A method and device for adjusting fraud scoring threshold

Abstract

The present invention provides a method and device for adjusting a fraud score threshold. The method utilizes an anti-fraud model to obtain a fraud score, and makes a risk decision based on the fraud score and the rejection threshold in the anti-fraud model. The method calculates the proportion of fraudulent customers in a sample to obtain the sample black density. The original assumption is that the baseline black density corresponding to the rejection threshold and the sample black density belong to the same distribution. If the original assumption is not true, then when the baseline black density corresponding to the rejection threshold is greater than the sample black density, the rejection threshold is modified based on the lower limit of the current score range, which is increased by one level. Otherwise, the rejection threshold is modified based on the lower limit of the current score range, which is decreased by one level, until the original assumption is true. In this solution, the sample black density within an adjustment period is calculated, and the original assumption is that the baseline black density and the sample black density belong to the same distribution. If the original assumption is not true, the rejection threshold is adjusted until it is true, thereby ensuring a stable rejection rate for the anti-fraud model and avoiding the rejection of a large number of legitimate customers.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for adjusting a fraud scoring threshold. Background Art

[0002] The pre-loan approval process for auto finance typically requires anti-fraud identification and credit rating. While both rely on algorithmic models, the actual audiences they serve are completely different. Generally, over 99.9% of loan applicants received by financial institutions are legitimate customers. By using statistical modeling to obtain basic customer information, credit data, and historical performance data, a scorecard is generated, ranking customers' credit ratings from high to low. Lower credit scores indicate a higher probability of default, allowing financial institutions to dynamically adjust credit score thresholds as needed. While the quality of legitimate loan applicants is consistently high, the rare fraudulent customer base, which accounts for less than 0.1% of the total, presents a significant sample imbalance that prevents credit rating models from identifying and resulting in low scores. Therefore, the core reason for the differences between anti-fraud identification technology and credit rating systems lies in the inherently different customer bases they serve.

[0003] After the anti-fraud model and credit rating model are launched, the Population Stability Index (PSI) is usually used to monitor whether the model score distribution and feature distribution are stable. When the model's indicators are abnormal, it means that the model's effectiveness is declining, and the model needs to be retrained and replaced in a timely manner. In the automotive finance scenario, the fraud ratio usually does not exceed 1:1000, the sources of fraudulent customers are wide, and there are strong deviations in the data characteristics, which leads to the instability of the anti-fraud model. Model overfitting often triggers certain feature values, causing the model to judge the customer's fraud risk as high and frequently make risky rejection decisions. Through post-loan AB testing, it can be found that the actual fraud probability of a large number of suspected fraudulent customers in a short period of time does not increase significantly, which indirectly shows that the anti-fraud model often overfits, and the disadvantage of monitoring based on the PSI indicator is lag. Therefore, models with overfitting risks may often intercept too many customers in a short period of time, causing business damage. Summary of the Invention

[0004] In view of this, an embodiment of the present invention provides a method and apparatus for adjusting a fraud scoring threshold, so as to ensure a stable rejection rate of an anti-fraud model and avoid rejecting a large number of legitimate customers.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of an embodiment of the present invention discloses a method for adjusting a fraud scoring threshold, the method comprising:

[0007] The anti-fraud model predicts a fraud score for incoming customers and makes risk decisions for them based on the fraud score and a rejection threshold set in the anti-fraud model. The rejection threshold is the lower limit of the current score range, which is one of multiple score ranges included in a preset score mapping table. Each score range corresponds to a baseline black density, which is the percentage of actual fraudulent customers among customers with a fraud score within the corresponding score range over a past period of time.

[0008] Whenever the number of customers who have completed the risk decision reaches a preset number, the proportion of customers determined to be fraudulent among the preset number of customers is calculated to obtain the sample black concentration;

[0009] Based on a hypothesis testing method, taking the reference black density and the sample black density corresponding to the rejection threshold as belonging to the same distribution as an original hypothesis, and judging whether the original hypothesis is established;

[0010] If the original hypothesis is not true, then when the benchmark black concentration corresponding to the rejection threshold is greater than the sample black concentration, the current scoring range is increased by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the increase; when the benchmark black concentration corresponding to the rejection threshold is less than the sample black concentration, the current scoring range is reduced by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the reduction; and the process returns to the hypothesis testing method, taking the benchmark black concentration corresponding to the rejection threshold and the sample black concentration as the original hypothesis that they belong to the same distribution, and judging whether the original hypothesis is true, until the original hypothesis is true.

[0011] Preferably, in the anti-fraud model, a corresponding general rejection threshold is set for each score range. Accordingly, the anti-fraud model predicts a fraud score for the incoming customer and makes a risk decision for the customer based on the fraud score and the rejection threshold set in the anti-fraud model, including:

[0012] The anti-fraud model predicts fraud scores for incoming customers;

[0013] determining whether the fraud score is greater than the rejection threshold;

[0014] If the fraud score is greater than the rejection threshold, querying the target score range corresponding to the fraud score from the multiple score ranges, and determining whether the normal rejection amount corresponding to the target score range is greater than 0;

[0015] If it is greater than 0, the customer is determined to be a fraudulent customer, and the normal rejection amount corresponding to the target score range is reduced by 1;

[0016] If it is not greater than 0, the customer is determined to be a normal customer;

[0017] If the fraud score is not greater than the rejection threshold, the customer is determined to be a normal customer.

[0018] Preferably, if, in the anti-fraud model, a corresponding priority rejection amount is set for each score range, and the priority rejection amount is half of the general rejection amount, then before determining whether the general rejection amount corresponding to the target score range is greater than 0, the method further includes:

[0019] If the customer is a mass-sending risk customer, determining whether the priority rejection amount corresponding to the target score range is greater than 0;

[0020] If it is greater than 0, the customer is determined to be a fraudulent customer, and the priority rejection amount corresponding to the target score range is reduced by 1;

[0021] If it is not greater than 0, the step of determining whether the ordinary rejection amount corresponding to the target score range is greater than 0 is performed.

[0022] Preferably, after modifying the rejection threshold, the method further comprises:

[0023] The ordinary rejection amount is adjusted to the lower limit of the score range corresponding to the modified rejection threshold, and the priority rejection amount is adjusted to half of the ordinary rejection amount.

[0024] Preferably, after determining that the customer is a fraudulent customer, the method further includes:

[0025] A fraud tag is added to the loan information of the customer.

[0026] A second aspect of an embodiment of the present invention discloses a device for adjusting a fraud scoring threshold, the device comprising:

[0027] A risk decision unit, wherein the anti-fraud model predicts a fraud score for an incoming customer and makes a risk decision for the customer based on the fraud score and a rejection threshold set in the anti-fraud model; the rejection threshold is the lower limit of the current score range, and the current score range is one of multiple score ranges included in a preset score mapping table. Each score range corresponds to a baseline black density, and the baseline black density is the proportion of actual fraudulent customers among customers with a fraud score within the corresponding score range in the past time period;

[0028] A calculation unit, configured to calculate the proportion of fraudulent customers among the preset number of customers whenever the number of customers who have completed the risk decision reaches a preset number, and obtain a sample black concentration;

[0029] a testing unit, configured to, based on a hypothesis testing method, take the fact that the reference black density and the sample black density corresponding to the rejection threshold belong to the same distribution as an original hypothesis, and determine whether the original hypothesis is established;

[0030] An adjustment unit is used to, if the original hypothesis is not established, increase the current scoring range by one level when the reference black concentration corresponding to the rejection threshold is greater than the sample black concentration, and modify the rejection threshold based on the lower limit of the current scoring range after the increase; when the reference black concentration corresponding to the rejection threshold is less than the sample black concentration, reduce the current scoring range by one level, and modify the rejection threshold based on the lower limit of the current scoring range after the decrease; return to the step of executing the hypothesis testing method, taking the reference black concentration corresponding to the rejection threshold and the sample black concentration belonging to the same distribution as the original hypothesis, and judging whether the original hypothesis is established, until the original hypothesis is established.

[0031] Preferably, in the anti-fraud model, a corresponding general rejection limit is set for each scoring range. Accordingly, the risk decision unit includes:

[0032] Scoring unit: The anti-fraud model predicts fraud scores for incoming customers;

[0033] a first judging unit, configured to judge whether the fraud score is greater than the rejection threshold;

[0034] a second judgment unit configured to, if the fraud score is greater than the rejection threshold, query and obtain a target score range corresponding to the fraud score from the multiple score ranges, and determine whether the normal rejection amount corresponding to the target score range is greater than 0;

[0035] A determination unit is configured to determine that the customer is a fraudulent customer if the score is greater than 0, and reduce the ordinary rejection amount corresponding to the target score range by 1; if the score is not greater than 0, determine that the customer is a normal customer; if the fraud score is not greater than the rejection threshold, determine that the customer is a normal customer.

[0036] Preferably, if in the anti-fraud model, a corresponding priority rejection amount is set for each scoring range, and the priority rejection amount is half of the ordinary rejection amount, then the second judgment unit is further configured to:

[0037] Before determining whether the ordinary rejection amount corresponding to the target score range is greater than 0, if the customer is a mass-sending risk customer, determine whether the priority rejection amount corresponding to the target score range is greater than 0; if it is greater than 0, determine that the customer is a fraudulent customer, and reduce the priority rejection amount corresponding to the target score range by 1; if it is not greater than 0, execute the step of determining whether the ordinary rejection amount corresponding to the target score range is greater than 0.

[0038] Preferably, the device further comprises:

[0039] The quota adjustment unit resets the common rejection quota corresponding to each scoring range after modifying the rejection threshold, and adjusts the priority rejection quota to half of the common rejection quota.

[0040] Preferably, the device further comprises:

[0041] The marking unit is used to add a fraud label to the loan information of the customer after determining that the customer is a fraud customer.

[0042] Based on the above-mentioned method and device for adjusting the fraud score threshold provided by the embodiment of the present invention, the anti-fraud model predicts a fraud score for the incoming customer, and makes a risk decision for the customer based on the fraud score and the rejection threshold set in the anti-fraud model; the rejection threshold is the lower limit of the current score range, and the current score range is one of the multiple score ranges contained in the preset score mapping table, each of the score ranges corresponds to a benchmark black density, and the benchmark black density is the proportion of real fraudulent customers among the customers whose fraud scores are within the corresponding score range in the past period; whenever the number of customers who have completed the risk decision reaches a preset number, the proportion of customers determined to be fraudulent in the preset number of customers is calculated to obtain the sample black density; based on the hypothesis testing method, the rejection threshold corresponding to the The original hypothesis that the benchmark black concentration and the sample black concentration belong to the same distribution is used to determine whether the original hypothesis is valid; if the original hypothesis is not valid, then when the benchmark black concentration corresponding to the rejection threshold is greater than the sample black concentration, the current scoring range is increased by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the increase; when the benchmark black concentration corresponding to the rejection threshold is less than the sample black concentration, the current scoring range is decreased by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the decrease; and the process returns to the step of determining whether the original hypothesis is valid based on the hypothesis testing method, and the benchmark black concentration and the sample black concentration corresponding to the rejection threshold belong to the same distribution as the original hypothesis, until the original hypothesis is valid. In this solution, a dynamic adjustment period is set, and the sample black concentration within the adjustment period is calculated. Based on the hypothesis testing method, the benchmark black concentration and the sample black concentration corresponding to the rejection threshold are used as the original hypothesis to belong to the same distribution. When the original hypothesis is not valid, the rejection threshold is adjusted until the original hypothesis is valid, so as to ensure the stability of the rejection rate of the anti-fraud model and avoid rejecting a large number of normal customers. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0044] Figure 1 This is a flowchart of a method for adjusting a fraud scoring threshold disclosed in an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of a scoring mapping table disclosed in an embodiment of the present invention;

[0046] Figure 3This is a flowchart of another method for adjusting a fraud scoring threshold disclosed in an embodiment of the present invention;

[0047] Figure 4 This is a structural diagram of a device for adjusting a fraud scoring threshold disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] In this application, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0050] like Figure 1 FIG. 1 is a flowchart of a method for adjusting a fraud scoring threshold disclosed in an embodiment of the present invention, which aims to solve the problem that a large number of customers hit the anti-fraud model in a short period of time, rejecting a large number of legitimate customers and causing business losses. The method mainly includes the following steps:

[0051] Step S101: The anti-fraud model predicts a fraud score for the incoming customer and makes a risk decision for the customer based on the fraud score and the rejection threshold set in the anti-fraud model.

[0052] Among them, the rejection threshold is the lower limit of the current scoring range. The current scoring range is one of the multiple scoring ranges contained in the preset scoring mapping table. Each scoring range corresponds to a benchmark black concentration. The benchmark black concentration is the proportion of real fraudulent customers among customers whose fraud scores are within the corresponding scoring range in the past period.

[0053] Among them, real fraud customers refer to customers who are confirmed to be fraud customers and are labeled as fraudulent.

[0054] In step S101, a customer applies for a car loan using an app or mini-program on a mobile device. The application communicates with a server over the network. The server collects the customer's information and processes it into features for the anti-fraud model. The anti-fraud model then provides feedback on whether the customer's application has been approved. The anti-fraud model can be any well-performing classification or anomaly detection model, such as a random forest, gradient boosted decision tree, or support vector machine. The anti-fraud model generates a fraud score, which increases with the probability of fraud.

[0055] like Figure 2 , which is a schematic diagram of a scoring mapping table disclosed in an embodiment of the present invention.

[0056] Exemplarily, the process of establishing a scoring mapping table is as follows:

[0057] We obtained 10,000 historical customer fraud risk scores and fraud labels from the past year and sorted them into tiers based on their fraud scores. The score range is [0, 200], with 10 tiers of 20 points. The number of fraudulent customers / customers in each tier represents the baseline black density for that group. For example, if there are 100 samples with scores in the range [120, 140) and 5 fraudulent customers, the corresponding baseline black density is 0.05. If there are 30 samples with scores in the range [140, 160) and 9 fraudulent customers, the corresponding baseline black density is 0.3. If there are 30 samples with scores in the range [160, 180) and 12 fraudulent customers, the corresponding baseline black density is 0.4. If there are 10 samples with scores in the range [180, 200] and 5 fraudulent customers, the corresponding baseline black density is 0.5.

[0058] It can be understood that the fraud label is a truly confirmed fraudulent customer, and the benchmark black density = the number of truly fraudulent customers / the number of historical customers.

[0059] In an embodiment of the present invention, the lower limit of the scoring range is selected to set the rejection threshold in the anti-fraud model. When the anti-fraud model is just launched and the rejection threshold has not yet been dynamically adjusted, the lower limit of the scoring range in which the baseline black concentration is the same as the target black concentration is selected as the rejection threshold based on the target black concentration set by the business side.

[0060] Assuming the target black density is 0.3, looking at the score mapping table, we get the score range for a base black density of 0.3: [140, 160). Therefore, we set 140 as the rejection threshold. If the target black density falls between two score ranges, for example, 0.35, we use the lower limit of the score range corresponding to the lower base black density as the rejection threshold, resulting in the corresponding group [140, 160). We also set 140 as the rejection threshold.

[0061] It should be noted that any expert rules, including risk decision rules, can be pre-set in the anti-fraud model. After the anti-fraud model obtains the customer's fraud score, it uses the risk decision rules to determine whether the customer is a fraudulent customer or a group fraud customer, and outputs the corresponding judgment result. For the specific risk decision process, please refer to Figure 3 Corresponding embodiments of the present invention.

[0062] Step S102: Whenever the number of customers who have completed risk decisions reaches a preset number, the proportion of customers determined to be fraudulent among the preset number of customers is calculated to obtain the sample black density.

[0063] Here, customers with a fraud score greater than the rejection threshold are considered fraudulent. Sample black density = number of identified fraudulent customers / number of sample customers.

[0064] In step S102, a preset number is used as a dynamic adjustment period. The preset number may be 5000, which means that after each risk decision is made on 5000 customers using the anti-fraud model, the sample black concentration is calculated, and based on the sample black concentration, it is determined whether the rejection threshold in the anti-fraud model needs to be adjusted.

[0065] In one embodiment, time can also be used as a period, such as a day, a week, etc., to calculate the proportion of customers determined to be fraudulent customers among the customers determined by the anti-fraud model risk decision within the time period to obtain the sample black concentration.

[0066] Step S103: Based on the hypothesis testing method, the original hypothesis is that the reference black density and the sample black density corresponding to the rejection threshold belong to the same distribution, and whether the original hypothesis is established is determined. If so, step S104 is executed; if not, step S105 is executed.

[0067] In the specific implementation of step S103, a hypothesis test is used to compare the significance level α (0.05) corresponding to the binomial distribution under the current rejection threshold (samples greater than the rejection threshold are 1, otherwise 0) using a two-sided test. If the statistic p is less than α, the H0 hypothesis of identical distribution is considered to be false and the threshold needs to be adjusted.

[0068] It should be noted that hypothesis testing: first set an original hypothesis (H0), which means that the baseline black concentration and the sample black concentration are the same, and it is necessary to test whether this original hypothesis is true.

[0069] Binomial distribution: In this test, the sample data will be classified according to the current rejection threshold: samples greater than the threshold are recorded as 1, and samples less than or equal to the threshold are recorded as 0. The samples formed in this way can be described by the binomial distribution.

[0070] Significance level α: The significance level set here is 0.05, which means that in 5% of cases, the original hypothesis may be wrongly rejected.

[0071] Two-sided test: means that you focus on the difference between the reference black density and the sample black density, whether the reference black density is higher or lower than the sample black density.

[0072] p-value: If the calculated p-value is less than the significance level α (0.05), it means that there is sufficient evidence to reject the original hypothesis. This indicates that there is a significant difference between the baseline black concentration and the sample black concentration.

[0073] Step S104: Determine not to adjust the rejection threshold.

[0074] Step S105: When the reference black concentration corresponding to the rejection threshold is greater than the sample black concentration, the current scoring range is increased by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the increase; when the reference black concentration corresponding to the rejection threshold is less than the sample black concentration, the current scoring range is reduced by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the decrease; and the process returns to step S103 until the original hypothesis is established.

[0075] In step S105 , if the reference black density is greater than the sample black density, it indicates that the reference black density corresponding to the current rejection threshold is too low and the rejection threshold needs to be increased; otherwise, it indicates that the sample black density is too high and the rejection threshold needs to be lowered.

[0076] Assume that the current sample black density is 0.19 and 140 points is the rejection threshold. Adjust the current score range by one level, and use the lower limit of the score range after the increase (160 points) as the rejection threshold. Then perform the hypothesis test again. If H0 is true, the adjustment is terminated; otherwise, the adjustment continues. Similarly, if the current sample black density is 0.41 and 160 points is the rejection threshold, lower the rejection threshold to 140 points. If H0 is true, the adjustment is terminated; otherwise, the adjustment continues.

[0077] For example, in actual application scenarios, after the anti-fraud model is launched, the model rejection rate may deviate, and the rejection rate needs to be dynamically adjusted to ensure business stability. For example, the historical benchmark [140,160) segment black concentration is 0.3, the [160,180) segment black concentration is 0.4, and the [180,200] segment black concentration is 0.5. If the newly launched anti-fraud model is abnormal, its ranking is poor after running for a period of time. For example, the [140,160] segment black concentration is 0.4, the [160,180) segment black concentration is 0.3, and the [180,200] segment black concentration is 0.2, indicating that the model has a large deviation and needs to be updated to ensure its ranking ability. Before updating the model, the model rejection threshold is adjusted through the fraud scoring threshold adjustment method disclosed in the embodiment of the present invention to ensure the stability of the overall rejection rate. In addition, when the newly launched anti-fraud model has run for a period of time and its recognition ability declines, it enters an underfitting state and the rejection rate decreases, the rejection threshold is automatically lowered through the fraud scoring threshold adjustment method disclosed in the embodiment of the present invention, so that the model can ensure the stability of the overall rejection rate.

[0078] For example, in a real-world application scenario, when a dealership, due to inventory pressure or other factors, needs to promote a particular car model, they may attract a large number of customers from other channels in addition to their original channel customers within a short period of time, leading to a surge in applications. However, the overall risk profile of these new channel customers differs significantly from that of existing channel customers, resulting in a high number of misjudgments in the anti-fraud model. Rules that can increase the hit rate of risk profile rules include: dealers increasing sales in a short period of time, highly consistent financing plans, and unusual customer locations and sources. Specific input variables are developed to target these characteristics, and the model will frequently reject customers that match these rule sets. When a dealership triggers a large number of rejections within a short period of time, and the sample black density exceeds the deviation from the baseline black density, the fraud scoring threshold adjustment method disclosed in this embodiment of the present invention raises the rejection threshold to match the baseline black density, thereby ensuring that legitimate customers are approved. For example, a suspected fraudulent application may have the following common features: same salesperson, same car model, same financing plan, and out-of-town purchase. If the model outputs a fraud score of 150 for this type of customer, and the current rejection threshold is set at 140, the application will be rejected. If a dealership in a certain area organizes promotions and marketing campaigns to clear inventory, a large number of customers may apply in a short period of time. In this case, making risk decisions based on the rejection threshold would be unreasonable. By dynamically adjusting the rejection threshold to 160 points, these customers will not be identified as high-risk customers, thus preventing a large number of normal customers from being rejected in a short period of time, which could affect business development.

[0079] Based on the above-mentioned method for adjusting the fraud scoring threshold disclosed in the embodiment of the present invention, in this solution, a dynamic adjustment period is set, and the sample black concentration within the adjustment period is calculated. Based on the hypothesis testing method, the baseline black concentration corresponding to the rejection threshold and the sample black concentration belong to the same distribution as the original hypothesis. When the original hypothesis is not true, the rejection threshold is adjusted until the original hypothesis is true, so as to achieve the purpose of ensuring the stability of the rejection rate of the anti-fraud model and avoiding the rejection of a large number of normal customers.

[0080] To explain in detail the specific implementation process of step S101 in the embodiment of the present invention, Figure 3 , which is a flow chart of another method for adjusting a fraud scoring threshold value disclosed in an embodiment of the present invention.

[0081] The specific implementation process of step S101 includes the following steps:

[0082] Step S301: The anti-fraud model predicts the fraud score of the incoming customer.

[0083] Step S302: Determine whether the fraud score is greater than the rejection threshold. If so, proceed to step S304; if not, proceed to step S303.

[0084] Step S303: Determine whether the customer is a normal customer.

[0085] Step S304: query and obtain a target score range corresponding to the fraud score from multiple score ranges.

[0086] Step S305: Determine whether the customer is a mass-sending risk customer. If so, proceed to step S306; if not, proceed to step S308.

[0087] Step S306: Determine whether the priority rejection amount corresponding to the target score range is greater than 0. If so, execute step S307; if not, execute step S308.

[0088] Step S307: Determine that the customer is a fraudulent customer, and reduce the priority rejection amount corresponding to the target score range by 1.

[0089] Step S308: Determine whether the normal rejection amount corresponding to the target score range is greater than 0. If so, execute step S309; ​​if not, execute step S303.

[0090] Step S309: Determine that the customer is a fraudulent customer, and reduce the normal rejection amount corresponding to the target score range by 1.

[0091] In an embodiment of the present invention, a normal queue and a priority queue are provided in the anti-fraud model. Customers who are not identified as mass-messaging risk customers by the identification rules for mass-messaging risk customers preset in the anti-fraud model enter the normal queue, and are determined whether they are fraudulent customers based on the customer's fraud score and the margin of the normal rejection amount set for each score range in the anti-fraud model.

[0092] For customers identified as mass-messing risk customers by the identification rules for mass-messing risk customers preset in the anti-fraud model, they will first enter the priority queue. Based on the customer's fraud score and the margin of the priority rejection amount set for each score range in the anti-fraud model, it is determined whether the customer is a fraudulent customer. If the margin of the priority rejection amount is 0, the customer will enter the ordinary queue.

[0093] The ordinary rejection amount corresponding to each scoring range is set according to the benchmark black density corresponding to the scoring range. For example, if there are 30 people in the scoring range [140,160) and the benchmark black density is 0.3, then the ordinary rejection amount of the scoring range is 30*0.3=9 people; if there are 30 people in the scoring range [160,180) and the benchmark black density is 0.4, then the ordinary rejection amount of the scoring range is 30*0.4=12 people; if there are 10 people in the scoring range [180,200] and the benchmark black density is 0.5, then the ordinary rejection amount of the scoring range is 10*0.5=5 people.

[0094] Preferably, the priority rejection amount corresponding to any scoring range is half of the ordinary rejection amount corresponding to the scoring range. If the ordinary rejection amount is an odd number, the larger value is rounded up. For example, if the ordinary rejection amount is 5, then the priority rejection amount is 3.

[0095] In one embodiment, after the rejection threshold is modified, the normal rejection amount corresponding to each scoring range is reset, and the priority rejection amount is adjusted to half of the normal rejection amount.

[0096] In an embodiment of the present invention, when the customer fraud score is higher than the rejection threshold and there is a remaining rejection amount in the score range corresponding to the customer fraud score, the request may be rejected, otherwise it may be released.

[0097] For example, if a customer enters the normal queue and the rejection threshold is 140 points, and their fraud score is greater than 140, the system will determine which score range the customer falls into. If the normal rejection quota for that score range is available, the customer is rejected (i.e., determined to be a fraudulent customer); if not, the customer is allowed through (i.e., determined to be a normal customer). If the customer enters the priority queue, the priority queue quota is used first to execute the rejection. Once the priority queue quota is exhausted, the customer enters the normal queue.

[0098] For example, in real-world applications, when addressing suspected mass fraud, it's necessary to temporarily lower the fraud threshold to ensure that currently suspected fraudulent customers aren't approved. A recent surge in fraud cases for a particular car model suggests a risk control vulnerability that has been discovered by the black market, often leading to a concentrated attack within a short period of time, triggering a mass outbreak. Common risk characteristics analyzed include: out-of-region vehicle purchases (inconsistencies between registration location, purchase location, and permanent residence), excessive recent credit history (multiple institutional inquiries or loan application records), abnormal sales data (dealers ramping up sales within a short period of time, no promotions, and the same salesperson), and obvious signs of uniform packaging (highly consistent financial plans and customer information). Specific modeling variables are developed to target these characteristics and reject customers who meet the above rule set. If a customer is identified as a mass fraud risk, the priority rejection limit is used for rejection. Once the priority rejection limit is exhausted, the customer enters the standard queue. If there's still a standard rejection limit within the corresponding fraud score range, the customer remains rejected.

[0099] Based on the method for adjusting the fraud score threshold disclosed in the above-mentioned embodiment of the present invention, in this solution, when a customer's fraud score is higher than the rejection threshold and the customer is not a mass-spam risk customer, if there is a remaining normal rejection amount within the corresponding score range, the customer can be rejected, otherwise they can be released. If the customer is determined to be a mass-spam risk customer, they will be placed in the priority queue and use the priority rejection amount within the score range. If there is no priority rejection amount, they will be placed in the normal queue. While ensuring that customers with relatively high fraud risks are rejected first, it prevents mass-spam risk events from causing the normal rejection amount to be consumed, thereby preventing ordinary suspected fraud customers from being released.

[0100] Based on the above-mentioned method for adjusting the fraud scoring threshold disclosed in the embodiment of the present invention, Figure 4 , which is a structural diagram of a device for adjusting a fraud scoring threshold disclosed in an embodiment of the present invention, includes: a risk decision unit 401 , a calculation unit 402 , a verification unit 403 and an adjustment unit 404 .

[0101] The risk decision unit 401 is used to predict the fraud score of the incoming customer based on the anti-fraud model, and make a risk decision on the customer based on the fraud score and the rejection threshold set in the anti-fraud model; the rejection threshold is the lower limit of the current scoring range, and the current scoring range is one of the multiple scoring ranges contained in the preset scoring mapping table. Each scoring range corresponds to a benchmark black density, and the benchmark black density is the proportion of real fraudulent customers among customers whose fraud scores are within the corresponding scoring range in the past time period.

[0102] The calculation unit 402 is configured to calculate the proportion of fraudulent customers among the preset number of customers whenever the number of customers who have completed risk decisions reaches a preset number, and obtain a sample black density.

[0103] The testing unit 403 is configured to determine whether the original hypothesis is established by taking the fact that the reference black density and the sample black density corresponding to the rejection threshold belong to the same distribution as the original hypothesis based on a hypothesis testing method.

[0104] The adjustment unit 404 is used to, if the original hypothesis is not established, increase the current score range by one level when the reference black concentration corresponding to the rejection threshold is greater than the sample black concentration, and modify the rejection threshold based on the lower limit of the current score range after the increase; when the reference black concentration corresponding to the rejection threshold is less than the sample black concentration, reduce the current score range by one level, and modify the rejection threshold based on the lower limit of the current score range after the decrease; return to execute the hypothesis testing method, take the reference black concentration corresponding to the rejection threshold and the sample black concentration as belonging to the same distribution as the original hypothesis, and judge whether the original hypothesis is established, until the original hypothesis is established.

[0105] In one embodiment, in the anti-fraud model, a corresponding general rejection limit is set for each scoring range. Accordingly, the risk decision unit includes:

[0106] Scoring unit: The anti-fraud model predicts fraud scores for incoming customers.

[0107] The first judgment unit is configured to judge whether the fraud score is greater than a rejection threshold.

[0108] The second judgment unit is configured to query and obtain a target score range corresponding to the fraud score from multiple score ranges if the fraud score is greater than the rejection threshold, and to determine whether the normal rejection amount corresponding to the target score range is greater than 0.

[0109] The judgment unit is used to judge the customer as a fraudulent customer if the score is greater than 0 and reduce the normal rejection amount corresponding to the target score range by 1; if it is not greater than 0, the customer is judged as a normal customer; if the fraud score is not greater than the rejection threshold, the customer is judged as a normal customer.

[0110] In one embodiment, if a corresponding priority rejection amount is set for each scoring range in the anti-fraud model, and the priority rejection amount is half of the normal rejection amount, the second judgment unit is further configured to:

[0111] Before determining whether the ordinary rejection amount corresponding to the target score range is greater than 0, if the customer is a mass-messaging risk customer, determine whether the priority rejection amount corresponding to the target score range is greater than 0; if it is greater than 0, the customer is determined to be a fraudulent customer, and the priority rejection amount corresponding to the target score range is reduced by 1; if it is not greater than 0, execute the step of determining whether the ordinary rejection amount corresponding to the target score range is greater than 0.

[0112] In one embodiment, the adjusting device further comprises:

[0113] The quota adjustment unit resets the normal rejection quota for each scoring range and adjusts the priority rejection quota to half of the normal rejection quota after modifying the rejection threshold.

[0114] In one embodiment, the adjusting device further comprises:

[0115] The marking unit is used to add a fraud label to the customer's loan information after determining that the customer is a fraud customer.

[0116] Based on the above-mentioned method for adjusting the fraud scoring threshold disclosed in the embodiment of the present invention, in this solution, a dynamic adjustment period is set, and the sample black concentration within the adjustment period is calculated. Based on the hypothesis testing method, the baseline black concentration corresponding to the rejection threshold and the sample black concentration belong to the same distribution as the original hypothesis. When the original hypothesis is not true, the rejection threshold is adjusted until the original hypothesis is true, so as to achieve the purpose of ensuring the stability of the rejection rate of the anti-fraud model and avoiding the rejection of a large number of normal customers.

[0117] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0118] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0119] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for adjusting a fraud scoring threshold, characterized in that: The method comprises: The anti-fraud model predicts a fraud score for incoming customers and makes risk decisions for them based on the fraud score and a rejection threshold set in the anti-fraud model. The rejection threshold is the lower limit of the current score range, which is one of multiple score ranges included in a preset score mapping table. Each score range corresponds to a benchmark black density, which is the percentage of actual fraudulent customers among customers with fraud scores within the corresponding score range in the past period. When the anti-fraud model is just launched and the rejection threshold has not yet been dynamically adjusted, the lower limit of the score range where the benchmark black density is the same as the target black density is selected as the rejection threshold based on the target black density set by the business side. Whenever the number of customers who have completed the risk decision reaches a preset number, the proportion of customers determined to be fraudulent among the preset number of customers is calculated to obtain the sample black concentration; Based on a hypothesis testing method, taking the reference black density and the sample black density corresponding to the rejection threshold as belonging to the same distribution as an original hypothesis, and judging whether the original hypothesis is valid; If the original hypothesis is not true, then when the benchmark black concentration corresponding to the rejection threshold is greater than the sample black concentration, the current scoring range is increased by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the increase; when the benchmark black concentration corresponding to the rejection threshold is less than the sample black concentration, the current scoring range is reduced by one level, and the rejection threshold is modified based on the lower limit of the current scoring range after the reduction; and the process returns to the hypothesis testing method, taking the benchmark black concentration corresponding to the rejection threshold and the sample black concentration as the original hypothesis that they belong to the same distribution, and judging whether the original hypothesis is true, until the original hypothesis is true.

2. The method according to claim 1, characterized in that In the anti-fraud model, a corresponding general rejection threshold is set for each score range. Accordingly, the anti-fraud model predicts a fraud score for the incoming customer and makes a risk decision for the customer based on the fraud score and the rejection threshold set in the anti-fraud model, including: The anti-fraud model predicts fraud scores for incoming customers; determining whether the fraud score is greater than the rejection threshold; If the fraud score is greater than the rejection threshold, querying the target score range corresponding to the fraud score from the multiple score ranges, and determining whether the normal rejection amount corresponding to the target score range is greater than 0; If it is greater than 0, the customer is determined to be a fraudulent customer, and the normal rejection amount corresponding to the target score range is reduced by 1; If it is not greater than 0, the customer is determined to be a normal customer; If the fraud score is not greater than the rejection threshold, the customer is determined to be a normal customer.

3. The method according to claim 2, characterized in that If, in the anti-fraud model, a corresponding priority rejection amount is set for each scoring range, and the priority rejection amount is half of the general rejection amount, then before determining whether the general rejection amount corresponding to the target scoring range is greater than 0, the method further includes: If the customer is a mass-sending risk customer, determining whether the priority rejection amount corresponding to the target score range is greater than 0; If it is greater than 0, the customer is determined to be a fraudulent customer, and the priority rejection amount corresponding to the target score range is reduced by 1; If it is not greater than 0, the step of determining whether the ordinary rejection amount corresponding to the target score range is greater than 0 is performed.

4. The method according to claim 3, characterized in that After modifying the rejection threshold, the method further comprises: The ordinary rejection amount corresponding to each scoring range is reset, and the priority rejection amount is adjusted to half of the ordinary rejection amount.

5. The method according to claim 3, characterized in that After determining that the customer is a fraudulent customer, the method further includes: Add a fraud tag to the customer's loan information.

6. A device for adjusting a fraud scoring threshold, characterized in that: The device comprises: The risk decision unit uses an anti-fraud model to predict a fraud score for an incoming customer and makes a risk decision for the customer based on the fraud score and a rejection threshold set in the anti-fraud model. The rejection threshold is the lower limit of the current score range, which is one of multiple score ranges included in a preset score mapping table. Each score range corresponds to a benchmark black density, which is the proportion of actual fraudulent customers among customers with fraud scores within the corresponding score range in the past period. When the anti-fraud model is just launched and the rejection threshold has not yet been dynamically adjusted, the lower limit of the score range where the benchmark black density is the same as the target black density is selected as the rejection threshold based on the target black density set by the business side. A calculation unit, configured to calculate the proportion of fraudulent customers among the preset number of customers whenever the number of customers who have completed the risk decision reaches a preset number, and obtain a sample black concentration; a testing unit, configured to, based on a hypothesis testing method, take the fact that the reference black density and the sample black density corresponding to the rejection threshold belong to the same distribution as an original hypothesis, and determine whether the original hypothesis is established; An adjustment unit is used to, if the original hypothesis is not established, increase the current scoring range by one level when the reference black concentration corresponding to the rejection threshold is greater than the sample black concentration, and modify the rejection threshold based on the lower limit of the current scoring range after the increase; when the reference black concentration corresponding to the rejection threshold is less than the sample black concentration, reduce the current scoring range by one level, and modify the rejection threshold based on the lower limit of the current scoring range after the decrease; return to the step of executing the hypothesis testing method, taking the reference black concentration corresponding to the rejection threshold and the sample black concentration belonging to the same distribution as the original hypothesis, and judging whether the original hypothesis is established, until the original hypothesis is established.

7. The device according to claim 6, characterized in that In the anti-fraud model, a corresponding general rejection limit is set for each scoring range. Accordingly, the risk decision unit includes: Scoring unit: The anti-fraud model predicts fraud scores for incoming customers; a first judging unit, configured to judge whether the fraud score is greater than the rejection threshold; a second judgment unit configured to, if the fraud score is greater than the rejection threshold, query and obtain a target score range corresponding to the fraud score from the multiple score ranges, and determine whether the normal rejection amount corresponding to the target score range is greater than 0; A determination unit is configured to determine that the customer is a fraudulent customer if the score is greater than 0, and reduce the ordinary rejection amount corresponding to the target score range by 1; if the score is not greater than 0, determine that the customer is a normal customer; if the fraud score is not greater than the rejection threshold, determine that the customer is a normal customer.

8. The device according to claim 7, characterized in that If, in the anti-fraud model, a corresponding priority rejection amount is set for each scoring range, and the priority rejection amount is half of the ordinary rejection amount, the second judgment unit is further configured to: Before determining whether the ordinary rejection amount corresponding to the target score range is greater than 0, if the customer is a mass-sending risk customer, determining whether the priority rejection amount corresponding to the target score range is greater than 0; if it is greater than 0, determining that the customer is a fraudulent customer, and reducing the priority rejection amount corresponding to the target score range by 1; If it is not greater than 0, the step of determining whether the ordinary rejection amount corresponding to the target score range is greater than 0 is performed.

9. The device according to claim 8, characterized in that The device further comprises: The quota adjustment unit resets the common rejection quota corresponding to each scoring range after modifying the rejection threshold, and adjusts the priority rejection quota to half of the common rejection quota.

10. The device according to claim 8, characterized in that The device further comprises: The marking unit is used to add a fraud label to the loan information of the customer after determining that the customer is a fraud customer.

Images