A vehicle finance credit method

Abstract

The application provides a vehicle financial credit method, relates to the technical field of fleet credit data processing, and comprises the following steps: in response to a data deletion request, determining the data fingerprint of a vehicle to be deleted in a target fleet; determining first retained data based on a data fingerprint list and initial data of the target fleet; determining the influence level of data deletion on a credit model based on the data fingerprint list and by using a simulation algorithm; determining the total matrix of fleet batch contribution degrees of the vehicle to be deleted to the credit model based on the multi-dimensional features of each vehicle of the target fleet and the dimensional features of the vehicle to be deleted; performing cooperative counter-training on the credit model based on the total matrix of fleet batch contribution degrees to generate a counter-training credit model; and inputting the first retained data of the target fleet into the counter-training credit model to generate a first credit result. While completely removing target data, the application retains the common features of retained vehicles of the fleet, and can maintain the discriminant stability of the credit model under local data loss.

Description

Technical Field

[0001] This invention relates to the field of fleet credit data processing technology, and in particular to a method for vehicle financial credit granting. Background Technology

[0002] With the rapid development of the new energy commercial vehicle industry, big data-based intelligent credit granting models have become a core tool for financial institutions to assess fleet operational risks and determine credit limits. These models typically rely on long-term accumulated fleet operation data (such as driving trajectories, charging and battery swapping records, and freight documents) for training to capture the fleet's operational characteristics and risk preferences. However, with increasingly stringent global data privacy regulations (such as GDPR and related domestic laws), the "right to be forgotten" held by data subjects, including fleets and enterprises, has led to a surge in requests for bulk revocation of authorizations or deletion of data.

[0003] In existing technologies, the common practice is to simply retrain the model after physically deleting the original data. This method is not only computationally expensive and time-consuming, making it difficult to meet the real-time requirements of business operations, but it can also lead to "catastrophic forgetting" when faced with large-scale data loss at the fleet level. That is, while the model forgets specific data during retraining, it also loses its general judgment ability for such scenarios. Secondly, there is a serious conflict between "feature entanglement" and "data dependency" in fleet credit granting scenarios. Vehicles within the same fleet often have highly consistent operating patterns (e.g., the same fixed routes, uniform refueling times), and these common characteristics are used by the model as an important basis for determining credit limits. When a fleet requests the batch deletion of some vehicle data, conventional machine forgetting algorithms often cannot distinguish between "unique features of the vehicles to be deleted" and "common features shared by the fleet," causing the model to mistakenly eliminate the weights of the common features of the remaining vehicles when removing the weights of the deleted data. Such accidental deletion can directly cause a sharp drop in the model's score for compliant vehicles in the fleet that were not requested for deletion, leading to abnormal fluctuations in credit limits and thus disrupting the continuity and fairness of financial services. Furthermore, in actual operation, there is also the issue of risk control blind spots caused by "data sparsity." For fleets located in specific regions or with specific vehicle types, industry samples are already scarce. Once a mass data deletion event occurs, the reference data for that dimension may instantly become zero, creating an information vacuum. Existing technical means lack effective data filling and risk hedging mechanisms, resulting in the system being unable to find a reasonable alternative reference after data deletion. This not only makes it impossible to maintain the original credit strategy but may even force a complete interruption of financial services to the fleet due to insufficient data. Summary of the Invention

[0004] This invention aims to at least solve the aforementioned technical problems existing in the prior art. To this end, this invention proposes a vehicle financial credit granting method, the method comprising:

[0005] In response to a data deletion request, the data fingerprints of the vehicles to be deleted in the target fleet are determined based on a preset association graph, and a data fingerprint list is generated; the data fingerprint is a fixed-length string generated by salting and hashing the original data of the vehicles to be deleted.

[0006] Based on the data fingerprint list and the initial data of the target fleet, the first retained data of the target fleet is determined;

[0007] Based on the data fingerprint list, a simulation algorithm is used to determine the impact level of data deletion on the credit model;

[0008] Based on the multi-dimensional characteristics of each vehicle in the target fleet and the multi-dimensional characteristics of the vehicles to be deleted, the total fleet batch contribution matrix of the vehicles to be deleted to the credit model is determined.

[0009] Based on the total contribution matrix of the fleet batch, collaborative anti-training is performed on the credit granting model to generate an anti-trained credit granting model;

[0010] The first retained data of the target fleet is input into the anti-training trust model to generate the first trust result; the first retained data is the original feature data of the vehicles retained by the target fleet.

[0011] Optionally, after inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate the first trust result, the method further includes:

[0012] Based on the impact level, determine whether to adjust the credit limit in the first credit granting result to determine the second credit granting result.

[0013] Optionally, determining the impact level of data deletion on the credit model based on the data fingerprint list and using a simulation algorithm includes:

[0014] Based on the data fingerprint list, and using a probability distribution simulation algorithm to simulate a data deletion scenario, the change in the credit model prediction error rate and the fluctuation in the credit limit before and after data deletion are determined.

[0015] The impact of data deletion on the credit granting model is determined based on the change in the prediction error rate of the credit granting model, the fluctuation of the credit limit, and the preset impact level; the impact level includes mild, moderate, and severe.

[0016] Optionally, the step of performing collaborative inverse training on the credit granting model based on the fleet batch contribution matrix includes:

[0017] The overall operational profile of the target fleet is corrected based on the fleet batch contribution matrix to obtain the first profile of the target fleet; the overall operational profile is a vector formed by aggregating the features of all vehicles in the target fleet.

[0018] Adjust the first profile until the deviation between the adjusted first profile and the overall business profile generated from the first retained data is less than a preset deviation threshold, and then generate the second profile of the target fleet.

[0019] Determine the second retained data corresponding to the second profile; the second retained data is the corrected retained data corresponding to the first retained data.

[0020] Optionally, determining the fleet batch contribution matrix of the vehicles to be deleted to the credit model based on the multi-dimensional features of each vehicle in the target fleet and the multi-dimensional features of the vehicles to be deleted includes:

[0021] Determine the multi-dimensional characteristics of each vehicle in the target fleet based on the initial data of the target fleet;

[0022] The vehicle to be deleted is selected from the multi-dimensional features of each vehicle in the target fleet.

[0023] Determine the characteristics of all vehicles in the target fleet in each dimension;

[0024] Based on the features of the vehicle to be deleted in each dimension and the sum of features in the corresponding dimension, the feature ratio of the vehicle to be deleted in each dimension is obtained, and the feature ratio is used as the single vehicle feature contribution of the vehicle to be deleted in each dimension.

[0025] The contribution of each vehicle's features across all dimensions is arranged in a matrix to obtain the total contribution matrix of the fleet.

[0026] Optionally, the step of performing collaborative detraining on the credit granting model based on the fleet batch contribution total matrix to generate a detrained credit granting model further includes:

[0027] A preset number of samples are randomly selected from the first retention data as a retention anchor set. The loss function of the retention anchor set under the credit model is determined. The average gradient is obtained based on the loss function and model parameters. The average gradient is used as the retention gradient vector.

[0028] Determine the gradient vector of the vehicle data to be deleted under the credit model, and use the gradient vector as the original deletion gradient vector;

[0029] Determine the projection component of the original deletion gradient vector in the direction of the retained gradient vector, and use the projection component as the feature entanglement gradient;

[0030] A purified inverse gradient is generated based on the original deletion gradient vector and the feature entanglement gradient.

[0031] Based on the purified reverse gradient, aggregated reverse gradient descent is performed, and combined with regularization constraints, to generate the reverse training confidence model.

[0032] Optionally, before determining the data fingerprint of the vehicle to be deleted in the target fleet, the method further includes:

[0033] The percentage of vehicles requested for deletion is determined based on the number of vehicles currently requested for deletion in the target fleet and the total number of vehicles in the target fleet.

[0034] If the proportion of the number of vehicles requested for deletion is greater than or equal to the preset batch deletion trigger threshold, or if the data deletion request is a full fleet data deletion request for the target fleet, a fleet batch deletion event identifier is generated.

[0035] Optionally, the step of inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate a first trust result includes:

[0036] The overall business profile generated based on the first retained data and the first profile are used to determine the second profile, and the corresponding second retained data are determined based on the second profile.

[0037] Build industry average data;

[0038] Extract the target feature dimensions from the second retained data that correspond to the industry average data;

[0039] The industry average data and the second retained data are weighted and fused along the target feature dimension, while the data of the second retained data in the non-target feature dimension are retained to obtain fused data;

[0040] The fused data is input into the anti-training trust model to obtain the first trust result.

[0041] Optionally, the construction of industry average data includes:

[0042] Extract data units that match the dimensions of the target fleet from a pre-defined multidimensional data pool;

[0043] Extract the average operating mileage, revenue fluctuation range, loan repayment rate, and vehicle failure rate from the data unit;

[0044] Determine whether the number of valid samples within the data unit is less than the statistical confidence threshold;

[0045] If so, based on a pre-constructed regional economic similarity matrix, a set of associated regions is identified; the set of associated regions is a set of regions whose similarity to the region where the target fleet is located is higher than a preset similarity threshold.

[0046] Extract alternative data from the associated region set that are consistent with the vehicle type and fleet size dimensions of the target fleet; the alternative data includes average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate;

[0047] The regional correction coefficient is determined based on the logistics activity index of the first region where the target fleet is located and the logistics activity index of the second region of the associated region set.

[0048] The substitute data is corrected based on the regional correction coefficient;

[0049] The revised alternative data will be used as the industry average data.

[0050] Otherwise, the average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate in the data unit will be used as the industry average data.

[0051] Optionally, the logistics activity index of the first region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of the region where the target fleet is located; the logistics activity index of the second region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of each associated region in the associated region set.

[0052] The technical solution provided in this application may include the following beneficial effects:

[0053] By employing a collaborative inverse training method based on the contribution matrix, the feature weights of data to be deleted in the model can be accurately located and removed, significantly reducing computational costs. Specifically, a gradient orthogonal purification mechanism is introduced during the inverse training process, decomposing the gradients of the data to be deleted into unique gradients and common gradients, and updating only the unique gradients. This mechanism ensures that while thoroughly removing the target data, the common feature structure of the remaining vehicles in the fleet is well preserved, avoiding the misallocation of credit scores for remaining vehicles due to "feature entanglement," and maintaining the discrimination stability of the credit granting model under local data loss. Attached Figure Description

[0054] Figure 1 This is a flowchart of a vehicle financial credit granting method provided in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0057] This invention provides a vehicle financial credit granting method, such as... Figure 1 As shown, the method may include the following steps:

[0058] Step 101: In response to the data deletion request, determine the data fingerprints of the vehicles to be deleted in the target fleet based on the preset association graph, and generate a data fingerprint list; the data fingerprint is a fixed-length string generated by salting and hashing the original data of the vehicles to be deleted.

[0059] Specifically, the system identifies the data fingerprints of vehicles to be deleted from a target fleet based on a pre-defined multi-level dynamic association graph. This graph's topology includes four core levels of nodes: fleet nodes, vehicle nodes, stop point nodes, and associated entity nodes. The fleet nodes store macro-level attributes of the fleet, such as business registration information, overall credit line, and other information. Vehicle nodes represent specific physical vehicles, storing their VIN codes, vehicle models, battery parameters, and other information. Stop point nodes map the high-frequency activity locations of vehicles in physical space, such as specific logistics parks, charging stations, or unloading points. Associated entity nodes represent third parties that have economic dealings with the fleet or vehicles, such as cargo owners, repair shops, or insurance companies.

[0060] In response to a data deletion request, the system locates the affiliation of the target fleet through fleet nodes and uses the edge attributes between fleet nodes and vehicle nodes to determine all vehicles under the target fleet to be deleted; where the edge attributes include at least the association strength between the fleet and the vehicle. The system then iterates through the vehicle operation data, business transaction data, and historical credit data of the vehicles to be deleted, extracting a unique data fingerprint for each data entry to generate a data fingerprint list.

[0061] For example, vehicles to be removed from a target fleet can be identified in the following way:

[0062] First, the affiliation of the target fleet is located by identifying the fleet node, thus pinpointing the unique position of the fleet node in the graph. Then, the edge attributes between the fleet node and vehicle nodes are used for traversal. In this embodiment, the edge attributes include at least the association strength between the fleet and the vehicles. This association strength is a weighted value that comprehensively reflects the stability of the vehicle affiliation, the frequency of operational data uploads, and historical repayment contributions. Following these edge attributes, all vehicles belonging to the target fleet that need to be deleted can be identified. This includes not only the vehicles that directly initiated the request, but also all associated vehicles registered under that fleet name in the event of a fleet-wide deletion command.

[0063] The association strength between the fleet and vehicles can be obtained through a commonly used weighted summation method. Specifically, the vehicle affiliation stability factor, operational data upload frequency factor, and historical repayment contribution factor are extracted, normalized, and then summed according to preset weights. The expression is as follows:

[0064]

[0065] in, Indicates the strength of the association between the fleet and the vehicles. Indicates the stability factor of the attribution relationship. This indicates the frequency factor for uploading operational data. Indicates the historical repayment contribution factor. , , These are the preset weighting coefficients for the three factors mentioned above.

[0066] After identifying all vehicles slated for deletion, the system iterates through all dimensions of their raw data stored in the database, including vehicle operational data (e.g., daily mileage, energy consumption), business transaction data (e.g., waybill logs, charging payment records), and historical credit data (e.g., past repayment records, overdue payments). To ensure data security and privacy, plaintext data is not extracted directly; instead, a unique data fingerprint is extracted for each data entry. This data fingerprint is a fixed-length string generated by salting and hashing the raw data. Finally, all relevant data fingerprints are aggregated to generate a data fingerprint list. This list will serve as the sole index for all subsequent deletion, verification, and detraining operations.

[0067] In one possible implementation, prior to determining the data fingerprint of the vehicle to be deleted from the target fleet, the method further includes:

[0068] The percentage of vehicles requested for deletion is determined based on the number of vehicles currently requested for deletion in the target fleet and the total number of vehicles in the target fleet.

[0069] If the proportion of the number of vehicles requested for deletion is greater than or equal to the preset batch deletion trigger threshold, or if the data deletion request is a full fleet data deletion request for the target fleet, a fleet batch deletion event identifier is generated.

[0070] Specifically, upon receiving a deletion command, the system first determines whether the triggering conditions for batch deletion are met based on preset logic. It monitors in real-time the percentage of vehicles within the same fleet that initiate data deletion requests. This percentage is calculated by dividing the total number of vehicles requesting deletion by the total number of all registered vehicles with valid credit limits under that fleet. If the percentage is greater than or equal to the preset batch deletion trigger threshold, or if a fleet-wide data deletion command is received from the target fleet entity via API, the event is determined to be a fleet batch deletion event. Furthermore, to facilitate subsequent end-to-end tracking and auditing, a globally unique fleet batch deletion event identifier is generated, triggering the step of determining the data fingerprint or data fingerprint list of vehicles to be deleted. The target fleet entity is the account with the highest management authority in the fleet; the fleet batch deletion event identifier is generated by a hash algorithm using a timestamp, fleet code, and random salt value.

[0071] For example, the batch deletion trigger threshold can be set to 30%. That is, once more than 30% of the vehicle data in the fleet faces removal, it is considered that this will have a significant structural impact on the existing credit model, thereby triggering subsequent deep processing procedures.

[0072] After confirming the triggering of a batch deletion event, an identification program based on a preset multi-level dynamic association graph is started to accurately locate the data fingerprints of the vehicles to be deleted in the target fleet, thereby determining the data fingerprint list.

[0073] Step 102: Based on the data fingerprint list and the initial data of the target fleet, determine the first retained data of the target fleet.

[0074] Specifically, retained data refers to the relevant data of the remaining vehicles in the target fleet.

[0075] Step 103: Based on the data fingerprint list, use a simulation algorithm to determine the impact level of data deletion on the credit model.

[0076] In one possible implementation, determining the impact level of data deletion on the credit model based on the data fingerprint list and using a simulation algorithm includes:

[0077] Based on the data fingerprint list, and using a probability distribution simulation algorithm to simulate a data deletion scenario, the change in the credit model prediction error rate and the fluctuation in the credit limit before and after data deletion are determined.

[0078] The impact of data deletion on the credit granting model is determined based on the change in the prediction error rate of the credit granting model, the fluctuation of the credit limit, and the preset impact level; the impact level includes mild, moderate, and severe.

[0079] After obtaining the data fingerprint list, to scientifically assess the consequences of the deletion operation, this embodiment of the invention does not perform physical deletion first, but instead conducts a virtual impact assessment. To achieve accurate quantification, the simulation algorithm in this embodiment employs the Monte Carlo simulation method. Specifically, a sandbox environment is constructed in memory, and a copy of the current trusted model is loaded. Then, based on the data fingerprint list, the corresponding feature inputs are masked in the sandbox environment, simulating the model inference process when these data are absent. This simulation process is repeated multiple times, introducing small random perturbations each time to test the model's robustness. Through the above simulation, two key quantitative indicators can be calculated: the change in the model prediction error rate before and after deletion. and credit limit fluctuation value These two metrics are quantifications of the difference between the model output before and after the simulated deletion, specifically calculated through the following comparative simulation process:

[0080] Before the simulated deletion, using complete fleet history data as input, the mean absolute percentage error of the current credit model on the validation set is denoted as . After masking the data of the vehicles to be deleted at the input layer based on the data fingerprint list (simulating deletion), the credit model is run again, and the new prediction error is denoted as . The expression for the change in the model prediction error rate is as follows:

[0081]

[0082] Extract the original credit limit prediction value output by the current credit granting model for the target fleet. After performing the simulated deletion operation, the predicted new credit limit output by the model is extracted and recalculated. The credit limit fluctuation value is the absolute difference or relative fluctuation range between the two, expressed as follows:

[0083]

[0084] The change in model prediction error rate reflects the degree to which the model's prediction accuracy for the remaining samples (or validation set) decreases after this portion of data is lost. The credit limit fluctuation reflects the degree to which the total credit limit for the fleet calculated by the model deviates from the original limit under the new data environment.

[0085] Based on these two indicators, the impact level of the deletion event on the current credit granting model is determined. This embodiment of the invention presets three impact levels: mild, moderate, and severe. The determination logic is as follows:

[0086] If the change in the model prediction error rate or the fluctuation in the credit limit is within the first preset range, the impact level is determined to be mild. This means that although the deletion operation will cause some data loss, the basic performance of the model remains stable, and business can continue normally.

[0087] If the above indicators fall within the second preset range, the impact level is determined to be moderate. This indicates a significant decline in model performance, requiring specific compensation measures to correct it.

[0088] If the above indicators exceed the upper limit of the second preset range, the impact level is determined to be severe. This usually means that a catastrophic loss of information has occurred, and the model is no longer able to maintain normal service through self-correction, requiring a high-level manual review or circuit breaker mechanism to be triggered.

[0089] Unlike existing technologies that only focus on physical deletion or simple retraining, this invention first establishes a quantitative impact assessment system. Through matrix aggregation algorithms and probabilistic simulations, it can accurately predict the specific impact level of batch deletion on model errors and credit limits before the deletion action is executed. This proactive risk quantification method transforms data deletion from a simple IT operation into a controllable financial risk control process, providing a basis for subsequent tiered processing strategies.

[0090] Step 104: Determine the total contribution matrix of the fleet batch of vehicles to be deleted to the credit model based on the multi-dimensional features of each vehicle in the target fleet and the multi-dimensional features of the vehicles to be deleted.

[0091] In one possible implementation, determining the fleet batch contribution matrix of the vehicles to be deleted to the credit model based on the multi-dimensional features of each vehicle in the target fleet and the multi-dimensional features of the vehicles to be deleted includes:

[0092] Determine the multi-dimensional characteristics of each vehicle in the target fleet based on the initial data of the target fleet;

[0093] The vehicle to be deleted is selected from the multi-dimensional features of each vehicle in the target fleet.

[0094] Determine the characteristics of all vehicles in the target fleet in each dimension;

[0095] Based on the features of the vehicle to be deleted in each dimension and the sum of features in the corresponding dimension, the feature ratio of the vehicle to be deleted in each dimension is obtained, and the feature ratio is used as the single vehicle feature contribution of the vehicle to be deleted in each dimension.

[0096] The contribution of each vehicle's features across all dimensions is arranged in a matrix to obtain the total contribution matrix of the fleet.

[0097] The fleet batch contribution matrix is ​​used to quantify the overall impact of these vehicles to be deleted on the current credit model parameters. Specifically, the process for obtaining the fleet batch contribution matrix is ​​as follows:

[0098] Assume that the fleet feature vectors upon which the current credit granting model relies contain Features in each dimension, denoted as The set of vehicles to be deleted includes Vehicle, recorded as Extract each vehicle to be deleted. ( In each feature dimension ( The original indicator data is used to calculate the ratio of the indicator to the total number of corresponding features of the entire target fleet (including all vehicles), thus obtaining the feature contribution of a single vehicle. .Will Vehicles to be deleted The contribution values ​​of individual vehicle features across each feature dimension are arranged in a matrix to construct a... A matrix of dimension is the total contribution matrix of the fleet batch. .

[0099]

[0100] Using this matrix, the model can clearly identify which specific feature dimensions caused the greatest data loss due to the batch deletion behavior.

[0101] Step 105: Perform collaborative anti-training on the credit model based on the total contribution matrix of the fleet batch, and generate an anti-trained credit model.

[0102] In one possible implementation, the step of performing collaborative inverse training on the credit granting model based on the fleet batch contribution matrix includes:

[0103] The overall operational profile of the target fleet is corrected based on the fleet batch contribution matrix to obtain the first profile of the target fleet; the overall operational profile is a vector formed by aggregating the features of all vehicles in the target fleet.

[0104] Adjust the first profile until the deviation between the adjusted first profile and the overall business profile generated from the first retained data is less than a preset deviation threshold, and then generate the second profile of the target fleet.

[0105] Determine the second retained data corresponding to the second profile; the second retained data is the corrected retained data corresponding to the first retained data.

[0106] Specifically, the collaborative anti-training in this embodiment of the invention includes operations at both the profiling and prediction layers. Anti-training is essentially memory erasure at the algorithm level. Before anti-training, this embodiment first performs a storage layer deletion operation on the data fingerprint list, i.e., physical data erasure. This specifically involves overwriting all original data blocks corresponding to the fingerprint list in the underlying database, backup server, and log files to ensure that the data cannot be recovered and fully meets the compliance requirements of the "right to be forgotten." At the profiling layer, the overall operational profile of the target fleet is corrected based on the fleet batch contribution matrix to generate the first profile. The fleet profile is a vector representation aggregated from the features of all vehicles within the fleet. When some vehicle data is deleted, the profile must be adjusted accordingly. However, simply subtracting the features of the deleted vehicles may distort the profile. Therefore, this embodiment utilizes the operational correlation between the remaining vehicles in the target fleet and the vehicles to be deleted to offset cross-vehicle impact. The profile quality is iteratively adjusted, i.e., the first profile is adjusted, until the deviation between the corrected first profile and the theoretical profile regenerated using only the first remaining data is less than a preset profile deviation threshold. The final corrected first profile is the second profile. It should be noted that in actual logistics operations, cross-vehicle impact means that when some vehicles in the fleet are removed, such as through lease termination or scrapping, the freight volume originally undertaken by these vehicles does not disappear into thin air, but is absorbed by other remaining vehicles in the fleet through methods such as increasing their shifts.

[0107] Simply subtracting from the vehicle profile would severely underestimate the fleet's future operational capabilities, leading to a distorted profile. Therefore, this embodiment of the invention uses the operational correlation between retained vehicles and vehicles to be deleted to offset cross-vehicle impact. For example, firstly, the Pearson correlation coefficient between the vehicle to be deleted and each retained vehicle on historically frequent routes and cargo types is calculated. Subsequently, during iterative adjustments, some feature values ​​of the vehicle to be deleted, such as average monthly mileage and average daily revenue, are used as "business transfer volume" according to the proportion of the correlation coefficient, and are added to the feature vectors of the highly correlated retained vehicles to offset the impact of a sharp drop in data caused by the reduction of vehicles.

[0108] In this embodiment of the invention, a bias measure is used to determine the degree of difference between the iteratively adjusted first portrait and the portrait regenerated using only the first retained data. Both the iteratively adjusted first portrait and the portrait regenerated from the first retained data are represented as n-dimensional feature vectors, denoted as follows: and Specifically, this deviation is calculated using the Euclidean distance formula. The expression is as follows:

[0109]

[0110] in, This represents the first image after iterative adjustments. This represents the feature of the k-th dimension in the first portrait of the iteratively adjusted image, where k ranges from 1 to n; This indicates that the profile was regenerated using only the first retained data. This indicates the feature of the k-th dimension in the regenerated image using only the first retained data, where k ranges from 1 to n.

[0111] This deviation will be continuously monitored during the iterative allocation of "business transfer volume". The process continues until the image converges and falls below the preset image deviation threshold.

[0112] The resulting profile, or second profile, determined through iterative adjustments, actually represents the fleet's true remaining operational capacity after vehicle reductions and internal business integration. In subsequent steps, this second profile will serve as a characteristic representation of the "retained data of the target fleet," and will be weighted and fused with extracted industry average data to form the final model input data.

[0113] It should be noted that the profiling layer is located at the front end of the core prediction model. It is responsible for aggregating and processing the massive and heterogeneous raw vehicle operation data from the underlying layer, and combining it with business logic. For example, it transforms the compensation business transfer volume into a high-dimensional feature vector that the model can directly read and understand, i.e., a profile of the overall fleet operation. The feature vector output by the profiling layer serves as the input to the prediction layer.

[0114] After generating the second profile of the target vehicle fleet, the second retention data corresponding to the second profile is determined; here, the second profile is the characteristic representation of the second retention data; the second retention data corresponds to the first retention data, the first retention data is the initial retention data, and the second retention data is the retention data after the initial retention data has been corrected.

[0115] In one possible implementation, the step of performing collaborative anti-training on the credit granting model based on the fleet batch contribution matrix to generate an anti-trained credit granting model further includes:

[0116] A preset number of samples are randomly selected from the first retention data as a retention anchor set. The loss function of the retention anchor set under the credit model is determined. The average gradient is obtained based on the loss function and model parameters. The average gradient is used as the retention gradient vector.

[0117] Determine the gradient vector of the vehicle data to be deleted under the credit model, and use the gradient vector as the original deletion gradient vector;

[0118] Determine the projection component of the original deletion gradient vector in the direction of the retained gradient vector, and use the projection component as the feature entanglement gradient;

[0119] A purified inverse gradient is generated based on the original deletion gradient vector and the feature entanglement gradient.

[0120] Based on the purified reverse gradient, aggregated reverse gradient descent is performed, and combined with regularization constraints, to generate the reverse training confidence model.

[0121] In the prediction layer, this embodiment of the invention performs a gradient orthogonal purification step to preserve the common feature structure of the target fleet while eliminating the influence of the unique features of the vehicles to be deleted, thus preventing the accidental deletion of common features. Then, an algorithm combining aggregated backward gradient descent with regularization constraints is used to adjust the parameters of the current credit model based on the data features of the vehicles to be deleted, generating a reverse-training baseline version model, which is the reverse-training credit model.

[0122] For gradient orthogonal purification, the first step is to identify the common features of the vehicle fleet. To do this, a predetermined number of samples are randomly selected from the retention data of the target vehicle fleet as a set of retention anchors. These retention anchors represent information that should not be deleted and should be retained by the model. It should be noted that the retention data here refers to the initial retention data, i.e., the first retention data. If the retention data corresponding to the corrected profile, i.e., the second retention data, is used, it will disrupt the true mathematical distribution of the original features of each vehicle because it has been artificially accumulated with virtual business transfer volume. This will lead to a deviation in the calculated retention gradient direction, making it impossible to accurately extract the true common features.

[0123] Next, the average gradient of the loss function of the anchor point set with respect to the model parameters under the current credit granting model is calculated, and this gradient is defined as the retention gradient vector, denoted as . The direction of this vector represents the direction in which the parameters should be updated or maintained in order for the model to fit the retained data. Simultaneously, the original deletion gradient vector of the data for the vehicles to be deleted under the current credit model is calculated, denoted as... In traditional machine forgetting, it usually proceeds directly along... The direction of the updated parameters. However, in this embodiment of the invention, it is considered that... The data contains both unique features of the data to be deleted and common features of the fleet. Directly updating in reverse would destroy these common features. Therefore, the original deletion gradient vector needs to be calculated. Retaining gradient vectors The projection component in the direction is defined as the feature entanglement gradient, denoted as . According to the vector projection formula, this projection component can be calculated using the following expression:

[0124]

[0125] in, This represents the dot product operation of vectors. This represents the magnitude of the retained gradient vector. The entangled gradient represents the part of the gradient of the data to be deleted that is in the same direction as the gradient of the retained data, i.e., the part with common features.

[0126] Subsequently, a gradient purification formula is constructed. By subtracting the feature entanglement gradient from the original deletion gradient vector, the purified inverse gradient, orthogonal to the retained gradient vector, is obtained, denoted as . The expression is as follows:

[0127]

[0128] Among them, due to and It is orthogonal (perpendicular), which means along Adjusting the model parameters in terms of direction should theoretically not affect the model's ability to retain data. (Direction) Fitting ability.

[0129] Finally, the purified reverse gradient was used. Instead of the original deleted gradient vector, perform an aggregated backward gradient descent operation. Along... The direction of the regularization constraint is determined, and the model parameters are fine-tuned. For example, L2 regularization can be selected to prevent excessive parameter drift.

[0130] By performing the above collaborative anti-training, the feature weights of vehicles to be deleted are removed, and an anti-training trust model is generated.

[0131] Through the above steps, the influence of the unique features of the vehicles to be deleted was successfully eliminated while the common feature structure of the target fleet was preserved, generating the anti-training baseline version model, i.e., the anti-training trust model.

[0132] To address the problem of accidental deletion caused by the high homogeneity of fleet data, this invention introduces a gradient orthogonal purification technique in the prediction layer. By calculating the projection of the gradient of the data to be deleted onto the gradient direction of the retained data, "common features" are separated from "specific noise," and only specific gradient components orthogonal to the retained data are removed. This fundamentally solves the "catastrophic forgetting" problem in the field of machine forgetting, ensuring that the model can remove the influence of specific vehicles without disrupting the cognitive structure of the entire fleet operation model.

[0133] Step 106: Input the first retained data of the target fleet into the anti-training trust model to generate the first trust result; the first retained data is the original feature data of the retained vehicles of the target fleet.

[0134] In one possible implementation, the step of inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate a first trust result includes:

[0135] The overall business profile generated based on the first retained data and the first profile are used to determine the second profile, and the corresponding second retained data are determined based on the second profile.

[0136] Build industry average data;

[0137] Extract the target feature dimensions from the second retained data that correspond to the industry average data;

[0138] The industry average data and the second retained data are weighted and fused along the target feature dimension, while the data of the second retained data in the non-target feature dimension are retained to obtain fused data;

[0139] The fused data is input into the anti-training trust model to obtain the first trust result.

[0140] Specifically, the input to the anti-training credit model includes industry average data and second retention data, where the second retention data is the retention data corresponding to the corrected profile. The industry average data and the second retention data are weighted and fused according to a preset fusion weight ratio, and then input into the anti-training credit model to obtain the initial credit result, i.e., the first credit result. The weight ratio can be dynamically adjusted according to the richness of the fleet retention data; the less retention data, the higher the weight of the industry data. It should be noted that since the dimensions of the industry average data and the second retention data are inconsistent, in this embodiment of the invention, the weighted fusion of the industry average data and the second retention data means first extracting the target feature dimension corresponding to the industry average data from the second retention data, then weighting and fusioning the data of the industry average data and the second retention data on the target feature dimension according to a preset fusion weight ratio, while also retaining the data of the second retention data on non-target feature dimensions, ultimately obtaining the fused data. The non-target feature dimensions and the target feature dimensions constitute all the feature dimensions of the second retention data.

[0141] In one possible implementation, the construction of industry average data includes:

[0142] Extract data units that match the dimensions of the target fleet from a pre-defined multidimensional data pool;

[0143] Extract the average operating mileage, revenue fluctuation range, loan repayment rate, and vehicle failure rate from the data unit;

[0144] Determine whether the number of valid samples within the data unit is less than the statistical confidence threshold;

[0145] If so, based on a pre-constructed regional economic similarity matrix, a set of associated regions is identified; the set of associated regions is a set of regions whose similarity to the region where the target fleet is located is higher than a preset similarity threshold.

[0146] Extract alternative data from the associated region set that are consistent with the vehicle type and fleet size dimensions of the target fleet; the alternative data includes average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate;

[0147] The regional correction coefficient is determined based on the logistics activity index of the first region where the target fleet is located and the logistics activity index of the second region of the associated region set.

[0148] The substitute data is corrected based on the regional correction coefficient;

[0149] The revised alternative data will be used as the industry average data.

[0150] Otherwise, the average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate in the data unit will be used as the industry average data.

[0151] After detraining, although the model eliminates dirty data, it still faces the risk of an information vacuum, especially when the remaining data of the fleet is insufficient to support effective credit granting. Therefore, this embodiment introduces a data hedging mechanism. Specifically, industry average data matching the target fleet dimension is extracted from a pre-built multi-dimensional alternative data pool. The multi-dimensional alternative data pool is a pre-constructed, large-scale industry statistical database, whose dimensions include at least regional, vehicle type, and fleet size dimensions. For example, the dimension could be divided into "East China region - 4.2-meter refrigerated trucks - 50 to 100 vehicles".

[0152] First, analyze the target fleet's operating area, vehicle type, and current fleet size to determine the target fleet's multidimensional labels.

[0153] Then, data units matching the multidimensional labels are retrieved from the multidimensional alternative data pool. If the retrieval is successful, the average operating mileage, revenue fluctuation range, repayment fulfillment rate, vehicle failure rate, and other key indicators from the data unit are extracted as industry average data. It should be noted that the data in the multidimensional alternative data pool is updated according to a preset period (e.g., monthly), and outliers have been removed before being stored to ensure data representativeness.

[0154] However, in actual operation, extreme cases of data sparsity may be encountered, meaning that there are very few samples for a specific combination of dimensions. To address this issue, this embodiment of the invention includes a data sparsity verification and homogenization compensation step before determining the industry average data.

[0155] Specifically, it is determined whether the number of valid samples within the data units that perfectly match the target fleet dimension retrieved from the multidimensional alternative data pool is lower than a preset statistical confidence threshold. Considering the stability requirements of the financial risk control model, this statistical confidence threshold can be set to 30 to meet the lower limit of small sample statistical inference. Alternatively, the statistical confidence threshold can be set according to the sparsity of the fleet data. For example, the range of the statistical confidence threshold can be set to 5%–10% of the total number of data unit samples.

[0156] If the effective sample size is lower than the statistical confidence threshold, it indicates that directly using the data unit carries a risk of statistical bias. Since the impact of vehicle type and fleet size on operational characteristics is usually rigid, keeping the vehicle type and fleet size dimensions unchanged triggers a relaxation mechanism for the region dimension.

[0157] The construction of the regional economic similarity matrix involves: firstly extracting macroeconomic indicators for each region in the country from a macroeconomic database, including but not limited to regional GDP, industrial added value growth rate, and population density, to construct an economic feature vector for each region; calculating the cosine similarity between any two regional vectors, and combining the similarity scores of all regional pairs into a two-dimensional symmetric matrix, which is the regional economic similarity matrix, denoted as Meco. The elements m in the matrix... x,y This represents the degree of economic similarity between region x and region y.

[0158] Then, based on a pre-constructed regional economic similarity matrix, a set of associated regions with a similarity higher than a preset similarity threshold to the region where the target convoy is located is identified. For example, if the target region is "Mountainous County A" and there is insufficient data, "Mountainous County B" and "Mountainous County C" with similar economic and geographical characteristics can be found as associated regions.

[0159] Preliminary replacement data for the associated regional set is extracted in terms of vehicle type and fleet size. To ensure that the borrowed data accurately reflects the local situation, the preliminary replacement data needs to be homogenized. The first regional logistics activity index for the target fleet's location and the second regional logistics activity index for the associated regional set are obtained. This index can be a comprehensive index determined based on local GDP, total logistics volume, or highway traffic volume. The ratio of the first regional logistics activity index to the second regional logistics activity index is used as a regional correction coefficient. This coefficient is then used to weight and correct the average operating mileage, revenue fluctuation range, loan repayment rate, and vehicle failure rate in the preliminary replacement data. In other words, the preliminary replacement data is multiplied by the regional correction coefficient to obtain the corrected replacement data. The corrected data is then used as the final industry average data. This process solves the problem of scarce local data.

[0160] If the effective sample size is greater than or equal to the statistical confidence threshold, the average operating mileage, income fluctuation range, repayment performance rate, and vehicle failure rate in the data unit will be used as the final industry average data.

[0161] In one possible implementation, the logistics activity index of the first region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of the region where the target fleet is located; the logistics activity index of the second region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of each associated region in the associated region set.

[0162] First Regional Logistics Activity Index The calculation is performed using a multi-dimensional linear weighted method, and its expression is as follows:

[0163]

[0164] in, This represents the normalized index of the total local GDP of the target region. The normalized index representing the total logistics volume of the target region. The normalized index representing the traffic volume at highway entrances and exits in the target area. , , These are preset weighting coefficients; for example, they can be... , , The values ​​were 0.3, 0.4, and 0.3 in sequence.

[0165] The logistics activity index for the second region is calculated using the same method as the logistics activity index for the first region. The only difference is that the data used for linear weighting is the average of the indicators of each region within the set of related regions.

[0166] In one possible implementation, after inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate the first trust result, the method further includes:

[0167] Based on the impact level, determine whether to adjust the credit limit in the first credit granting result to determine the second credit granting result.

[0168] Although industry data has been incorporated into the initial credit approval results, a final strategy calibration of the credit limit is still needed based on the impact level in order to further control risks.

[0169] Specifically, after determining the impact level of the deletion event on the current credit granting model, the system determines whether to adjust the credit limit in the initial credit granting result based on the impact level. If the impact level is mild, it indicates that the model fluctuation is within a controllable range, and the credit limit in the initial credit granting result remains unchanged. If the impact level is moderate, it indicates that there is some risk, and the credit limit in the initial credit granting result is reduced by a first preset percentage, for example, 5%. If the impact level is severe, it indicates that the risk is high, and the credit limit in the initial credit granting result is reduced by a second preset percentage, triggering a manual review process. For example, the second preset percentage can be set to 10%. It should be noted that in this embodiment of the invention, the second preset percentage is greater than the first preset percentage, reflecting the principle of matching risk with control measures.

[0170] The above process achieves credit policy calibration. The adjusted credit limit and related terms constitute the new credit policy, namely the second credit result, and are pushed to the business side for execution.

[0171] This invention constructs a multi-dimensional alternative data pool and designs a dimensional relaxation and homogenization compensation mechanism. In extreme cases where data is sparse for specific regions or vehicle models, it automatically reduces dimensional constraints, finds related regional data with similar economic and geographical characteristics, and performs homogenization correction using logistics activity indicators. This mechanism transforms "dead" historical data into "live" risk hedging tools. Combined with a tiered credit limit calibration strategy based on impact levels, it achieves a smooth transition of credit business and automatic convergence of risk exposure in situations of data gaps.

[0172] In one possible implementation, after determining the second credit result, the method further includes:

[0173] Generate a consistency verification report containing proof of storage layer deletion, proof of collaborative anti-training, and proof of trust policy calibration; construct a version chain containing the version before batch deletion, the anti-training baseline version, and the version after policy calibration; associate the consistency verification report and the version chain with the unique event identifier generated by the fleet batch deletion request, and store the evidence on the chain to achieve full-link audit traceability.

[0174] To meet the requirements of financial regulation and compliance auditing, this embodiment of the invention also includes a full-chain audit traceability step. Specifically, a consistency verification report is automatically generated, which includes a storage layer deletion proof, a collaborative detraining proof, and a policy calibration proof (proving the basis for adjusting the credit granting policy). The storage layer deletion proof demonstrates that the physical data has been cleared, the collaborative detraining proof demonstrates that the model parameters have been updated and specific features have been removed, and the policy calibration proof demonstrates that the output of the credit granting model has been calibrated, thus obtaining the final credit granting result.

[0175] At the same time, a version chain is constructed, which sequentially links the version before batch deletion, the anti-training benchmark version, and the version after policy calibration, recording the complete trajectory of model evolution.

[0176] After confirming the consistency verification report and version chain, the data hash values ​​of the consistency verification report and version chain are associated with the fleet batch deletion event identifier and stored on the blockchain. Here, "on-chain" refers to writing the hash digest of key data into the blockchain network to ensure the immutability and traceability of the data.

[0177] In summary, the above embodiments employ a collaborative inverse training method based on the contribution matrix, which can accurately locate the feature weights of the data to be deleted in the model and remove them, significantly reducing computational costs. Specifically, a gradient orthogonal purification mechanism is introduced during the inverse training process, decomposing the gradients of the data to be deleted into unique gradients and common gradients, and updating only the unique gradients. This mechanism ensures that while thoroughly removing the target data, the common feature structure of the vehicles remaining in the fleet is well preserved, avoiding the misallocation of credit scores for remaining vehicles due to "feature entanglement," and maintaining the discrimination stability of the credit granting model under local data loss.

[0178] This invention establishes a risk hedging and strategy calibration system based on multidimensional alternative data. When faced with an information vacuum left by data deletion, it can automatically extract industry average data from a multidimensional data pool for fusion and supplementation. It also implements dimensional relaxation and homogenization compensation for data-sparse scenarios and utilizes regional economic similarity to call reference data across regions, ensuring that there is still a basis for action even in extreme data scarcity situations. Combined with a dynamic strategy calibration mechanism, it can adjust credit limits in a step-by-step manner based on the impact level of the deletion event on the model, rather than abruptly cutting off loans. This soft-landing approach, while meeting compliance requirements, maximizes the continuity of financial services and customer experience, achieving a dynamic balance between privacy protection and financial risk control.

[0179] This invention addresses the challenges of credit model performance collapse, credit assessment distortion, and accidental deletion of common features when fleets initiate batch data deletion requests (e.g., for privacy protection, changes in operating entities, or compliance requirements). Specifically, it constructs a closed-loop technical solution encompassing multi-level graph recognition, contribution matrix calculation, gradient orthogonal purification and reverse training, and multi-dimensional alternative data hedging. This solution prevents such data loss from causing issues like erroneous deletion of common features. While meeting compliant deletion requirements, this invention maximizes the stability of the credit model and the continuity of the strategy.

[0180] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for vehicle financial credit granting, characterized in that, include: In response to a data deletion request, the data fingerprints of the vehicles to be deleted in the target fleet are determined based on a preset association graph, and a list of data fingerprints is generated. The data fingerprint is a fixed-length string generated by salting and hashing the original data of the vehicle to be deleted. Based on the data fingerprint list and the initial data of the target fleet, the first retained data of the target fleet is determined; Based on the data fingerprint list, a simulation algorithm is used to determine the impact level of data deletion on the credit model; Based on the multi-dimensional characteristics of each vehicle in the target fleet and the multi-dimensional characteristics of the vehicles to be deleted, the total fleet batch contribution matrix of the vehicles to be deleted to the credit model is determined. Based on the total contribution matrix of the fleet batch, collaborative anti-training is performed on the credit granting model to generate an anti-trained credit granting model; The first retained data of the target fleet is input into the anti-training trust model to generate the first trust result; the first retained data is the original feature data of the vehicles retained by the target fleet.

2. The vehicle financial credit granting method according to claim 1, characterized in that, After inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate the first trust result, the method further includes: Based on the impact level, determine whether to adjust the credit limit in the first credit granting result to determine the second credit granting result.

3. The vehicle financial credit granting method according to claim 1, characterized in that, The determination of the impact level of data deletion on the credit model based on the data fingerprint list and using a simulation algorithm includes: Based on the data fingerprint list, and using a probability distribution simulation algorithm to simulate a data deletion scenario, the change in the credit model prediction error rate and the fluctuation in the credit limit before and after data deletion are determined. The impact of data deletion on the credit granting model is determined based on the change in the prediction error rate of the credit granting model, the fluctuation of the credit limit, and the preset impact level; the impact level includes mild, moderate, and severe.

4. The vehicle financial credit granting method according to claim 1, characterized in that, The step of performing collaborative inverse training on the credit granting model based on the fleet batch contribution matrix includes: The overall operational profile of the target fleet is corrected based on the fleet batch contribution matrix to obtain the first profile of the target fleet; the overall operational profile is a vector formed by aggregating the features of all vehicles in the target fleet. Adjust the first profile until the deviation between the adjusted first profile and the overall business profile generated from the first retained data is less than a preset deviation threshold, and then generate the second profile of the target fleet. Determine the second retained data corresponding to the second profile; the second retained data is the corrected retained data corresponding to the first retained data.

5. The vehicle financial credit granting method according to claim 1, characterized in that, The determination of the fleet batch contribution matrix of the vehicles to be deleted to the credit model based on the multi-dimensional features of each vehicle in the target fleet and the multi-dimensional features of the vehicles to be deleted includes: Determine the multi-dimensional characteristics of each vehicle in the target fleet based on the initial data of the target fleet; The vehicle to be deleted is selected from the multi-dimensional features of each vehicle in the target fleet. Determine the characteristics of all vehicles in the target fleet in each dimension; Based on the features of the vehicle to be deleted in each dimension and the sum of features in the corresponding dimension, the feature ratio of the vehicle to be deleted in each dimension is obtained, and the feature ratio is used as the single vehicle feature contribution of the vehicle to be deleted in each dimension. The contribution of each vehicle's features across all dimensions is arranged in a matrix to obtain the total contribution matrix of the fleet.

6. The vehicle financial credit granting method according to claim 1, characterized in that, The step of performing collaborative anti-training on the credit granting model based on the fleet batch contribution matrix to generate an anti-trained credit granting model further includes: A preset number of samples are randomly selected from the first retention data as a retention anchor set. The loss function of the retention anchor set under the credit model is determined. The average gradient is obtained based on the loss function and model parameters. The average gradient is used as the retention gradient vector. Determine the gradient vector of the vehicle data to be deleted under the credit model, and use the gradient vector as the original deletion gradient vector; Determine the projection component of the original deletion gradient vector in the direction of the retained gradient vector, and use the projection component as the feature entanglement gradient; A purified inverse gradient is generated based on the original deletion gradient vector and the feature entanglement gradient. Based on the purified reverse gradient, aggregated reverse gradient descent is performed, and combined with regularization constraints, to generate the reverse training confidence model.

7. The vehicle financial credit granting method according to claim 1, characterized in that, Before determining the data fingerprint of the vehicle to be deleted in the target fleet, the process also includes: The percentage of vehicles requested for deletion is determined based on the number of vehicles currently requested for deletion in the target fleet and the total number of vehicles in the target fleet. If the proportion of the number of vehicles requested for deletion is greater than or equal to the preset batch deletion trigger threshold, or if the data deletion request is a full fleet data deletion request for the target fleet, a fleet batch deletion event identifier is generated.

8. The vehicle financial credit granting method according to claim 4, characterized in that, The step of inputting the first retained data of the target vehicle fleet into the anti-training trust model to generate the first trust result includes: The overall business profile generated based on the first retained data and the first profile are used to determine the second profile, and the corresponding second retained data are determined based on the second profile. Build industry average data; Extract the target feature dimensions from the second retained data that correspond to the industry average data; The industry average data and the second retained data are weighted and fused along the target feature dimension, while the data of the second retained data in the non-target feature dimension are retained to obtain fused data; The fused data is input into the anti-training trust model to obtain the first trust result.

9. The vehicle financial credit granting method according to claim 8, characterized in that, The industry average data mentioned includes: Extract data units that match the dimensions of the target fleet from a pre-defined multidimensional data pool; Extract the average operating mileage, revenue fluctuation range, loan repayment rate, and vehicle failure rate from the data unit; Determine whether the number of valid samples within the data unit is less than the statistical confidence threshold; If so, based on a pre-constructed regional economic similarity matrix, a set of associated regions is identified; the set of associated regions is a set of regions whose similarity to the region where the target fleet is located is higher than a preset similarity threshold. Extract alternative data from the associated region set that are consistent with the vehicle type and fleet size dimensions of the target fleet; the alternative data includes average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate; The regional correction coefficient is determined based on the logistics activity index of the first region where the target fleet is located and the logistics activity index of the second region of the associated region set. The substitute data is corrected based on the regional correction coefficient; The revised alternative data will be used as the industry average data. Otherwise, the average operating mileage, revenue fluctuation range, repayment fulfillment rate, and vehicle failure rate in the data unit will be used as the industry average data.

10. The vehicle financial credit granting method according to claim 9, characterized in that, The logistics activity index of the first region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of the target fleet's region; the logistics activity index of the second region is determined based on the total GDP, total logistics volume, and highway entrance and exit traffic volume of each associated region in the associated region set.

Images