:

[0007] 1. High subjectivity and low efficiency: Manual evaluation relies on personal experience, and the evaluation results for the same vehicle can vary by 10%-15%. Moreover, the evaluation of a single vehicle usually takes several minutes to more than ten minutes, which cannot meet the efficiency requirements of large-scale online used car transactions.

[0008] 2. Weak feature cross-referencing capability: Traditional machine learning methods (such as XGBoost, Extreme Gradient Boosting, a classic traditional machine learning algorithm based on the gradient boosting framework, which improves prediction accuracy by constructing multiple decision trees and iteratively optimizing the loss function) can only achieve low-order cross-referencing (such as "vehicle age × mileage") through "feature combination", and cannot handle third-order or higher feature cross-referencing (such as "vehicle series - province - 50% of 7-day residual value"), resulting in low fitting accuracy for segmented market conditions.

[0009] 3. Insufficient utilization of sample correlation: Existing single-tower deep learning methods and Siamese network solutions either fail to consider price differences between samples or only measure similarity through simple vector distance, failing to use "sample price difference" as the training sample price difference. Furthermore, they do not introduce Deep & Cross Network (DCN, a deep learning architecture for handling high-dimensional sparse features, which explicitly learns high-order interaction relationships between features through "cross layers" and learns complex nonlinear features through "deep layers," suitable for scenarios such as price evaluation that require precise feature cross-tracing) architecture to optimize feature cross-tracing. As a result, the model cannot accurately learn the mapping relationship between "sample correlation - price difference." Summary of the Invention

[0010] To address this, this application provides a training method and apparatus for a used car price assessment model based on deep learning, in order to solve the problems of weak feature crossover and insufficient utilization of sample correlation in existing used car price prediction methods.

[0011] To achieve the above objectives, this application provides the following technical solution:

[0012] Firstly, a deep learning-based method for training a used car price assessment model includes:

[0013] Step 1: Obtain historical transaction data for used cars, and extract basic attribute features, usage attribute features, and market trend features from the historical transaction data for used cars;

[0014] Step 2: Perform data preprocessing on the basic attribute features, usage attribute features, and market trend features to obtain training samples; the data preprocessing includes missing value imputation, outlier removal, and feature encoding;

[0015] Step 3: Select one sample from the training samples as the sample to be evaluated, and select a sample from the training samples that is the same brand, model, and year as the sample to be evaluated as the real sample 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 transaction price to the official guide price.

[0016] Step 4: Input the real sample and the sample to be evaluated from the sample pair into the first sub-network and the second sub-network of the pre-constructed dual-tower model, respectively, to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross-network architecture;

[0017] Step 5: Perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and map the interaction feature vector to a scalar through the output layer;

[0018] Step 6: Using the Huber loss between the scalar and the sample price difference of the sample pair as the loss function, update all parameters of the dual-tower model through backpropagation;

[0019] Step 7: Repeat steps 3 to 6 until the preset training termination condition is met, and obtain the trained used car price evaluation model.

[0020] Optionally, in step 2, the feature encoding specifically includes: embedding and encoding the discrete features among the basic attribute features, usage attribute features, and market trend features; encoding the binary features among the basic attribute features, usage attribute features, and market trend features using 0-1 encoding; and normalizing the continuous features among the basic attribute features, usage attribute features, and market trend features.

[0021] Optionally, in step 2, outliers are identified and removed using the interquartile range method.

[0022] Optionally, in step 4, the structures of the first sub-network and the second sub-network both include: an input layer, an embedding layer, a feature concatenation layer, a cross layer, a deep layer, and a feature fusion layer.

[0023] Optionally, the cross layer has three layers, and the depth layer adopts three fully connected layers.

[0024] Optionally, in step 6, the loss function is:

[0025]

[0026] Among them, Z i For scalars, label iFor the sample pair residual difference, This is a hyperparameter.

[0027] Secondly, a training device for a deep learning-based used car price evaluation model includes:

[0028] The historical transaction data acquisition module is used to acquire historical transaction data of used cars and extract basic attribute features, usage attribute features and market trend features from the historical transaction data of used cars.

[0029] The historical transaction data preprocessing module is used to preprocess the basic attribute features, usage attribute features, and market trend features to obtain training samples; the data preprocessing includes missing value imputation, outlier removal, and feature encoding.

[0030] The sample pair construction module is used to select one sample from the training samples as the sample to be evaluated, and to select samples of the same brand, model and year as the sample to be evaluated as real samples from the training 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 transaction price to the official guide price.

[0031] The training module is used to input the real sample and the sample to be evaluated in the sample pair into the first sub-network and the second sub-network of the pre-built dual-tower model, respectively, to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross-network architecture.

[0032] The scalar output module is used to perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and to map the interaction feature vector into a scalar through the output layer;

[0033] The parameter update module is used to update all parameters of the dual-tower model through backpropagation, using the Huber loss between the scalar and the sample price difference of the sample pair as the loss function.

[0034] The loop module is used for repeated training until the preset training termination condition is met, resulting in a trained used car price evaluation model.

[0035] Thirdly, a deep learning-based method for used car price evaluation includes:

[0036] Step 1: Obtain the used car data to be predicted, and extract basic attribute features, usage attribute features, and market trend features from the used car data to be predicted;

[0037] Step 2: Perform data preprocessing on the basic attribute features, usage attribute features, and market condition features to obtain preprocessed used car data to be predicted;

[0038] Step 3: Filter multiple data points from the actual transaction data according to the specified conditions to obtain a reference set;

[0039] Step 4: Input each data point in the reference set into the first sub-network of the used car price assessment model, and input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; the used car price assessment model is trained according to the deep learning-based used car price assessment model training method described above;

[0040] Step 5: Obtain multiple residual differences based on the multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the mean residual difference;

[0041] Step 6: Calculate the mean of the true residual value of the reference set, and calculate the sum of the mean of the true residual value and the mean of the residual value difference to obtain the residual value of the used car to be predicted;

[0042] Step 7: Obtain the predicted transaction price of the used car based on the predicted residual value and the official guide price.

[0043] Optionally, in step 7, when calculating the predicted transaction price of the used car based on the predicted residual value and the official guide price, the calculation formula is as follows:

[0044]

[0045] Among them, P X R represents the transaction price. X P represents the estimated residual value of the used car. guide This indicates the official guide price.

[0046] Fourthly, a deep learning-based used car price evaluation device includes:

[0047] The module for acquiring data to be predicted is used to acquire data on used cars to be predicted and to extract basic attribute features, usage attribute features, and market trend features from the data on used cars to be predicted.

[0048] The data preprocessing module is used to preprocess the basic attribute features, usage attribute features and market trend features to obtain the preprocessed used car data to be predicted.

[0049] The reference set construction module is used to filter multiple data points from real transaction data according to conditions to obtain a reference set.

[0050] The scalar prediction module is used to input each data point in the reference set into the first sub-network of the used car price assessment model, and to input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; the used car price assessment model is trained according to the deep learning-based used car price assessment model training method described above.

[0051] The residual difference mean calculation module is used to obtain multiple residual differences based on multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the residual difference mean.

[0052] The used car residual value calculation module is used to calculate the average true residual value of the reference set, and to calculate the sum of the average true residual value and the average residual value difference to obtain the predicted used car residual value.

[0053] The used car transaction price prediction module is used to obtain the transaction price of the used car to be predicted based on the residual value of the used car to be predicted and the official guide price.

[0054] Compared with the prior art, this application has at least the following beneficial effects:

[0055] 1. This application provides a deep learning-based training method for a used car price evaluation model. The method involves acquiring historical used car transaction data and performing data preprocessing to obtain training samples. Samples are selected from the training samples to construct sample pairs between the sample to be evaluated and the real sample. The sample price difference between the sample pairs is the difference between the residual value of the sample to be evaluated and the residual value of the real sample. The real sample and the sample to be evaluated are input into the first and second sub-networks of a dual-tower model, respectively, to obtain a first sample vector and a second sample vector. Both the first and second sub-networks employ a deep cross-network architecture. The first and second sample vectors are element-wise multiplied to obtain an interaction feature vector, which is then mapped to a scalar through the output layer. The Huber loss between the scalar and the sample price difference of the sample pairs is used as the loss function. Backpropagation is then used to update all parameters of the dual-tower model, resulting in a trained used car price evaluation model. This method employs a deep cross-network architecture as a sub-network of the dual-tower model to explicitly learn high-order feature cross relationships, thereby improving the fitting accuracy of subdivided scenarios. It uses real samples and samples to be evaluated as training labels to guide the dual-tower model to learn the correspondence between sample correlation and price difference, enabling the model to accurately learn the mapping relationship between "sample correlation and price difference".

[0056] 2. This application also provides a deep learning-based method for used car price evaluation. The method involves acquiring used car data to be predicted and preprocessing it to obtain preprocessed used car data; filtering multiple data points from real transaction data according to conditions to obtain a reference set; inputting each data point in the reference set into the first sub-network of the used car price evaluation model, and inputting the preprocessed used car data to be predicted into the second sub-network of the same model to obtain multiple prediction scalars; obtaining multiple residual value differences based on these prediction scalars, and calculating the mean of these residual value differences to obtain the residual value difference mean; calculating the true residual value mean of the reference set, and summing the true residual value mean and the residual value difference mean to obtain the residual value of the used car to be predicted; and obtaining the transaction price of the used car to be predicted based on the residual value and the official guide price. This method, through a fully automated deep learning model, achieves the output of the evaluation price within seconds after the input of used car features, avoiding manual intervention and improving prediction efficiency. Attached Figure Description

[0057] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0058] Figure 1 A flowchart illustrating a deep learning-based used car price evaluation model training method provided in Embodiment 1 of this application;

[0059] Figure 2 This is a schematic diagram of the twin-tower model structure provided in Embodiment 1 of this application;

[0060] Figure 3 This is a flowchart of a deep learning-based used car price evaluation method provided in Embodiment 3 of this application. Detailed Implementation

[0061] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0063] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to indicate the general relative positional relationship for the purpose of intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationship in the actual product.

[0064] Example 1

[0065] Please see Figure 1 This embodiment provides a method for training a deep learning-based used car price evaluation model, including:

[0066] S1: Obtain historical transaction data of used cars, and extract basic attribute features, usage attribute features and market trend features from the historical transaction data of used cars;

[0067] Specifically, the historical transaction data for used cars obtained in this embodiment covers the past three years to ensure the timeliness of market information. After obtaining the historical transaction data for used cars, basic attribute features, usage attribute features, and market information features need to be extracted from it, including:

[0068] Basic attributes include: brand ID, vehicle series ID, model ID, country of origin (China / USA / Germany, etc.), model year (e.g., 2021 model), registration year, sales year, sales month, vehicle type (sedan / SUV / MPV), and whether it is a new energy vehicle (yes / no), etc.

[0069] The attributes used include: mileage (unit: 10,000 kilometers), vehicle age (calculated as: year of sale - year of registration, unit: year), average annual mileage (calculated as: mileage / vehicle age, unit: 10,000 kilometers / year), number of transfers (unit: times), color (black / white / silver, etc.), province (province of vehicle registration), sales cycle (number of days from listing to transaction, unit: days), etc.

[0070] Market characteristics include: residual value (transaction price / guideline price) statistical percentiles (25%, 50%, 75%), mean residual value, and quantity sold for data sold within 7 / 14 / 30 / 90 / 120 days.

[0071] The above three features include calculated derived features and residual statistical quantiles, where:

[0072] The calculation of derived characteristics includes: vehicle age = year of sale - year of registration; average annual mileage = mileage / vehicle age (0 is used when vehicle age = 0); residual value = transaction price / guide price;

[0073] Residual value statistical quantile calculation: Using "vehicle series ID + model year" as the subdivision dimension, and with sliding time windows of 7 days, 14 days, 30 days, 90 days, and 120 days (excluding the sample day), the 25%, 50%, and 75% quantiles, mean, and number of items sold are calculated for each window.

[0074] It should be noted that the residual value statistical quantile is the ratio of the "transaction price / guide price" of a used car (residual value). After sorting from smallest to largest within a specified time window (7 days, 14 days, etc.), the values ​​corresponding to the 25th, 50th, and 75th percentiles are taken, which represent the lower limit, median, and upper limit of the residual value within that window, respectively, to reflect market fluctuations.

[0075] S2: Data preprocessing is performed on basic attribute features, usage attribute features, and market trend features to obtain training samples; data preprocessing includes missing value imputation, outlier removal, and feature encoding.

[0076] Specifically, missing value imputation: for missing features such as "average mileage per year" (when the vehicle age is 0, i.e., a new car that has not been registered, take 0) and "color" (marked as "other" when missing), imputation is performed according to business logic; samples with more than 30% missing values ​​are directly removed.

[0077] Outlier Removal: Obvious outliers such as "mileage" and "residual value" are identified and removed using the IQR (Interquartile Range) method. IQR is a statistical indicator calculated as "the third quartile (Q3, the value at the 75th percentile after sorting) minus the first quartile (Q1, the value at the 25th percentile after sorting)," i.e., IQR = Q3 - Q1. In the data preprocessing stage of this embodiment, outliers (such as mileage and residual value) are typically identified and removed if they exceed the range of "[Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]" to ensure sample quality.

[0078] Feature encoding: For discrete features such as “brand ID” and “model ID”, Embedding encoding is used (dimension set to 8-64, which can be adjusted according to the feature cardinality, such as: if the brand ID cardinality is 200, the Embedding dimension is set to 16); for binary features such as “whether it is a new energy vehicle”, 0-1 encoding is used; for continuous features such as “residual value quantile” and “average annual mileage”, Min-Max normalization is used (mapped to the [0,1] interval).

[0079] It should be noted that embedding encoding is a coding method that maps high-dimensional discrete features (such as brand ID, vehicle model ID, and other classification features with large cardinality) into low-dimensional continuous vectors. In this embodiment, for discrete features such as "brand ID" and "vehicle model ID", the embedding dimension is set according to the feature cardinality (e.g., if the cardinality of brand ID is 200, the embedding dimension is set to 16), transforming discrete values ​​into continuous vectors that can participate in deep learning calculations, thus providing adapted input for feature cross-interaction in the DCN architecture.

[0080] The core of this step is to acquire high-quality used car feature data and complete preprocessing and feature engineering to provide input for model training and prediction. It should be noted that in this embodiment, each training sample contains a set of discrete features and a set of continuous features. The discrete and continuous features are obtained by classifying the aforementioned basic attribute features, usage attribute features, and market condition features.

[0081] S3: Select one sample from the training sample as the sample to be evaluated, and select a sample from the training sample that is the same brand, model and year as the sample to be evaluated as the real sample 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 transaction price to the official guide price.

[0082] Specifically, this step performs the following iterative operation on the preprocessed training samples: Randomly select one sample at a time as the evaluation sample Bi (i=1, 2, 3, ..., n), and filter out all samples from the training samples that are "of the same brand, model, and year" as Bi as real samples, denoted as: Ai1, Ai2, ..., Ai1. i The sample pairs (Bi, Ai1), (Bi, Ai2), ..., (Bi, Ai1) are formed. i After traversing all the samples to be evaluated, a complete set of sample pairs is finally obtained; where the sample price difference (label) of each sample pair is set as the difference between the residual value of the sample to be evaluated and the residual value of the real sample, that is, "B residual value - A residual value".

[0083] It should be noted that the sample price difference (Label) is a sample pair of "used car sample (A)" and "used car sample (B)" constructed during the model training phase. The "residual value of B - residual value of A" (or vice versa) is used as the training label to guide the dual-tower model to learn the mapping relationship between the feature associations between samples and the price differences.

[0084] S4: Input the real sample and the sample to be evaluated in the sample pair into the first sub-network and the second sub-network of the pre-built dual-tower model to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross network architecture;

[0085] Specifically, in this embodiment, the pre-built two-tower model includes a "real sample tower (tower A, i.e., the first sub-network)" and a "sample tower to be evaluated (tower B, i.e., the second sub-network)". Each tower adopts a deep cross network architecture (i.e., DCN architecture) and is trained with "sample price difference" as the label. It should be noted that the two-tower model is a deep learning architecture containing two sub-networks ("towers") with identical (or similar) structures. In this embodiment, feature extraction and vector mapping are performed on the two types of input samples to capture the feature correlation and price difference between the two samples. The core is to achieve adaptive learning of the correlation between samples. DCN (Deep & Cross Network) is a deep learning architecture for processing high-dimensional sparse features. It explicitly learns high-order interaction relationships between features through "cross layers" and learns complex nonlinear features through "deep layers," making it suitable for scenarios requiring precise feature cross-referencing, such as price evaluation.

[0086] Please see Figure 2 In this embodiment, tower A and tower B are completely symmetrical (they run in parallel), each containing six modules: "input layer → embedding layer → feature concatenation layer → (cross layer + depth layer) → feature fusion layer". The output vector dimension is set to 64 (balancing accuracy and computational efficiency). More specifically:

[0087] a. Input layer: Used to receive two types of preprocessed raw features: ① discrete features (brand ID, vehicle model ID, etc.) and ② continuous features (mileage, vehicle age, residual value quantiles, etc., 0-1 encoding is also assigned here).

[0088] b. Embedding layer: used to encode discrete features, setting the dimension according to the feature cardinality (e.g., brand ID cardinality 200 → 16-dimensional vector, car model ID cardinality 45000 → 64-dimensional vector).

[0089] c. Feature concatenation layer: The "embedded discrete feature vector" is concatenated with the "Min-Max normalized continuous feature vector" to form a unified input feature;

[0090] d. Cross Layer (Explicit Feature Cross): Using the features output from the feature concatenation layer as input, and applying the formula... (in, Indicates the initial splicing features. Represents the cross features of the l-th layer. (Represents trainable parameters) to learn higher-order cross relationships. This layer is set to 3 layers and is used to output explicit cross features;

[0091] e. Deep layer (implicit feature cross): Also using the features output by the feature concatenation layer as input, it adopts 3 fully connected layers (activation function: ReLU, number of neurons: 128→64→32) to learn complex nonlinear relationships and output 32-dimensional implicit cross features;

[0092] f. Feature Fusion Layer: The 32-dimensional explicit features output from the cross layer are concatenated with the 32-dimensional implicit features output from the deep layer to form a 64-dimensional fused feature. Then, a fully connected layer is used to perform dimensionality mapping on the 64-dimensional fused feature (preserving 64 dimensions), outputting the sample vector representation of the tower (i.e., tower A outputs the first sample vector V). A Tower B outputs the second sample vector V. B ).

[0093] This step inputs the "sample to be evaluated (B)" from the sample pair into tower B (i.e., the second sub-network) to obtain the second sample vector V. B The first sample vector V is obtained by inputting the "real sample (A)" into tower A (i.e., the first sub-network). A .

[0094] S5: Perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and then map the interaction feature vector to a scalar through the output layer;

[0095] For details, please continue reading. Figure 2 This step utilizes the feature interaction layer to process the first sample vector V output by tower A. A The second sample vector V output by tower B B To perform element-wise multiplication, that is: This outputs a 64-dimensional interaction vector, which is then input into the output layer (a single-neuron fully connected layer) to output a scalar z.

[0096] S6: Use the Huber loss between the sample price difference of the scalar and the sample pair as the loss function, and update all parameters of the dual-tower model through backpropagation;

[0097] In this embodiment, when training the model, the optimizer is Adam, the learning rate is set to 0.005, the batch size is 256, the training epochs are set to 50, and the loss function is Huber Loss, i.e.:

[0098]

[0099] Among them, Zi The model outputs a scalar, the label. i For the sample pair residual difference, is a hyperparameter (valued at 1.0, balancing the sensitivity of MSE to outliers with the convergence speed of MAE).

[0100] S7: Repeat steps S3 to S6 until the preset training termination condition is met, and obtain the trained used car price evaluation model.

[0101] This embodiment provides a deep learning-based training method for used car price evaluation models. It employs a DCN architecture as a sub-network of a dual-tower model, explicitly learning high-order feature cross-relationships such as "vehicle series - province - residual value quantile," improving the fitting accuracy for subdivided scenarios and solving the problem of "insufficient high-order feature cross-processing capability." Furthermore, this embodiment uses "sample A and sample B" as training labels to guide the dual-tower model in learning the correspondence between sample correlation and price differences. (Note: In the actual model, to ensure convergence, the price is processed into residual value, where residual value = price / guide price, which is essentially no different from the price. Therefore, the entire text still uses price as the term for ease of explanation.) This solves the problem of "inaccurate mapping between sample correlation and price differences."

[0102] In summary, the deep learning-based used car price evaluation model training method provided in this embodiment has the following advantages:

[0103] 1. Improved feature cross-cutting accuracy: By using the "DCN architecture", the problem of discrete feature processing is solved and higher-order feature cross-cutting is explicitly learned, thereby reducing prediction error.

[0104] 2. Improved prediction accuracy: The dual-tower model, which uses "sample price difference" as the label and Huber Loss as the loss, combined with the prediction logic of "multiple reference sample means", reduces the overall evaluation error.

[0105] 3. Strong market adaptability: The introduction of the "7-120 day residual value statistical quantile" feature can capture market fluctuations in real time; at the same time, the embedding encoding has strong compatibility with new brands / models, and the evaluation scope can be expanded without reconstructing the model, making it more adaptable than traditional models.

[0106] 4. Improved objectivity and efficiency: The fully automated process requires no human intervention, reducing the evaluation time for a single vehicle from minutes to less than seconds; at the same time, outliers (such as abnormally high mileage samples) are removed through the IQR method, reducing the standard deviation of sample data and the standard deviation of evaluation results for the same vehicle.

[0107] Example 2

[0108] This embodiment provides a training device for a deep learning-based used car price evaluation model, including:

[0109] The historical transaction data acquisition module is used to acquire historical transaction data of used cars and extract basic attribute features, usage attribute features and market trend features from the historical transaction data of used cars.

[0110] The historical transaction data preprocessing module is used to preprocess the basic attribute features, usage attribute features, and market trend features to obtain training samples; the data preprocessing includes missing value imputation, outlier removal, and feature encoding.

[0111] The sample pair construction module is used to select one sample from the training samples as the sample to be evaluated, and to select samples of the same brand, model and year as the sample to be evaluated as real samples from the training 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 transaction price to the official guide price.

[0112] The training module is used to input the real sample and the sample to be evaluated in the sample pair into the first sub-network and the second sub-network of the pre-built dual-tower model, respectively, to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross-network architecture.

[0113] The scalar output module is used to perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and to map the interaction feature vector into a scalar through the output layer;

[0114] The parameter update module is used to update all parameters of the dual-tower model through backpropagation, using the Huber loss between the scalar and the sample price difference of the sample pair as the loss function.

[0115] The loop module is used for repeated training until the preset training termination condition is met, resulting in a trained used car price evaluation model.

[0116] For details on the specific implementation of each module in the training device for a deep learning-based used car price evaluation model, please refer to the above description of the limitations of the training method for a deep learning-based used car price evaluation model, which will not be repeated here.

[0117] Example 3

[0118] Please see Figure 3 This embodiment provides a deep learning-based method for evaluating the price of used cars, including:

[0119] S1: Obtain the used car data to be predicted, and extract basic attribute features, usage attribute features and market trend features from the used car data to be predicted;

[0120] S2: Perform data preprocessing on basic attribute features, usage attribute features, and market condition features to obtain preprocessed used car data to be predicted;

[0121] Specifically, the preprocessing process for predicting used car data X is the same as the model training process, and will not be repeated here.

[0122] S3: Filter multiple data points from real transaction data according to conditions to obtain a reference set;

[0123] Specifically, this step filters data from real transaction data based on certain conditions (e.g., data with the same brand, model, age difference ≤ 1 year, and mileage difference ≤ 10,000 km as X, assuming an official guide price P). guide =349,900 yuan), forming a reference set S (containing at least 5 data points, denoted as S1, S2, ..., S5, with their actual residual values ​​being R1=0.85, R2=0.82, R3=0.88, R4=0.83, and R5=0.81 respectively).

[0124] S4: Input each data point in the reference set into the first sub-network of the used car price assessment model, and input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; wherein, the used car price assessment model is trained according to the deep learning-based used car price assessment model training method provided in Example 1;

[0125] Specifically, this step inputs the used car data X to be predicted into tower B; and inputs each data S in the reference set S. i Input tower A, model output prediction scalar z i .

[0126] S5: Obtain multiple residual differences based on multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the mean residual difference;

[0127] Specifically, this step involves predicting the scalar z. i The corresponding residual value difference can be obtained directly. (For example: , , , , ).

[0128] Then, by calculating the average of multiple residual differences, the mean of the residual difference is obtained, i.e.:

[0129] .

[0130] S6: Calculate the mean of the true residual value of the reference set, and calculate the sum of the mean of the true residual value and the mean of the residual value difference to obtain the residual value of the used car to be predicted;

[0131] Specifically, this step requires first calculating the mean of the true residuals of the reference set, that is:

[0132] ;

[0133] Then, the sum of the mean of the actual residual value and the mean of the residual value difference is calculated to obtain the predicted residual value of the used car, i.e.:

[0134] .

[0135] S7: The predicted transaction price of the used car is obtained based on the predicted residual value and the official guide price.

[0136] Specifically, in this step, when calculating the predicted transaction price of the used car based on its residual value and the official guide price, the formula is as follows:

[0137] ;

[0138] Among them, P X R represents the transaction price. X P represents the estimated residual value of the used car. guide This indicates the official guide price;

[0139] For example:

[0140] Ten thousand yuan.

[0141] This embodiment provides a deep learning-based used car price evaluation method that can use a trained dual-tower model, combined with multiple sets of real transaction samples, to calculate the final price of the used car to be evaluated.

[0142] This embodiment provides a deep learning-based used car price evaluation method that uses a fully automated deep learning model to output the evaluated price within seconds after the input of used car features, avoiding manual intervention, improving prediction efficiency, and solving the problems of "low efficiency and strong subjectivity of manual evaluation"; and by using the calculation method of "mean of price difference of multiple real samples + mean of real price", it reduces the error of a single sample.

[0143] Example 4

[0144] This embodiment provides a deep learning-based used car price evaluation device, including:

[0145] The module for acquiring data to be predicted is used to acquire data on used cars to be predicted and to extract basic attribute features, usage attribute features, and market trend features from the data on used cars to be predicted.

[0146] The data preprocessing module is used to preprocess the basic attribute features, usage attribute features and market trend features to obtain the preprocessed used car data to be predicted.

[0147] The reference set construction module is used to filter multiple data points from real transaction data according to conditions to obtain a reference set.

[0148] The scalar prediction module is used to input each data point in the reference set into the first sub-network of the used car price assessment model, and to input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; the used car price assessment model is trained according to a deep learning-based used car price assessment model training method provided in Embodiment 1;

[0149] The residual difference mean calculation module is used to obtain multiple residual differences based on multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the residual difference mean.

[0150] The used car residual value calculation module is used to calculate the average true residual value of the reference set, and to calculate the sum of the average true residual value and the average residual value difference to obtain the predicted used car residual value.

[0151] The used car transaction price prediction module is used to obtain the transaction price of the used car to be predicted based on the residual value of the used car to be predicted and the official guide price.

[0152] For details on the specific implementation of each module in a deep learning-based used car price evaluation device, please refer to the above description of the limitations of a deep learning-based used car price evaluation method, which will not be repeated here.

[0153] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A training method for a used car price evaluation model based on deep learning, characterized in that, include: Step 1: Obtain historical transaction data for used cars, and extract basic attribute features, usage attribute features, and market trend features from the historical transaction data for used cars; Step 2: Perform data preprocessing on the basic attribute features, usage attribute features, and market trend features to obtain training samples; the data preprocessing includes missing value imputation, outlier removal, and feature encoding; Step 3: Select one sample from the training samples as the sample to be evaluated, and select a sample from the training samples that is the same brand, model, and year as the sample to be evaluated as the real sample 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 transaction price to the official guide price. Step 4: Input the real sample and the sample to be evaluated from the sample pair into the first sub-network and the second sub-network of the pre-constructed dual-tower model, respectively, to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross-network architecture; Step 5: Perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and map the interaction feature vector to a scalar through the output layer; Step 6: Using the Huber loss between the scalar and the sample price difference of the sample pair as the loss function, update all parameters of the dual-tower model through backpropagation; Step 7: Repeat steps 3 to 6 until the preset training termination condition is met, and obtain the trained used car price evaluation model.

2. The training method for a deep learning-based used car price evaluation model according to claim 1, characterized in that, In step 2, the feature encoding specifically includes: embedding and encoding the discrete features among the basic attribute features, usage attribute features, and market trend features; encoding the binary features among the basic attribute features, usage attribute features, and market trend features using 0-1 encoding; and normalizing the continuous features among the basic attribute features, usage attribute features, and market trend features.

3. The training method for a deep learning-based used car price evaluation model according to claim 1, characterized in that, In step 2, outliers are identified and removed using the interquartile range method.

4. The training method for a deep learning-based used car price evaluation model according to claim 1, characterized in that, In step 4, the structures of the first sub-network and the second sub-network both include: an input layer, an embedding layer, a feature concatenation layer, a cross layer, a deep layer, and a feature fusion layer.

5. The training method for a deep learning-based used car price evaluation model according to claim 4, characterized in that, The cross layer has three layers, and the depth layer uses three fully connected layers.

6. The training method for a deep learning-based used car price evaluation model according to claim 1, characterized in that, In step 6, the loss function is: ； Among them, Z i For scalars, label i For the sample pair residual difference, This is a hyperparameter.

7. A training device for a deep learning-based used car price evaluation model, characterized in that, include: The historical transaction data acquisition module is used to acquire historical transaction data of used cars and extract basic attribute features, usage attribute features and market trend features from the historical transaction data of used cars. The historical transaction data preprocessing module is used to preprocess the basic attribute features, usage attribute features, and market trend features to obtain training samples; the data preprocessing includes missing value imputation, outlier removal, and feature encoding. The sample pair construction module is used to select one sample from the training samples as the sample to be evaluated, and to select samples of the same brand, model and year as the sample to be evaluated as real samples from the training 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 transaction price to the official guide price. The training module is used to input the real sample and the sample to be evaluated in the sample pair into the first sub-network and the second sub-network of the pre-built dual-tower model, respectively, to obtain the first sample vector and the second sample vector; both the first sub-network and the second sub-network adopt a deep cross-network architecture. The scalar output module is used to perform element-wise multiplication of the first sample vector and the second sample vector to obtain the interaction feature vector, and to map the interaction feature vector into a scalar through the output layer; The parameter update module is used to update all parameters of the dual-tower model through backpropagation, using the Huber loss between the scalar and the sample price difference of the sample pair as the loss function. The loop module is used for repeated training until the preset training termination condition is met, resulting in a trained used car price evaluation model.

8. A deep learning-based method for evaluating the price of used cars, characterized in that, include: Step 1: Obtain the used car data to be predicted, and extract basic attribute features, usage attribute features, and market trend features from the used car data to be predicted; Step 2: Perform data preprocessing on the basic attribute features, usage attribute features, and market condition features to obtain preprocessed used car data to be predicted; Step 3: Filter multiple data points from the actual transaction data according to the specified conditions to obtain a reference set; Step 4: Input each data point in the reference set into the first sub-network of the used car price assessment model, and input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; the used car price assessment model is trained by the deep learning-based used car price assessment model training method according to any one of claims 1 to 6; Step 5: Obtain multiple residual differences based on the multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the mean residual difference; Step 6: Calculate the mean of the true residual value of the reference set, and calculate the sum of the mean of the true residual value and the mean of the residual value difference to obtain the residual value of the used car to be predicted; Step 7: Obtain the predicted transaction price of the used car based on the predicted residual value and the official guide price.

9. The deep learning-based used car price evaluation method according to claim 8, characterized in that, In step 7, when calculating the predicted transaction price of the used car based on the predicted residual value and the official guide price, the formula is as follows: ； Among them, P X R represents the transaction price. X P represents the estimated residual value of the used car. guide This indicates the official guide price.

10. A used car price evaluation device based on deep learning, characterized in that, include: The module for acquiring data to be predicted is used to acquire data on used cars to be predicted and to extract basic attribute features, usage attribute features, and market trend features from the data on used cars to be predicted. The data preprocessing module is used to preprocess the basic attribute features, usage attribute features and market trend features to obtain the preprocessed used car data to be predicted. The reference set construction module is used to filter multiple data points from real transaction data according to conditions to obtain a reference set. The scalar prediction module is used to input each data point in the reference set into the first sub-network of the used car price assessment model, and to input the preprocessed used car data to be predicted into the second sub-network of the used car price assessment model to obtain multiple prediction scalars; the used car price assessment model is trained by the deep learning-based used car price assessment model training method according to any one of claims 1 to 6. The residual difference mean calculation module is used to obtain multiple residual differences based on multiple predicted scalars, and calculate the mean of the multiple residual differences to obtain the residual difference mean. The used car residual value calculation module is used to calculate the average true residual value of the reference set, and to calculate the sum of the average true residual value and the average residual value difference to obtain the predicted used car residual value. The used car transaction price prediction module is used to obtain the transaction price of the used car to be predicted based on the residual value of the used car to be predicted and the official guide price.

Images