Consumption finance credit mediation processing method and device

Abstract

The invention discloses a consumer financial credit mediation processing method and device. Data information is stored based on a structured database, and operations such as data processing, data analysis, feature derivation, feature screening and model training are carried out by using a structured query language (SQL) and a Python language. The model algorithm is based on LightGBM (lightweight, fast, high-performance and distributed decision tree-based gradient lifting tool). The feature selection reference indexes comprise an IV value (Infromation Value which can be used for representing the predictive ability of variables), a missing rate (the ratio of missing samples to total samples), a maximum single value ratio (the maximum value of the ratio of samples with the same value to the total samples), PSI (Population Stabilization Index which is used for evaluating the distribution change between two data sets or time points) and the like.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a consumer finance credit intermediary processing method and device. Background Art

[0002] In the field of machine learning, the consumer finance credit intermediary identification model obtains information such as the device, location, image, etc. of the consumer finance credit applicant (hereinafter referred to as the customer), combines the customer's historical credit and public credit information, and inputs it into the relevant machine learning model for training. The model is fine-tuned according to the characteristics of the credit intermediary identification task to achieve the purpose of improving the accuracy of credit intermediary identification.

[0003] Existing technology: In order to identify credit intermediaries to the greatest extent, existing technology usually collects customer equipment, location, image and other information, including basic equipment information, equipment installation APP information, call, text message, address book and other information stored on the device, as well as equipment location information and live image information, etc., and processes the above information through natural language recognition and association network analysis technologies, and then converts the above information into scores or lists through models or rules to identify high-risk customers of credit intermediaries.

[0004] Defects of existing technologies: When collecting customer device information, there is a risk of excessive information collection, and there are high requirements for the availability of device information. If the device information cannot be collected, all relevant data will be missing, which will greatly affect the effect of the intermediary identification model. At the same time, some methods adopt the form of blacklist + association network analysis. When performing association network calculations, higher computing resources are required, and the consumed computing resources will increase exponentially with the increase in the overall sample size, thereby greatly increasing the development and maintenance costs of the intermediary identification model. Summary of the invention

[0005] The technical problem to be solved by the present invention is to solve the deficiencies of the above-mentioned prior art and provide a consumer finance credit intermediary processing method and device.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A consumer finance credit intermediary processing method, characterized by comprising the following steps: S1. Process the features of the recognition model. The processing method is as follows: S11, data acquisition step: extracting data of multiple dimensions of the entire process of customer loan application from the historical database. When the customer is in the process of loan operation, the system automatically obtains the relevant information data of the current loan customer and transfers it to the database for structured storage; S12, data processing step: processing the acquired data information through statistical transformation and summarizing it into the customer dimension, corresponding to the input of the intermediary identification model; S13, sample selection step: extract customer samples that have been confirmed to have been processed through intermediaries and the corresponding features of the above samples as positive samples in subsequent model training samples, and select samples with a certain proportion of the number of positive samples from normal customer samples as negative samples in subsequent model training samples, and remove abnormal samples; S14, merging the positive and negative samples to obtain sample set 1, wherein all features included in the sample set 1 are used as feature pool 1; S2. Train the recognition model. The training method is as follows: S21, sample processing: divide sample set 1 into two parts, training samples and out-of-time verification samples, according to the time when the customer applied for the loan, and calculate the IV value, missing rate, maximum single value ratio and PSI index of the features in feature pool 1 on the training samples and out-of-time verification samples respectively; S22. Initial feature screening: Based on the above indicators, the features of feature pool 1 are screened, and features with IV < 0.01, missing rate > 95%, or maximum single value ratio > 95% are eliminated to obtain feature pool 2; S23, feature fine screening: Eliminate the features whose risk performance is inconsistent with business experience, whose risk performance is inconsistent between training samples and out-of-time validation samples, or whose PSI>0.1, and obtain feature pool 3; S24, model training: input the training samples containing the features of feature pool 3 into the LightGBM binary classification model for training to obtain the trained model file; in this process, adjust the model effect by adjusting the model parameters and input features, and perform scoring verification on the out-of-time verification samples; S3. Take the latest customer application samples containing the final use characteristics of the model, score them using the aforementioned model, and perform customer score distribution analysis; determine the score threshold of the intermediary identification model based on the risk distribution of different scores; S4. All credit applicants are scored using the identification model. Customers whose scoring results exceed the risk threshold are considered to be at risk. Customers within the threshold enter the normal review process and ultimately obtain the review results.

[0007] Furthermore, the data extracted in step S11 includes ID number, mobile phone number, self-filled contact information, customer credit information, and historical lending behavior information.

[0008] Furthermore, the data extracted in step S11 also includes the number of contacts, the relationship between the contacts and the applicant, verification of the contact’s mobile phone number, comparison of the location of the contact’s mobile phone number with the location of the applicant’s mobile phone number, loan history, repayment behavior, whether there is corresponding post-loan behavior after application, the number of credit applications and the number of institutions in the past month, and the number of applications and trends at different levels of institutions.

[0009] Furthermore, the number of positive samples and negative samples are both greater than 2000, and the ratio of positive samples to negative samples is 1:20.

[0010] Furthermore, in step S24, Python is used to call the lightgbm module for model training, the decision tree construction method is selected as "gbdt", the objective function is selected as "binary", and the evaluation index is selected as "auc".

[0011] Furthermore, parameter adjustment: use the Hyperopt module for automatic hyperparameter selection; when adjusting features, remove features that have no gain or negative gain on the model's predictive ability.

[0012] Furthermore, scoring verification was performed on the out-of-time validation samples, with the model score PSI not exceeding 0.25 and the changes in KS and AUC on the training and validation samples not exceeding 0.03.

[0013] In addition, the present invention also provides a consumer finance credit intermediary processing device, including the following: One or more processors or processing units that can perform data processing, feature calculation, and machine learning model training by executing computer executable instructions; One or more memories, including internal memory and external memory, wherein the internal memory is used to temporarily store the operation data in the CPU and the data exchanged with the external memory; the external memory is used to store data features and to execute program codes, and has corresponding storage space; The input / output device enables the user to interact with the computer device and visualize the output of the intermediary identification model scoring results and corresponding customer labels and other information.

[0014] From the above technology, it can be seen that the present invention has the following advantages: its resource consumption for data processing and model training is low, the intermediary risk scoring is simple to use, which lowers the usage threshold of the entire system and is very friendly to small and medium-sized financial institutions with limited relevant software and hardware resources. DETAILED DESCRIPTION

[0015] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0016] The consumer finance credit intermediary processing method of the present invention comprises the following steps: First, the features of the recognition model are processed. The specific processing method is as follows: 1) Data acquisition step: Summarize and extract data from the historical database on multiple dimensions of the entire process of customer loan applications, including: self-filled identity information (including but not limited to ID number, mobile phone number, etc.), self-filled contact information, customer credit information, historical lending behavior information (this institution and other institutions) and other public credit information. It should be noted that during the customer's loan operation, the system automatically obtains the current loan customer related information data and transfers it to the database for structured storage.

[0018] 2) Data processing steps: Through statistical transformations, such as sum, difference, count, maximum, minimum, average, ratio, variance, etc., the acquired data information is processed and aggregated to the customer dimension, corresponding to the input of the intermediary identification model.

[0019] The feature derivation methods involved in the present invention include: self-filled information mining (such as: number of contacts, relationship with the person, verification of contact mobile phone numbers, comparison of the location of the contact mobile phone number with the location of the applicant mobile phone number, etc.), credit information mining (such as: borrowing history, repayment behavior, whether there is corresponding post-loan behavior after application, etc.), historical borrowing behavior (such as: number of credit applications and number of institutions in the past month, number of applications and trends at different levels of institutions (banks, consumer finance, and mutual finance), etc.

[0020] 3) Sample selection step: extract customer samples that have been confirmed as intermediary applications (including intermediary applications and intermediary agents, the same below) and the corresponding features mentioned above as positive samples in the subsequent model training samples, and select samples of a certain proportion of the number of positive samples from the normal customer samples as negative samples in the subsequent model training samples. Merge the positive and negative samples to obtain sample set 1, and all the features contained in the sample set 1 are used as feature pool 1. It should be noted that the selected positive and negative samples need to reach a certain order of magnitude (theoretically, a certain number of samples are statistically significant. One case of this patent: either the positive / negative samples is not less than 2000. It can be adjusted according to actual conditions) and the ratio is kept within an appropriate range (a positive-negative sample ratio that is too high or too low will affect the machine learning effect. One case of this patent: the positive-negative sample ratio is 1:20, which can be adjusted according to actual conditions), and at the same time, remove abnormal samples (including statistically defined outliers and a small number of isolated cases based on business experience, to avoid the model fitting these small number of outliers resulting in a decrease in the prediction ability of normal samples) to ensure the effect of subsequent model training.

[0021] After feature processing, model training is performed. The training method is as follows: 1) Sample processing: Sample set 1 is divided into two parts, training samples and out-of-time verification samples, according to the time when the customer applies for a loan. This is used to verify the effect and stability of features and models in different time periods. It is necessary to ensure that the two parts of the samples do not overlap. The specific division nodes are determined according to business needs and actual data conditions. For example: customers who apply between January 1, 2023 and December 31, 2023 are used as training samples, and customers who apply between January 1, 2024 and May 31, 2024 are used as out-of-time verification samples. Calculate the IV value, missing rate, maximum single value ratio, PSI and other indicators of the features in feature pool 1 on the training samples and out-of-time verification samples respectively.

[0022] 2) Initial feature screening: Based on the above indicators, the features of feature pool 1 are screened to obtain feature pool 2. In this embodiment, features with IV < 0.01, missing rate > 95% or maximum single value ratio > 95% are eliminated.

[0023] 3) Feature fine screening: Based on the interpretability, risk monotonicity, and stability of each feature, the features of feature pool 2 are further screened into feature pool 3. In this embodiment, features whose risk performance is inconsistent with business experience, whose risk performance is inconsistent between training samples and out-of-time verification samples, or whose PSI>0.1 are eliminated.

[0024] 4) Model training: Input the training samples containing the features of feature pool 3 into the LightGBM binary classification model for training to obtain the trained model file. Use python to call the lightgbm module for model training, select "gbdt" for decision tree construction, select "binary" for objective function, and select "auc" for evaluation index. The output result of the model is a probability value between 0 and 1 (hereinafter referred to as the score). The higher the score, the greater the possibility that it is an intermediary. In this process, the model effect is optimized through model parameter adjustment and input feature adjustment (parameter adjustment: use the Hyperopt module for automatic hyperparameter selection, and then manually adjust the parameters based on expert experience. Feature adjustment: eliminate features that have no gain or negative gain for the model's predictive ability), and score verification on the out-of-time validation samples to ensure that the model's ability to distinguish has not decreased significantly (usually the model score PSI does not exceed 0.25 and the changes in KS and AUC on training and validation samples do not exceed 0.03. The latter is not a hard indicator, usually determined based on expert experience and actual business conditions, and can be appropriately relaxed); 5) Risk assessment: Take the latest customer application samples that contain the final usage characteristics of the model, use the aforementioned model to score, and conduct customer score distribution analysis based on the intermediary application labels confirmed by manual investigation. Determine the score threshold of the intermediary identification model based on the risk distribution of different scores. Divide customers into n grades according to the model score (n is determined according to business needs, usually n>=10), calculate the lift value of each grade, and then use the lowest score with a lift value greater than m (m is determined according to business needs, usually m>=2) as the risk threshold based on the actual business situation.

[0025] After the model training is completed, all credit application customers are scored by the intermediary identification model. For customers whose scoring results exceed the risk threshold, they are considered to have intermediary risks and their applications are transferred to manual processing. Through manual review, it is determined whether the customer is an intermediary or an agent of an intermediary. If not, they will be processed and enter the normal review process; if they are, they will be rejected. For customers within the threshold, they will enter the normal review process and finally get the review result.

[0026] Among them, KS (Kolmogorov-Smirnov, used to evaluate the model's ability to distinguish between good and bad customers, and calculate the maximum difference between the cumulative percentage of bad customers and the cumulative percentage of good customers. The value range of KS is 0 to 1. The closer the value is to 1, the better the model's fitting ability and prediction effect. Generally speaking, a KS value greater than 0.2 is considered a qualified prediction model), AUC (Area Under the ROC Curve, refers to the area under the ROC curve. The ROC curve is based on the false positive rate (False Positive Rate, FPR) as the horizontal axis and the true positive rate (True Positive Rate, FPR) as the horizontal axis. The curve reflects the proportion of correctly sorted good and bad user pairs. The AUC value ranges from 0.5 to 1, and the higher the AUC value, the better the predictive ability of the model) and the Lift value (a lift index used to measure the effectiveness of an association rule or prediction model. In this patent, it represents the ratio between the probability of an intermediary submission for a certain segment of customers and the probability of random occurrence. The higher the Lift value, the higher the probability of an intermediary submission for the segment is much higher than the probability of random occurrence, that is, the model has a high effectiveness).

[0027] In addition, the present invention also provides a device using the above method, comprising: 1) One or more processors or processing units that can perform operations such as data processing, feature calculation, and machine learning model training by executing computer executable instructions (program code); 2) One or more memories, including internal memory (RAM, ROM, Cache, etc.) and external memory (SSD, HHD, HDD, etc.). The former is used to temporarily store the calculation data in the CPU and the data exchanged with the external memory; the latter is used to store the data features involved in each embodiment and the program code for executing any method step in each embodiment, and has corresponding storage space; 3) Input / output devices, such as keyboards, pointing devices, displays, etc., which enable users to interact with the computer device and visualize the output of the intermediary recognition model scoring results and corresponding customer labels and other information.

[0028] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A consumer finance credit intermediary processing method The consumer finance credit intermediary processing method is characterized by: The steps include: S1. Process the features of the recognition model. The processing method is as follows: S11, data acquisition step: extracting data of multiple dimensions of the entire process of customer loan application from the historical database. When the customer is in the process of loan operation, the system automatically obtains the relevant information data of the current loan customer and transfers it to the database for structured storage; S12, data processing step: processing the acquired data information through statistical conversion and summarizing it into the customer dimension, corresponding to the input of the recognition model; S13, sample selection step: extract customer samples that have been confirmed to have been processed through intermediaries and the corresponding features of the above samples as positive samples in subsequent model training samples, and select samples with a certain proportion of the number of positive samples from normal customer samples as negative samples in subsequent model training samples, and remove abnormal samples; S14, merging the positive and negative samples to obtain sample set 1, wherein all features included in the sample set 1 are used as feature pool 1; S2. Train the recognition model. The training method is as follows: S21, sample processing: divide sample set 1 into two parts, training samples and out-of-time verification samples, according to the time when the customer applied for the loan, and calculate the IV value, missing rate, maximum single value ratio and PSI index of the features in feature pool 1 on the training samples and out-of-time verification samples respectively; S22. Initial feature screening: Based on the above indicators, the features of feature pool 1 are screened, and features with IV < 0.01, missing rate > 95%, or maximum single value ratio > 95% are eliminated to obtain feature pool 2; S23, feature fine screening: Eliminate the features whose risk performance is inconsistent with business experience, whose risk performance is inconsistent between training samples and out-of-time validation samples, or whose PSI>0.1, and obtain feature pool 3; S24, model training: input the training samples containing the features of feature pool 3 into the LightGBM binary classification model for training to obtain the trained model file; in this process, adjust the model effect by adjusting the model parameters and input features, and perform scoring verification on the out-of-time verification samples; S3. Take the latest customer application samples containing the final use characteristics of the model, score them using the aforementioned model, and perform customer score distribution analysis; determine the score threshold of the intermediary identification model based on the risk distribution of different scores; S4. All credit applicants are scored using the identification model. Customers whose scoring results exceed the risk threshold are considered to be at risk. Customers within the threshold enter the normal review process and ultimately obtain the review results.

2. The consumer finance credit intermediary processing method according to claim 1, characterized in that: The data extracted in step S11 includes ID number, mobile phone number, self-filled contact information, customer credit information, and historical lending behavior information.

3. The consumer finance credit intermediary processing method and device according to claim 2, characterized in that: The data extracted in step S11 also includes the number of contacts, the relationship between the contacts and the applicant, verification of the contact's mobile phone number, comparison of the location of the contact's mobile phone number with the location of the applicant's mobile phone number, loan history, repayment behavior, whether there is corresponding post-loan behavior after application, the number of credit applications and the number of institutions in the past month, and the number of applications and trends at different levels of institutions.

4. The consumer finance credit intermediary processing method according to claim 1, characterized in that: The number of positive samples and negative samples is greater than 2000, and the ratio of positive samples to negative samples is 1:

20.

5. The consumer finance credit intermediary processing method according to claim 1, characterized in that: In step S24, Python is used to call the lightgbm module for model training, and the decision tree construction method is selected as "gbdt", the objective function is selected as "binary", and the evaluation index is selected as "auc".

6. The consumer finance credit intermediary processing method according to claim 5, characterized in that: Parameter adjustment: Use the Hyperopt module for automatic hyperparameter selection; when adjusting features, remove features that have no or negative gains in model prediction ability.

7. The consumer finance credit intermediary processing method according to claim 1, characterized in that: Scoring validation was performed on the out-of-time validation samples, with the model score PSI not exceeding 0.25 and the changes in KS and AUC between training and validation samples not exceeding 0.

03.

8. A consumer finance credit intermediary processing device, using the method according to any one of claims 1 to 7, comprising the following One or more processors or processing units that can perform data processing, feature calculation, and machine learning model training by executing computer executable instructions; One or more memories, including internal memory and external memory, wherein the internal memory is used to temporarily store the operation data in the CPU and the data exchanged with the external memory; the external memory is used to store data features and to execute program codes, and has corresponding storage space; The input / output device enables the user to interact with the computer device and visualize the output of the recognition model scoring results and corresponding customer labels and other information.