Deep learning-based automatic auditing method, system and equipment for application materials of staged consumption products, and medium

Abstract

The invention discloses an automatic auditing method, system and equipment for application materials of staged consumption products based on deep learning, and a medium, relates to the technical field of financial risk control, and solves the technical problem that the analysis capability of automatic auditing of the application materials of the staged consumption products is relatively weak. According to the technical scheme, the method is characterized in that feature extraction and structuring are carried out on application materials in an image form, then the application materials are simply filtered based on basic rules, then deep feature analysis is carried out on the application materials through a deep cross network, and finally a credibility score is output. If the credibility score is higher than a preset safety threshold value, the application material passes auditing; if the credibility score is lower than a preset danger threshold value, rejecting the application material; and if the credibility score is between a preset danger threshold value and a preset safety threshold value, transferring the application material to manual auditing. The workload of manual auditing is further reduced, and the auditing efficiency and the application experience of installment consumption product applicants are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial risk control, and in particular to a deep learning-based automatic review method, system, device and medium for application materials of installment consumption products. BACKGROUND

[0002] In financial services, banks and financial institutions mainly rely on manual review to process application materials for installment consumption products. However, the quality of manual review is highly dependent on the auditors, which inevitably limits the efficiency of the review work by the number of auditors. In addition, due to the uneven quality of manual auditors, some complex or intentional fraud application materials often cannot be accurately identified. The above factors significantly hinder the efficiency of risk control of financial services and the application experience of installment consumption product users. In order to assist the manual review process, the concept of automatic review technology for application materials has been proposed. Using the consistent processing power of computer systems, the automatic review technology for materials can quickly and standardize the verification of application materials, thereby providing more efficient and reliable review technology for financial services. Automatic review technology and manual review together form a modern pre-loan risk control system.

[0003] Generally, after the automatic review system receives the application materials of the user, it performs format checking and basic information extraction on the application materials, and performs the most basic information review, and finally sends the passed application materials to manual review, thereby reducing the workload of manual review. However, this basic information review is limited by the rules set by humans, and due to the lack of deep analysis capability, the automatic review system cannot actively identify the internal logical contradictions or potential fraud patterns of the application materials, which limits the auxiliary effect of the automatic review system on manual review. Therefore, how to improve the automatic review capability of installment consumption product application materials needs to be solved by the present application. SUMMARY

[0004] The present application provides a deep learning-based automatic review method, system, device and medium for application materials of installment consumption products, which aims to improve the analysis capability of automatic review of installment consumption product application materials, further reduce the workload of manual review, improve the review efficiency and the application experience of installment consumption product applicants.

[0005] The above technical purpose of the present application is achieved by the following technical solutions: A deep learning-based automatic review method for application materials of installment consumption products, comprising: performing feature extraction on the application materials of installment consumption products to obtain extracted features; performing review on the extracted features through basic rules to obtain a first review result; Deep feature analysis is performed on extracted features that cannot be judged by basic rules through deep cross-network, and a confidence score is output. The second review result is obtained based on the credibility score; The final audit result is obtained by combining the results of the first and second audits; The deep cross-network includes a parallel deep network, a cross-network, a gated attention unit, and an output layer; The deep network includes K fully connected layers and ReLU activation function layers. After extraction, the features are passed through the K fully connected layers and ReLU activation function layers in sequence to generate a low-dimensional deep feature vector. The cross-network comprises L cross-layers. The complete feature vector passes through L cross-layers sequentially to generate a cross-feature vector. The operation of each cross-layer is as follows: ; in, This represents the complete feature vector of the input. Indicates the first The output features of the layer Indicates the first The output features of the layer Indicates the first Layer weights Indicates the first Layer bias; The gated attention unit concatenates the low-dimensional deep feature vector and the cross feature vector by weighting the gate weight g to obtain the final combined vector. The gated attention unit first performs a nonlinear transformation through a fully connected layer with a tanh activation function, then outputs a scalar score through a linear layer, and then uses the sigmoid function on the scalar score to generate the gate weight g. The output layer is a fully connected layer. The output layer uses the Sigmoid activation function to map the combined vector into an output value between [0,1], which is the confidence score.

[0006] Furthermore, the step of extracting features from the application materials to obtain the extracted features includes: The received image materials are classified to obtain the classification type of the image materials; Perform full-page text recognition on all image materials and convert the image materials into strings; The BERT-BiLSTM-CRF sequence labeling model is used to identify custom named entities in strings. The strings corresponding to each image material are divided into different structured features based on the classification type. Visual features are extracted from all image materials, including the detection of tampering and forgery, the detection and comparison of seals and signatures, and image quality assessment. Based on the string, cross-validation features are extracted based on the detection and comparison of the seal and signature. The cross-validation features include whether the names are consistent, whether the companies are consistent, and whether the incomes are consistent. The final extracted features include structured features, visual features, and cross-validation features.

[0007] Furthermore, by using a CNN model to detect tampering and forgery, the extracted features include: whether it is a screen capture, whether it is a copy after printing, whether there are traces of image modification, and whether key information is obscured. The YOLOv5-based object detection model is used to detect and compare seals or signatures, and a consistency score is given for seals or signatures in all application materials. The features extracted for image quality assessment include: image blur, whether the lighting is too dark, and whether the edges and corners are missing.

[0008] Furthermore, the basic rules include rejection rules and approval rules. The approval basic rules review the extracted features to obtain a first review result, including: First, the extracted features are reviewed to see if they meet the rejection rules. If they do, the application is automatically rejected. Otherwise, the extracted features are reviewed to see if they meet the approval rules. If the application meets the approval criteria, the application will be automatically approved; otherwise, the extracted features will be transferred to a deep cross-network for deep feature analysis.

[0009] Furthermore, the extraction of features for which the basic rules cannot provide a judgment result undergoes deep feature analysis, and a confidence score is output, including: The extracted features are divided into dense features and sparse features; among them, dense features are continuous numerical features, and sparse features are categorical features or discrete features. Normalize the dense features to obtain normalized dense features; map the sparse features into low-dimensional vectors to obtain low-dimensional sparse features. The normalized dense features and the low-dimensional sparse features are concatenated to obtain the complete feature vector. Deep feature analysis of the complete feature vector is performed using a deep cross-network, and a confidence score is output.

[0010] Furthermore, the loss function of the deep cross-network is defined as follows: ,in: ; Indicates the batch size of the data; Indicates the first The true label of each sample; This represents the previously obtained credibility score; Represents the regularization coefficient; This represents the set of trainable weight parameters for all layers in a deep cross-network.

[0011] Furthermore, the second review result obtained based on the credibility score includes: When the credibility score is higher than the preset security threshold, the application material is deemed to have passed the review; when the credibility score is between the preset danger threshold and the preset security threshold, the application material is forwarded to the manual review system and the automatic review task ends; when the credibility score is lower than the preset danger threshold, the application material is deemed to have failed the review.

[0012] An automated review system for application materials of installment consumption products based on deep learning is disclosed. This automated review system utilizes a deep learning-based method for automatically reviewing application materials for installment consumption products. The system includes: The feature extraction unit extracts features from the application materials for installment consumption products to obtain the extracted features. The first review unit reviews the extracted features using basic rules to obtain the first review result; The deep feature analysis unit performs deep feature analysis on the extracted features that cannot be judged by the basic rules through a deep cross-network, and outputs a confidence score. The second review unit obtains the second review result based on the credibility score; The output unit combines the first and second review results to obtain the final review result.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a deep learning-based method for automatically reviewing application materials for installment consumption products.

[0014] A computer storage medium storing a computer program, which, when executed by a processor, implements a deep learning-based method for automatically reviewing application materials for installment consumption products.

[0015] The above technical solution can achieve at least some of the following technical effects: The method, system, device, and medium for automatic review of application materials for installment consumption products based on deep learning described in this application first extracts and structures features from the application materials in image form, then performs simple filtering on the application materials based on basic rules, then uses a deep cross-network to perform deep feature analysis on the application materials, and finally outputs a credibility score. If the credibility score is higher than a preset safety threshold, the application material passes the review; if the credibility score is lower than a preset danger threshold, the application material is rejected; if the credibility score is between the preset danger threshold and the preset safety threshold, the application material is transferred to manual review.

[0016] Compared with existing technologies, this application utilizes deep learning technology to further enhance the analytical capabilities of automated review, further reduce the workload of manual review, and improve review efficiency and the application experience for installment consumption product applicants. Attached Figure Description

[0017] Figure 1 This is a flowchart of the automatic review method for application materials of installment consumption products based on deep learning in this application embodiment. Detailed Implementation

[0018] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0019] like Figure 1 As shown, the method for automatically reviewing application materials for installment consumption products based on deep learning described in this application includes: Step 1: Extract features from the application materials for installment payment products to obtain the extracted features.

[0020] The automated review system first receives images (such as scanned copies, photos, etc.) of application materials submitted by users applying for installment payment products via the App, Web, or offline. The goal of this step is to transform the image-based application materials into structured information and further identify and extract key features.

[0021] Preferably, the step of extracting features from the application materials to obtain the extracted features includes: Step 11: Classify the received image materials to obtain the classification type of the image materials.

[0022] Specifically, the automated review system categorizes the received image files and identifies their material type. This mainly includes identity documents (such as resident ID cards, driver's licenses, etc.), income or asset proof documents (such as bank statements, pay slips, tax payment certificates, property ownership certificates, vehicle registration certificates, etc.), employment proof documents (such as employment certificates with official seals, labor contracts, etc.), application forms (such as application forms filled out by customers that include basic personal information, family information, contact information, vehicle information, loan amount and term, etc.), and credit reports.

[0023] Step 12: Perform full-page text recognition on all image materials and convert the image materials into strings.

[0024] Specifically, after classification, the automated review system uses a pre-trained OCR model to perform full-page text recognition on all image materials, aiming to convert the image materials into machine-readable characters. OCR models are typically end-to-end models based on CRNN or Transformer, capable of directly outputting the complete string corresponding to each image.

[0025] Step 13: Use the sequence labeling model BERT-BiLSTM-CRF to identify custom named entities of strings, and combine the classification type to divide the strings corresponding to each image material into different structured features.

[0026] Specifically, after obtaining the complete string corresponding to each image, the sequence labeling model BERT-BiLSTM-CRF is used to implement custom named entity recognition for application materials of installment consumption products. This can divide the complete string corresponding to each image into different structured features. Specifically, the following features will be classified: identity information (including name, ID number, gender, ethnicity, date of birth, address, issuing authority, validity period, and level in the bank, etc.), application form information (including mobile phone number, marital status, education level, employer, employer address, employer phone number, current address, emergency contact person's name, emergency contact person's phone number, emergency contact person's relationship, applied vehicle model, applied amount, application period, down payment ratio, etc.), proof of employment information (including monthly income, full name of employer, contract start and end dates, etc.), and bank information (including transaction date, transaction time, transaction amount, counterparty, transaction summary, account balance, etc.).

[0027] Step 14: Extract visual features from all image materials, including detection of tampering and forgery, detection and comparison of seals and signatures, and image quality assessment.

[0028] Preferably, tampering and forgery are detected using a CNN model or a generative adversarial network. The extracted features include: whether it is a screen capture, whether it is a copy after printing, whether there are traces of image modification, and whether key information is obscured. For example, the label for presence is 1, and the label for absence is 0. If the output feature is (1,0,0,1), it indicates that a screen capture exists, a copy after printing does not exist, there are no traces of image modification, and key information is obscured.

[0029] Preferably, the seals and signatures are detected and compared using a YOLOv5-based object detection model, and a consistency score is given for the seals or signatures in all application materials.

[0030] Preferably, the features extracted for image quality assessment include: image blurriness, whether the lighting is too dark, and whether corners are missing. For example, the "yes" label is 1, and the "no" label is 0. Image blurriness can be divided into three labels: no blurriness, slight blurriness, and severe blurriness, represented by [0,1,2]. If the output feature is (2,0,1), it indicates severe blurriness, moderate lighting, and missing corners. Image blurriness can also be directly output as a numerical value, which can be set according to actual needs.

[0031] Step 15: Based on the string, extract cross-validation features based on the detection and comparison of the seal and signature. The cross-validation features include whether the name, company, and income are consistent. For example, the "yes" label is 1 and the "no" label is 0. If the output feature is (1,1,0), it means that the name, company, and income are consistent.

[0032] Specifically, considering consistency checks, the automated review system will perform preliminary logical checks based on the extracted features, calculating and generating cross-validation features, including name consistency, company consistency, and income consistency. For name consistency, it needs to check whether all application materials maintain the same name, such as the ID card, application form, and bank account holder name. For company consistency, it needs to check whether the company name on the application form, the company stamp on the employment certificate, and the company that pays the salary on the bank statement are consistent. For income consistency, it needs to check whether the monthly income matches the average transaction amount marked as salary in the bank statement. Name consistency and company consistency only output "consistent" or "inconsistent." Income consistency compares the average monthly income with the salary portion of the bank statement; if the absolute value of the difference is too large, it returns "inconsistent"; otherwise, it returns "consistent."

[0033] Step 16: The final extracted features include structured features, visual features, and cross-validation features.

[0034] Step 2: Review the extracted features using basic rules to obtain the first review result.

[0035] Preferably, the basic rules include rejection rules and approval rules. The specific review includes: first, reviewing whether the extracted features meet the rejection rules; if so, the application material is automatically rejected; otherwise, reviewing whether the extracted features meet the approval rules; if the approval rules are met, the application material is automatically approved; otherwise, the extracted features are transferred to a deep cross-network for deep feature analysis.

[0036] Specifically, in this embodiment of the application, the rejection rules include: (1) Missing materials: Any key material is missing, such as ID card, bank statement, application form, etc.; (2) Age mismatch: The age calculated based on the ID number is either below 18 or above 65; (3) Identity fraud: The current time has exceeded the validity period given by the ID card, or the ID card number does not meet the verification rules; (4) Information forgery: There is any kind of image modification or forgery, or the consistency score of the application materials is lower than a security threshold given by a person. (5) Information contradiction: Any one of the following checks fails: name consistency, formula consistency, or income consistency. (6) External risks: The applicant’s mobile phone number, ID card number, etc. are on the blacklist, which is a third-party anti-fraud service platform.

[0037] The rules categorize applicants into two scenarios. The first scenario is when the applicant is deemed a high-quality customer of the bank; the second scenario is when the applicant's risk of default is deemed extremely low. The specific determination methods are as follows: (1) The conditions for determining an applicant to be a premium customer of the bank are as follows: the customer label is a senior member or above of the bank; the total assets of the applicant in the bank are higher than a certain given value; the amount of the application is lower than a certain given value; and all rejection rules have been passed.

[0038] (2) The conditions for determining that the applicant’s risk of default is extremely low are as follows: the client’s employer is on the white list of high-quality employers, such as Fortune 500 companies, large state-owned enterprises, public institutions, civil servants, etc.; the average monthly salary in the bank statement is higher than a given multiple of the monthly payment of this application; and all rejection rules have been passed.

[0039] The aforementioned strict rules can be used for preliminary review of application materials. First, the rejection rules are determined. If any rejection rule is met, the application is automatically rejected and the process ends. If no rejection rule is met, but any approval rule is met, the application is automatically approved and the process ends. If no rejection rule is met and no approval rule is met, the application is transferred to a deep cross-network for review.

[0040] Step 3: Perform deep feature analysis on the extracted features for which the basic rules cannot provide a judgment result through a deep cross-network, and output a confidence score.

[0041] Preferably, it includes: Step 31: Divide the extracted features into dense features and sparse features; among them, dense features are continuous numerical features, and sparse features are categorical features or discrete features.

[0042] Specifically, dense features refer to continuous numerical features, including the applicant's age, remaining days of ID card validity, application amount, down payment ratio, average monthly salary calculated from bank statements, income stability (such as monthly income variance), average monthly expenses, signature consistency score, etc.

[0043] Sparse features are categorical or discrete features, including the applicant's province of origin, city of residence, marital status, education level, type of employer, relationship with emergency contact, and cross-validation features generated in step 1.

[0044] Step 32: Normalize the dense features to obtain normalized dense features; map the sparse features into low-dimensional vectors to obtain low-dimensional sparse features.

[0045] Specifically, in this application, a normalization method is used to shrink all dense features to a fixed numerical range in order to eliminate the influence caused by the difference in dimensions between different features.

[0046] This application uses an embedding layer technique to map each high-dimensional sparse feature to a low-dimensional vector. This mapping process is automatically learned through model training, ensuring that categories that are semantically or risk-related are also close to each other in the vector space.

[0047] Step 33: Concatenate the normalized dense features and the low-dimensional sparse features to obtain the complete feature vector.

[0048] Specifically, after normalizing all dense features and mapping all sparse features to vectors, the system concatenates these vectors together to form a complete feature vector. This complete feature vector represents all the features in the application materials and will be used as the input to the deep cross-network.

[0049] Step 34: Perform deep feature analysis on the complete feature vector using a deep cross-network and output a confidence score.

[0050] Preferably, the deep cross-network includes a parallel deep network, a cross-network, a gated attention unit, and an output layer.

[0051] The deep network includes K fully connected layers and ReLU activation function layers. After extraction, the features are passed through the K fully connected layers and ReLU activation function layers in sequence to generate a low-dimensional deep feature vector.

[0052] Specifically, a deep network is a multilayer perceptron consisting of multiple fully connected layers, used to automatically discover nonlinear combination relationships between features.

[0053] The cross-network comprises L cross-layers. The complete feature vector passes through L cross-layers sequentially to generate a cross-feature vector. The operation of each cross-layer is as follows: ; in, This represents the complete feature vector of the input. Indicates the first The output features of the layer Indicates the first The output features of the layer; Indicates the first Layer weights Indicates the first To avoid the vanishing gradient problem, residual learning is introduced to help build deeper networks and enhance their expressive power. This structure allows each layer to interact with the initial input, enabling the network to efficiently learn feature combinations in a multinomial form and avoid parameter explosion. Finally, the output of the Lth intersecting layer of the cross-branch network becomes the branch vector output by the cross-branch network.

[0054] After the deep network and the cross-network respectively calculate the vectors of the two branches, the automatic review system concatenates these two branch vectors, representing the implicit mode and the explicit cross-network respectively, to form a more complete combined vector. In this process, an attention mechanism is introduced. Specifically, this application introduces a gated attention unit to calculate the gate weight g after concatenating the branch vectors. The gate weight g is between [0,1], and g and (1-g) can represent the importance of the deep network branch and the cross-network branch, respectively. The gated attention unit first performs a nonlinear transformation through a fully connected layer with a tanh activation function, then outputs a scalar score through a linear layer. The scalar score is then processed using a sigmoid function to generate the gate weight g. The branch vectors generated by the two branch networks are weighted using the gate weight g and then concatenated to form the final combined vector.

[0055] Unlike standard deep cross-networks, this application introduces a multimodal attention fusion mechanism, specifically a gated attention unit. This gated attention unit can dynamically evaluate the importance of the two branches of the deep network and the cross-network based on the input features, thereby achieving adaptive feature fusion. This significantly improves the model's risk quantification accuracy and robustness.

[0056] Finally, the combined vector is fed into a fully connected layer that acts as the output layer, using the sigmoid activation function. This output layer maps the combined vector to an output value between [0,1], which is the credibility score defined in this application. The credibility score intuitively quantifies the authenticity of the installment payment product application materials and the applicant's probability of fulfilling their obligations. The closer the score is to 1, the higher the credibility and the lower the risk; the closer the score is to 0, the higher the risk of fraud or default.

[0057] Preferably, the deep cross-interference network uses a binary cross-entropy loss function and an L2 regularization term as the main loss function, defined as follows: ,in: ; Indicates the batch size of the data; Indicates the first The true label of each sample is 0 or 1 in this application; This represents the previously obtained confidence score. The L2 regularization term exists to prevent the model from overfitting. Represents the regularization coefficient. This represents the set of trainable weight parameters for all layers in a deep cross-network.

[0058] The training process of the deep cross-network is a supervised learning process. The system utilizes a large amount of labeled historical application data as the training set, where the feature vector of each sample serves as the input, and the final human review result serves as the real label. During training, the network calculates the predicted confidence score through forward propagation, and then updates all weight parameters in the deep network branches, cross-network branches, and gated attention units simultaneously by minimizing the overall loss composed of the binary cross-entropy loss function and the L2 regularization term using the backpropagation algorithm. Through iterative optimization, the network gradually learns how to extract effective information from the given features and perform weighted fusion, ultimately making the output confidence score match the actual risk situation as closely as possible.

[0059] Step 4: Obtain the second review result based on the credibility score.

[0060] Preferably, obtaining the second audit result based on the credibility score includes: Step 41: When the credibility score is higher than the preset security threshold, the application materials are deemed to have passed the review.

[0061] Step 42: When the credibility score is between the preset danger threshold and the preset safety threshold, the application materials are forwarded to the manual review system, and the automatic review task ends.

[0062] Step 43: If the credibility score is lower than the preset danger threshold, the application materials are deemed unqualified.

[0063] Step 5: Combine the results of the first and second audits to obtain the final audit result.

[0064] The automatic review system for application materials of installment consumption products based on deep learning described in this application includes a feature extraction unit, a first review unit, a deep feature analysis unit, a second review unit, and an output unit.

[0065] The feature extraction unit is used to extract features from the application materials for installment consumption products to obtain the extracted features.

[0066] The first review unit is used to review the extracted features according to basic rules and obtain preliminary review results.

[0067] The deep feature analysis unit is used to perform deep feature analysis on extracted features that cannot be judged by basic rules through deep cross-networks, and output a confidence score.

[0068] The second review unit is used to obtain the second review result based on the credibility score.

[0069] The output unit is used to combine the first review result and the second review result to obtain the final review result.

[0070] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A deep learning-based automatic review method for consumer product application materials, characterized by, The application comprises the following steps: characteristic extraction is performed on the application materials of the staged consumption product to obtain extracted features; the extracted features are audited through basic rules to obtain a first audit result; deep feature analysis is performed on the extracted features for which the basic rules cannot give a judgment result through a deep cross network to output a credibility score; a second audit result is obtained based on the credibility score; a final audit result is obtained by combining the first audit result and the second audit result; wherein the deep cross network comprises a parallel deep network, a cross network, a gated attention unit and an output layer; the deep network comprises K layers of fully connected layers and ReLU activation function layers, and after the extracted features pass through the K layers of fully connected layers and the ReLU activation function layers in sequence, a low-dimensional deep feature vector is generated; the cross network comprises L layers of cross layers, and after a complete feature vector passes through the L layers of cross layers in sequence, a cross feature vector is generated; wherein the operation of each cross layer is as follows: ； in, This represents the complete feature vector of the input. Indicates the first The output features of the layer Indicates the first The output features of the layer Indicates the first Layer weights, Indicates the first Layer bias; the gated attention unit splices the low-dimensional deep feature vector and the cross feature vector after weighting by a gating weight g to obtain a final combined vector; wherein the gated attention unit first performs nonlinear transformation through a fully connected layer with a tanh activation function, then outputs a scalar score through a linear layer, and then uses a Sigmoid function on the scalar score to generate the gating weight g; the output layer is a fully connected layer, and the output layer uses a Sigmoid activation function to map the combined vector into an output value between [0, 1], which is the credibility score.

2. The periodic consumer product application material automatic review method of claim 1, wherein, The characteristic extraction on the application materials to obtain the extracted features comprises the following steps: classifying the received image materials to obtain the classification types of the image materials; performing full-page text recognition on all image materials to convert the image materials into strings; performing custom named entity recognition on the strings using a sequence labeling model BERT-BiLSTM-CRF, and dividing the string corresponding to each image material into different structured features in combination with the classification types; extracting visual features of all image materials, including detection of tampering and forgery, detection and comparison of seals and signatures, and image quality assessment; based on the strings, cross-validation features are extracted based on the detection and comparison of the seals and signatures, the cross-validation features including whether the names are consistent, whether the companies are consistent, and whether the incomes are consistent; the finally obtained extracted features include structured features, visual features and cross-validation features.

3. The periodic consumer product application material automatic review method of claim 2, wherein, tampering and forgery are detected through a CNN model, and the extracted features include whether it is a screen shot, whether it is a printed copy, whether there are traces of image modification, and whether there are key information obstructions; seals or signatures are detected and compared through a target detection model based on YOLOv5, and a consistency score of the seals or signatures in all application materials is given; the image quality assessment extracts features including image blurriness, whether the light is too dark, and whether the corners are missing.

4. The periodic consumer application materials automatic review method of claim 3 wherein, The basic rules include veto rules and pass rules, and the auditing of the extracted features through the basic rules to obtain the first audit result comprises the following steps: The post-extraction features are audited for whether they meet the rejection rules, and if yes, the application materials are automatically rejected, otherwise, the post-extraction features are audited for whether they meet the passing rules; If the passing rules are met, the application materials are automatically passed, otherwise, the post-extraction features are transferred to the deep cross network for deep feature analysis.

5. The periodic consumer application materials automatic review method of claim 4, wherein, The post-extraction features for which the basic rules cannot give a judgment result are subjected to deep feature analysis, and a credibility score is output, including: The post-extraction features are divided into dense features and sparse features; wherein the dense features are continuous numerical features, and the sparse features are categorical features or discrete features; The dense features are normalized to obtain normalized dense features; and the sparse features are mapped into low-dimensional vectors to obtain low-dimensional sparse features; The normalized dense features and the low-dimensional sparse features are spliced to obtain a complete feature vector; The complete feature vector is subjected to deep feature analysis by the deep cross network, and a credibility score is output.

6. The periodic consumer application materials automatic review method of claim 5, wherein, The loss function of the deep cross network is defined as wherein: ； denotes the data batch size; denotes the true label of the th sample; denotes the previously obtained confidence score; denotes the regularization coefficient; denotes the set of trainable weight parameters of all layers in the deep cross network.

7. The installment consumer products application material automatic review method of claim 6 wherein, The second audit result is obtained based on the credibility score, including: When the credibility score is higher than a preset safety threshold, it is determined that the application materials pass the audit; when the credibility score is between a preset danger threshold and the preset safety threshold, the application materials are forwarded to an artificial audit system, and the automatic audit task ends; and when the credibility score is lower than the preset danger threshold, it is determined that the application materials do not pass the audit.

8. A deep learning-based automatic review system for consumer product application materials, the automatic review system being used for the automatic review method according to any one of claims 1 to 7, characterized in that, The automatic audit system includes: A feature extraction unit extracts features from the application materials of installment consumption products to obtain post-extraction features; A first audit unit audits the post-extraction features by the basic rules to obtain a first audit result; A deep feature analysis unit subjects the post-extraction features for which the basic rules cannot give a judgment result to deep feature analysis by the deep cross network, and outputs a credibility score; A second audit unit obtains a second audit result based on the credibility score; An output unit obtains a final audit result by combining the first audit result and the second audit result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the automatic audit method of any one of claims 1 to 7.

10. A computer storage medium storing a computer program, the computer program comprising instructions, which, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 9. The computer program is executed by the processor to implement the automatic audit method of any one of claims 1 to 7.

Images