Customer risk identification method and computer readable storage medium

Abstract

The invention provides a customer risk identification method and a computer readable storage medium, and the method specifically comprises the following steps: (1) carrying out the first physical health condition evaluation based on a customer image of a trusted customer, and generating a physical health condition evaluation report and a health risk probability; and (2) when the health risk probability is greater than a set probability threshold value, carrying out secondary body health condition assessment based on the body health condition assessment report, and when the health risk score is greater than a set score threshold value, determining that the corresponding credible customer is a risk customer. According to the invention, through secondary body health condition assessment, the risk assessment of the credit auditing user is realized, the positive sample and the negative sample are automatically screened out, the related model is further trained based on the automatically screened positive sample and negative sample, and the manual intervention degree is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of credit review technology, and more specifically, this invention relates to a customer risk identification method and a computer-readable storage medium. Background Technology

[0002] With the rapid development of artificial intelligence, it is reshaping the global socio-economic landscape at an astonishing pace, becoming a core engine driving human civilization's progress. In key sectors such as finance, manufacturing, and healthcare, AI is evolving from single-point technological applications to systemic productivity transformations. In the financial sector, it has become a crucial hub connecting financial institutions and the real economy, undergoing a paradigm shift from "experience-dependent decision-making" to "digital intelligent governance."

[0003] Image-based credit review systems, as a practical and effective application of artificial intelligence technology in the financial sector, demonstrate their significant advantages in automating task processing. These systems can assist or even replace traditional credit reviewers by extracting and comparing key features from various materials submitted by clients, thereby verifying the authenticity of information and identifying client risks, greatly saving manpower and improving work efficiency.

[0004] However, image samples suffer from a natural lack of labels due to objective factors. Existing technical solutions that rely on manual annotation suffer from an extreme imbalance in the ratio of positive to negative samples, with the number of positive samples far exceeding that of negative samples, resulting in extremely high costs for obtaining negative samples. Summary of the Invention

[0005] In view of this, this application provides a customer risk identification method, which aims to improve at least one of the above-mentioned problems.

[0006] Specifically, the following technical solutions are included:

[0007] On the one hand, embodiments of this application provide a customer risk identification method, the method being as follows:

[0008] (1) Conduct the first physical health assessment based on the customer image of the credit review customer, and generate a physical health assessment report and health risk probability;

[0009] (2) When the probability of health risk is greater than the set probability threshold, a second health assessment is conducted based on the health status assessment report. When the health risk score is higher than the set score threshold, the corresponding credit review customer is identified as a risk customer.

[0010] In some embodiments of the present invention, the method for generating a physical health assessment report is as follows:

[0011] The client image is divided into a background area and a target area containing only the client image. Based on the background area, medical devices and medical facilities are extracted to form a first assessment report containing the type of medical device and / or the type of medical facility.

[0012] Disease detection is performed based on customer images in the target area to generate a second assessment report containing abnormal health status characteristics;

[0013] A physical health assessment report corresponding to the client's image is generated based on the first and second assessment reports.

[0014] In some embodiments of the present invention, the method for forming the probability of health risk is as follows:

[0015] (11) If the first assessment report contains a medical device type or medical facility type, generate number 1; otherwise, generate number 0.

[0016] (12) If there are abnormal health conditions in the second assessment report, generate the number 1; otherwise, generate the number 0.

[0017] (13) The weighted sum of the figures from all the first and second assessment reports is used to form the health risk probability of the first physical health assessment.

[0018] In some embodiments of the present invention, if the probability of health risk is 0, the customer image and its physical health assessment report are placed into the positive sample set as positive samples.

[0019] In some embodiments of the present invention, user images of at-risk customers and their corresponding physical health assessment reports are placed as negative samples into a negative sample set.

[0020] In some embodiments of the present invention, the first model evaluator, the second model evaluator, and the third model evaluator are trained periodically based on positive samples in the positive sample set and negative samples in the negative sample set. The third model evaluator is used to output the health risk score corresponding to the physical health status assessment report.

[0021] In some embodiments of the present invention, the first model evaluator and the second model evaluator adopt the Qwen2.5-VL-7B model, and the third model evaluator adopts the Qwen2.5-VL-72B model.

[0022] In some embodiments of the present invention, when the number of negative samples in the negative sample set is sufficient, the ratio of positive samples to negative samples participating in training is maintained at 1:1; when the number of negative samples in the negative sample set is small, the ratio of positive samples to negative samples participating in training does not exceed 3:1.

[0023] On the other hand, embodiments of this application provide a computer-readable storage medium including computer program instructions, which, when executed by a cluster of computing devices, enable the cluster of computing devices to perform the aforementioned customer risk identification method.

[0024] This invention not only achieves risk assessment of credit review users through a two-stage health status assessment, but also automatically filters out positive and negative samples. Based on the automatically filtered positive and negative samples, the relevant models are further trained, greatly reducing the degree of manual intervention. In addition, the first health status assessment mainly performs semantic-level feature recognition and extraction on customer images, while the second health status assessment infers customers with health risks based on the semantic features extracted in the first health status assessment, thereby improving the accuracy of identifying risky customers in the credit review process. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of a customer risk identification method provided in an embodiment of the present invention;

[0027] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Unless otherwise defined, all technical terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art.

[0030] Figure 1 The flowchart of the customer risk identification method provided in this embodiment of the invention is as follows:

[0031] (1) Conduct the first physical health assessment based on the customer image of the credit review customer, and generate a physical health assessment report and health risk probability;

[0032] In this embodiment of the invention, the customer image is divided into a background area and a target area containing only the customer image. Based on the background area, medical devices and medical facilities are extracted to form a first assessment report containing keywords, namely, the type of medical device and / or the type of medical facility. Based on the customer image in the target area, disease detection is performed to form a second assessment report containing abnormal health conditions, including: emaciation, abnormal complexion, and visible wounds. Based on the first and second assessment reports, a physical health assessment report corresponding to the customer image is formed.

[0033] In this embodiment of the invention, a customer image is input into a first model evaluator. The keywords in the first evaluation report output by the first model evaluator include medical device type and medical facility type. At this time, the first evaluation report can be generated based on one first evaluation model. When the keywords in the first evaluation report output by the first model evaluator include medical device type or medical facility type, at least two first evaluation models are required to generate the first evaluation report. A second evaluation report is generated based on the second model evaluation report. The keywords corresponding to medical device type and / or medical facility type are extracted from the first evaluation report, and abnormal health status features are extracted from the second evaluation report to form a physical health status evaluation report.

[0034] In this embodiment of the invention, the method for forming the probability of health risk is as follows:

[0035] (11) Generate a binary logic number based on each first assessment report. If the first assessment report contains a medical device type or a medical facility type, generate the number 1; otherwise, generate the number 0.

[0036] (12) Generate binary logic numbers based on the second assessment report. If there are abnormal health status characteristics in the second assessment report, generate number 1; otherwise, generate number 0.

[0037] (13) Set the weight values ​​of the first assessment report and the second assessment report, calculate the weighted sum of the numbers of all the first assessment reports and the second assessment reports, and form the health risk probability of the first physical health assessment.

[0038] Assuming the weight of the first assessment report is 0.22, there are two first assessment reports, and the score of the second assessment report is 0.5, if the first assessment report only mentions medical devices (IV drips), then a number 1 is generated based on the corresponding first assessment report; if it does not include medical facilities, then a number 0 is generated based on the corresponding first assessment report. Furthermore, if abnormal health status characteristics are detected in the second assessment report, a number 1 is generated based on that second assessment report. The corresponding health risk probability p = 0.25*1 + 0.25*0 + 0.5*1 = 0.75, where the health risk probability ranges from 0 to 1. When the health risk probability is 0, it indicates that the corresponding credit review user is healthy; when the health risk probability = 1, it indicates that the corresponding credit review user has a disease.

[0039] In this embodiment of the invention, if the health risk probability formed based on the first assessment report and the second assessment report is 0, then the customer image and its physical health assessment report are placed into the positive sample set as positive samples. In this embodiment of the invention, if the health risk probability formed based on the first assessment report and the second assessment report is greater than zero, but less than the set probability threshold, then the corresponding credit review customer is determined not to be a risk customer. The first physical health assessment is mainly to screen out customers with good physical health. Therefore, the probability threshold is low, and customers with poor or very poor physical health enter the second physical health assessment.

[0040] (2) When the probability of health risk is greater than the set probability threshold, a second health assessment is conducted based on the health status assessment report. When the health risk score is higher than the set score threshold, the corresponding credit review customer is identified as a risk customer.

[0041] The health status assessment report extracted from the customer's image is input into a third-model evaluator. The third-model evaluator outputs a health risk score corresponding to the health status assessment report. The health risk score ranges from 0 to 100. A health risk score of 0 indicates that the corresponding credit review customer is healthy, while a health risk score of 100 indicates that the corresponding credit review customer has an illness. This invention identifies the credit review customer as a high-risk customer when the health risk score exceeds a certain threshold. The scoring threshold is set according to actual needs. This invention sets the scoring threshold to 85. If the health risk score is higher than 85, the corresponding credit review customer is identified as a high-risk customer; if the health risk score is lower than 85, the corresponding credit review customer is identified as not a high-risk customer.

[0042] In this embodiment of the invention, the first physical health assessment is only a rough assessment of the health status of the credit review user. Therefore, the first and second model evaluators adopt a multimodal model with a small number of parameters (such as Qwen2.5-VL-7B). The second physical health assessment is a fine assessment of the credit review for those whose physical health assessment results are poor or extremely poor in the first physical health assessment. Therefore, the third model evaluator adopts a multimodal model with a large number of parameters (such as Qwen2.5-VL-72B).

[0043] (3) The user images of high-risk customers and their corresponding physical health assessment reports are used as negative samples and placed into the negative sample set.

[0044] In this embodiment of the invention, after step (3), the method further includes:

[0045] The first, second, and third model estimators are trained periodically based on positive samples from the positive sample set and negative samples from the negative sample set. In practical applications, the number of positive samples in the positive sample set and the number of negative samples in the negative sample set are still unbalanced, meaning the number of positive samples is much greater than the number of negative samples. Therefore, a random sampling strategy is implemented on the positive sample set based on the number of samples in the negative sample set: when the number of negative samples in the negative sample set is sufficient (greater than a set threshold), the ratio of positive to negative samples participating in training is maintained at 1:1; when the number of negative samples in the negative sample set is small, the ratio of positive to negative samples participating in training does not exceed 3:1. This sampling mechanism dynamically adjusts the sample ratio, balances the data distribution, provides a more balanced sample base for subsequent model training, and improves training effectiveness.

[0046] This invention not only achieves risk assessment of credit review users through a two-stage health status assessment, but also automatically filters out positive and negative samples. Based on the automatically filtered positive and negative samples, the relevant models are further trained, greatly reducing the degree of manual intervention. In addition, the first health status assessment mainly performs semantic-level feature recognition and extraction on customer images, while the second health status assessment infers customers with health risks based on the semantic features extracted in the first health status assessment, thereby improving the accuracy of identifying risky customers in the credit review process.

[0047] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center that includes one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct a computing device to perform a client risk identification method.

[0048] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only.

[0049] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A customer risk identification method, characterized in that, The method is as follows: (1) Conduct the first physical health assessment based on the customer image of the credit review customer, and generate a physical health assessment report and health risk probability; (2) When the probability of health risk is greater than the set probability threshold, a second health assessment is conducted based on the health status assessment report. When the health risk score is higher than the set score threshold, the corresponding credit review customer is identified as a risk customer.

2. The customer risk identification method as described in claim 1, characterized in that, The method for generating a physical health assessment report is as follows: The client image is divided into a background area and a target area containing only the client image. Based on the background area, medical devices and medical facilities are extracted to form a first assessment report containing the type of medical device and / or the type of medical facility. Disease detection is performed based on customer images in the target area to generate a second assessment report containing abnormal health status characteristics; A physical health assessment report corresponding to the client's image is generated based on the first and second assessment reports.

3. The customer risk identification method as described in claim 1, characterized in that, The specific methods for determining the probability of health risks are as follows: (11) If the first assessment report contains a medical device type or medical facility type, generate number 1; otherwise, generate number 0. (12) If there are abnormal health conditions in the second assessment report, generate the number 1; otherwise, generate the number 0. (13) The weighted sum of the figures from all the first and second assessment reports is used to form the health risk probability of the first physical health assessment.

4. The customer risk identification method as described in claim 1, characterized in that, If the probability of health risk is 0, then the customer's image and their physical health assessment report will be included as positive samples in the positive sample set.

5. The customer risk identification method as described in claim 1, characterized in that, User images of high-risk customers and their corresponding health status assessment reports are included as negative samples in the negative sample set.

6. The customer risk identification method as described in claim 5, characterized in that, The first model estimator, the second model estimator, and the third model estimator are trained periodically based on positive samples from the positive sample set and negative samples from the negative sample set. The third model estimator is used to output the health risk score corresponding to the physical health status assessment report.

7. The customer risk identification method as described in claim 6, characterized in that, The first and second model evaluators use the Qwen2.5-VL-7B model, and the third model evaluator uses the Qwen2.5-VL-72B model.

8. The customer risk identification method as described in claim 6, characterized in that, When there are enough negative samples in the negative sample set, the ratio of positive samples to negative samples participating in training is kept at 1:1; when there are few negative samples in the negative sample set, the ratio of positive samples to negative samples participating in training does not exceed 3:

1.

9. A computer-readable storage medium, characterized in that, It includes computer program instructions, which, when executed by a cluster of computing devices, perform the customer risk identification method as described in any one of claims 1 to 8.

Images