Intelligent collection case allocation method based on gain model

Through the intelligent allocation method of collection cases based on the gain model, combined with text tendency analysis and differential response model, the collection method and personnel matching are optimized, the problem of unoptimized resource allocation in the existing technology is solved, and efficient collection case allocation and Resource utilization.

CN119090207APending Publication Date: 2024-12-06CHINA BOHAI BANK CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411136716.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing collection case allocation methods ignore the differences in the fields and professional abilities of different collection officers, and fail to effectively consider the cost and response improvement relationship of collection methods, resulting in unoptimized resource allocation.

Method used

Using the intelligent allocation method of collection cases based on gain model, the account data is processed through text tendency analysis and feature engineering, and a differential response model is constructed to evaluate the gain effect of collection methods, and the optimal matching of cases and personnel portraits is combined, and the high allocation is given priority. Value cases to the right personnel.

Benefits of technology

The efficiency of collection cases and resource utilization are improved, the return on investment is maximized, the optimal matching between cases and personnel is achieved through greedy algorithms, and the overall resource allocation efficiency is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090207A_ABST
    Figure CN119090207A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent collection case allocation method based on a gain model. The method comprises the following steps: step 1, a case to be allocated is urged; step 2, account data in a database is processed through text tendency analysis and feature engineering; 3, building a gain model: obtaining a gain effect based on a differential response model of the two models; 4, distributing collection channels based on the result of the gain model; step 5, redistributing the cases based on the case / personnel portraits; and step 6, matching the operators with the to-be-allocated cases to complete case allocation. According to the method, the matching degree of the case and the personnel is calculated by combining the case repayment factor tag and the personnel skilled collection tag, the case traversal distribution is carried out according to the case value through the greedy algorithm, the to-be-distributed case is recommended to the most suitable collection personnel, and the optimal matching of the case and the personnel is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an intelligent allocation method for debt collection cases, and in particular to an intelligent allocation method for debt collection cases based on a gain model. Background Art

[0002] In the work of credit risk management, it is usually divided into pre-loan management, loan management and post-loan management according to the loan process. As the last link of credit risk management, post-loan management plays a vital role in the whole process management of credit risk. However, due to the complexity and diversity of the issues involved in post-loan management, the strong operation process, the reliance on human resources and the lack of attention to post-loan in the market expansion stage, post-loan management has long focused on operational management such as resource introduction and personnel management. However, with the slowdown in the growth rate of the credit market and the increasing standardization of post-loan management by regulatory agencies, the competition in the stock market has become a battlefield for various credit financial institutions, and post-loan collection management has gradually shifted to intelligent, digital and scientific. Relying on the development of intelligent risk control and artificial intelligence technology, post-loan collection management has gradually transformed from labor-intensive to digital. Related technologies such as text analysis, intelligent voice, knowledge graphs, collection score cards, etc. have also been gradually applied in the field of post-loan collection, saving a lot of human resources and improving the efficiency of post-loan collection management.

[0003] At present, there are two mainstream ideas in the allocation of collection cases: one is to evaluate the difficulty of collection cases through collection scoring cards or machine learning models, and allocate cases based on the matching degree of the two in combination with the collection ability of the collectors. This method pays too much attention to the collection ability reflected by historical performance and ignores the areas of expertise of different collectors. The other is to build a portrait label system through data analysis and business experience, respectively determine the repayment factor labels of collection cases and the collection expertise labels of collectors, and then match the characteristic labels of collection cases and collection specialists, and allocate cases based on the matching degree of the labels of both parties. This method is prone to excessive label concentration and ignores the differences in professional capabilities of collectors.

[0004] As the main link of post-loan management, post-loan collection completes the preservation and collection of overdue assets of financial institutions through collection methods such as SMS, letters, intelligent voice, manual telephone collection, door-to-door visits, arbitration, and litigation. However, since different collection methods have different costs and effects, how to determine the collection method of a case has always been a problem that has been continuously explored in the field of post-loan collection. The current mainstream collection case allocation plan is to assess the risk of customers by predicting customer response (usually return rate or repayment rate), and to build a waterfall collection case allocation plan based on the cost and resource situation of the collection method. This method only considers the customer's response and does not consider the improvement of different collection methods, so it ignores the relationship between cost increase and response improvement. For example, the repayment rate of case A without collection is 20%, and the repayment rate of manual telephone collection is 60%; the repayment rate of case B without collection is 59%, and the repayment rate of manual telephone collection is 60%. If only one customer can be selected for manual telephone collection, considering the customer's response, the effect of manual telephone collection on case A and case B is the same, so it is enough to randomly select one. However, if the gain is taken into consideration, with the same cost, the return on investment (ROI) of case A is higher than that of case B, so case A should be selected.

[0005] In actual business, when the collection method is determined to be manual telephone collection and door-to-door visits, it is more dependent on manual operations. Although the SOP process can narrow the differences between operators, as the difficulty of the case increases, the complexity of the entire communication process also gradually increases. The business experience and communication preferences of different operators will also have a greater impact on the results. When assigning cases to personnel, the mainstream idea is to assign cases by matching the difficulty of the case with the ability of the operator, or by matching the case repayment factor label with the collection expert label of the person. Both ideas have defects.

[0006] Finally, when evaluating customer response, structured data such as application data, credit data, and behavioral data are usually the main features, and unstructured data such as collection voice records are not used. Some speech recognition and text analysis technologies are mainly used in intelligent quality inspection and have not been fully applied to the evaluation of collection cases to dig out more customer overdue information. Summary of the invention

[0007] In order to solve the deficiencies of the above-mentioned technologies, the present invention provides a method for intelligent allocation of collection cases based on a gain model.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for intelligent allocation of collection cases based on a gain model, comprising the following steps:

[0009] Step 1: Cases to be assigned are entered into the collection database: Eligible risk accounts enter the collection database and wait for assignment;

[0010] Step 2: Process the account data in the database through text tendency analysis and feature engineering;

[0011] Step 3: Construct a gain model: Based on the differential response model of the two models, obtain the gain effect;

[0012] Step 4: Allocate collection channels based on the results of the gain model;

[0013] Step 5: Reassign cases based on case / personnel portraits: After completing the initial assignment in step 4, cases need to be reassigned based on the operator's portrait;

[0014] Step 6: Match the operators with the cases to be assigned to complete case allocation: prioritize high-value cases and assign them to appropriate operators.

[0015] Preferably, in step 2, text tendency analysis refers to using speech recognition to convert the call process into text, and then using text tendency analysis to identify the customer's current communication intention, repayment ability and repayment willingness, and adding the recognition results to the construction features of the gain model.

[0016] Preferably, in step three, one model G is used to estimate the user's response in the presence of intervention, and another model G' is used to learn the user's response in the absence of intervention, and then the outputs of the two models are subtracted to obtain a gain effect.

[0017] Preferably, in step 4, the costs of different collection methods are considered as weights, the weights are multiplied by the gain effect values ​​and then sorted, and the collection method corresponding to the maximum weighted gain effect is selected, that is, the collection method with the maximum return on investment (ROI) is selected. The calculation formula for ROI is:

[0018] ROI = benefit effect / collection cost.

[0019] Preferably, step five comprises the following steps:

[0020] First, the repayment factor profile of the collection case is made based on the case information;

[0021] Secondly, we will create a collection profiling for the operators based on the repayment factor profile records of historical collection cases;

[0022] Finally, the optimal allocation of cases is achieved by matching the repayment factor profile with the collection expertise profile.

[0023] Preferably, in step 5, when profiling repayment factors, the repayment factors of the case should be profiled in multiple dimensions based on customer information, account information, collection tendency information, and operator annotations, and the m labels with the highest impact on repayment should be retained, denoted as A i ,i∈1,…,m;

[0024] Then, based on the operator's historical performance, the operator's response rate for each repayment factor label is calculated, and the n labels with the highest response rate are selected as the collection expert labels, denoted as B j ,j∈1,…,n, the response rate corresponding to the tag, that is, the historical response rate of the operator to the tag, is recorded as C j ,j∈1,…,n; then the matching degree between the case and the operator can be defined as the sum of the response rates of the intersection labels of the case's repayment factor label set and the operator's collection expertise label set. The response rate formula of case A and operator B is as follows:

[0025]

[0026] Among them, ind(B j ∈A) means that if the operator is good at label B j If it is included in the case repayment label A, the value is 1, otherwise the value is 0.

[0027] Preferably, in step 6, the matching process between the operator and the case to be assigned is as follows:

[0028] ① Traverse and sort the collection cases in descending order according to the case value;

[0029] ② Select the top-ranked cases to be assigned;

[0030] ③ Calculate the matching degree of all operators to the case, that is, the response rate;

[0031] ④ Assign the case to the operator with the highest matching degree;

[0032] ⑤ Determine whether there are cases to be assigned. If there are cases to be assigned, remove the operators who have reached the case assignment quota and repeat the process ②-⑤. Otherwise, all cases have been assigned.

[0033] Preferably, in step ①, the case value is evaluated using the repayment expectation of the case, that is, the predicted repayment rate under the current collection method multiplied by the amount of debt owed.

[0034] The present invention calculates the matching degree between cases and personnel by combining the case repayment factor label and the personnel's collection expertise label, traverses and allocates cases according to case value through a greedy algorithm, and recommends the cases to be allocated to the most suitable collection personnel, thereby achieving the optimal matching between cases and personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The present invention is a flowchart of the collection case allocation scheme.

[0036] Figure 2 The present invention is a recommended flow chart for allocating cases to operators. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0038] The present invention proposes an intelligent allocation method for collection cases based on a gain model, and introduces text tendency analysis in the feature dimension, hoping to improve the circulation efficiency of post-loan collection case allocation.

[0039] First, the call records of debt collection voice are converted into text through speech recognition, and the communication intention, repayment willingness and repayment ability of overdue customers are mined through text tendency analysis. The structured data such as customer identity information, credit information, and behavior information are combined as independent variables for the construction of subsequent models.

[0040] Secondly, the gain model is used to evaluate the impact of different collection methods (such as text messages, letters, intelligent voice, manual telephone collection, door-to-door visits, arbitration, litigation, etc.) on the results of collection cases. The specific formula can be expressed as: τ(X) = P(Y|X, T = 1) - P(Y|X, T = 0). The gain model allocates the next flow channel of the collection case according to the degree of improvement of τ(X);

[0041] Finally, for manual debt collection (such as manual telephone debt collection, door-to-door visits, etc.), due to the differences in the communication ability and work efficiency of the personnel, when cases are assigned to personnel within the company, the allocation plan is also crucial. The present invention combines the case repayment factor label and the personnel's collection expertise label to calculate the matching degree between the case and the personnel, and uses a greedy algorithm to traverse and allocate cases according to the case value, recommending the cases to be assigned to the most suitable collection personnel to achieve the optimal matching of cases and personnel.

[0042] According to the allocation plan process Figure 1 It can be seen that the entire allocation scheme allocates collection methods with the gain model as the core, adds text tendency analysis in the feature engineering stage to improve the accuracy of customer evaluation, and adds recommendation matching stages in the two collection methods of manual telephone collection and door-to-door visits, so as to improve the efficiency of intelligent allocation of collection cases. The specific process is as follows:

[0043] 1. Cases on hold (cases to be assigned)

[0044] This node means that in actual business, eligible risk accounts will enter the collection system, and this solution will be applied to allocate the cases to be allocated in the collection system. Usually, the judgment rules of risk accounts are determined by financial institutions based on their own business strategies, which are not within the scope of this solution. This solution mainly focuses on the allocation of collection cases to be allocated.

[0045] 2. Text tendency analysis and feature engineering

[0046] This solution uses the gain model to allocate collection cases. The application of the gain model requires relevant customer data as input. Therefore, the data in the database is processed through text tendency analysis and feature engineering.

[0047] For feature engineering of post-loan customers, in addition to commonly used structured data such as application data, credit data, and behavioral data, customer feedback and communication processes are also crucial, which can directly reflect the customer's current communication intention, repayment ability, and willingness to repay. However, since customer feedback and communication processes are usually stored in the form of telephone recordings, they cannot be used directly in the modeling process. With the continuous development of speech recognition and natural language processing technology, it is possible to consider using speech recognition to convert the call process into text, and then use text tendency analysis to identify the customer's current communication intention, repayment ability, and willingness to repay, and add the recognition results to the construction features of the gain model to improve the recognition accuracy of the gain model.

[0048] 3. Gain model

[0049] The commonly used gain modeling method is a differential response model based on two models. Its form is very similar to the definition of the gain model introduced earlier. It includes two response models, one of which is G, which is used to estimate the user's response in the presence of intervention, and the other model G' is used to learn the user's response without intervention. Then, the output of the two models is subtracted to get the desired gain effect. The advantage of this modeling method is that it is relatively simple and easy to understand. At the same time, it can be applied to our common machine learning models, such as LR, SVM, GBDT, etc., so the implementation cost of this model is relatively low.

[0050] In collection management, we pay more attention to the repayment rate of collection cases. Usually, when building a collection scorecard, we mainly use a prediction model to predict the repayment of overdue loans in the next period of time. However, this method only focuses on the repayment rate of overdue loans and does not consider the improvement of repayment by the collection method. Take the following table as an example:

[0051] client No collection Telephone collection Gain Effect A 0.1 0.9 0.8 B 0.9 0.91 0.01 C 0.1 0.11 0.01 D 0.9 0.1 -0.8

[0052] If resources are limited and only one customer can be selected for telephone collection, it can be found that customer B has the highest repayment rate, so customer B will tend to be selected for telephone collection. However, customer B has better qualifications and is likely to repay even without collection. Therefore, this selection will result in customer B and customer D repaying. If the gain effect of telephone collection compared to no collection is considered, customer A will tend to be selected for telephone collection. In the end, customers A, B, and D will all repay, and the overall repayment rate will be higher.

[0053] 4. Allocate collection channels based on gain model results

[0054] In the actual post-loan collection business, there are a variety of collection methods to choose from. Since the costs and effects of different collection methods are different, when allocating collection methods, generally speaking, the higher the cost, the greater the gain effect. Therefore, if the collection method with the largest gain effect is directly selected, then most collection cases will be assigned to the collection method with the highest cost. Obviously, this is not conducive to the effective allocation of collection resources. Therefore, it is possible to consider using the cost of different collection methods as weights, multiplying the weights by the gain effect value, and then sorting them. Select the collection method corresponding to the maximum weighted gain effect, that is, select the collection method that maximizes the return on investment (ROI). The following is the calculation formula for ROI:

[0055] ROI = benefit effect / collection cost.

[0056] 5. Case / Personnel Profile

[0057] After the allocation of collection methods is completed through the gain model, the initial allocation of collection resources has been completed. However, in actual business, manual telephone collection and door-to-door visits are basically completed with the intensive participation of operating personnel. Although the SOP process can standardize the operating process and narrow the gap between operating methods and communication habits, due to the complexity of post-loan collection cases, many situations still need to rely on the experience and communication skills of operating personnel. Therefore, a scientific allocation method is also needed in the process of allocating cases to specific operating personnel.

[0058] When allocating collection cases, considering the large differences in the operating capabilities and types of expertise of operators, a reasonable allocation method is helpful for in-depth mining of operating resources. This solution first profiles the repayment factors of collection cases based on case information, then profiles the operators' expertise in collection based on the repayment factor profile records of historical collection cases, and finally achieves the optimal allocation of cases by matching the repayment factor profile with the expertise in collection profile. When profiling repayment factors, the repayment factors of the case should be profiled based on multiple dimensions such as customer information, account information, collection tendency information, and operator annotations, and the m labels with the highest impact on repayment should be retained, denoted as Ai, i∈1,…,m; then, based on the operator's historical work performance, the operator's response rate for each repayment factor label is calculated, and the n labels with the highest response rate are selected as the collection-good label, denoted as Bj, j∈1,…,n, and the corresponding response rate of the label (the operator's historical response rate for the label) is denoted as Cj, j∈1,…,n; then the matching degree between the case and the operator can be defined as the sum of the response rates of the intersection labels of the case's repayment factor label set and the operator's collection-good label set. The specific formula for the response rate of case A and operator B is as follows:

[0059]

[0060] ind(B j ∈A) means that if the operator is good at label B j If it is included in the case repayment label A, the value is 1, otherwise the value is 0.

[0061] 6. Case matching

[0062] The matching process between operators and cases to be assigned is as follows. For details of the overall process, see Figure 2 :

[0063] ① Traverse and sort the collection cases in descending order according to the case value (the case value can be evaluated using the repayment expectation of the case, that is, the predicted repayment rate under the current collection method multiplied by the amount of debt owed);

[0064] ② Select the top-ranked cases to be assigned;

[0065] ③ Calculate the matching degree of all operators to the case (i.e. the response rate index mentioned above);

[0066] ④ Assign the case to the operator with the highest matching degree;

[0067] ⑤ Determine whether there are cases to be assigned. If there are cases to be assigned, remove the operators who have reached the case assignment quota and repeat the process ②-⑤. Otherwise, all cases have been assigned.

[0068] Through this allocation method, it can be ensured that high-value cases are assigned to more suitable personnel first, thereby improving the overall resource allocation efficiency.

[0069] Taking a recently overdue debt collection client as an example, the specific steps are as follows:

[0070] Step 1: After a customer is overdue, we collect relevant information about the customer and construct feature engineering. During this process, we conduct textual tendency analysis on recent collection records.

[0071] Step 2: Calculate the weighted gain effect value of each collection method through the gain model, and select the collection method corresponding to the maximum value as the collection method for the next stage;

[0072] Step 3: If the collection method is manual collection methods such as manual telephone collection or door-to-door visits, the most suitable operator needs to be selected through the recommendation model;

[0073] Step 4: After the current collection method is used for N days, determine whether the customer meets the case closure requirements. If so, exit the collection process; if not, return to Step 1 and reallocate the case.

[0074]

Explanation of technical terms

[0075] Uplift Model: Uplift Model is an incremental model used to estimate the effect of intervention actions on user response behaviors. The purpose is to predict the causal effect of a certain intervention on individual status or behavior. It is usually used in smart marketing, personalized recommendation and other fields. Mathematically, it can be expressed as the difference between two conditional probabilities, as shown in the formula:

[0076] τ(X)=P(Y|X,T=1)-P(Y|X,T=0)

[0077] Y represents the result (such as overdue loan rollover, repayment, etc.), X is the feature of the user dimension, and T represents the variable of intervention (1 represents intervention, 0 represents no intervention). Therefore, this probability difference represents the change of users with and without intervention. The difficulty of modeling this model is that the training data we obtain is incomplete. For individuals, we cannot observe the performance with and without intervention at the same time, which is the counterfactual problem often mentioned in causal inference. How should we model it? We can change our thinking and make statistics on the average causal effect from the perspective of the population. Suppose we have two groups of homogeneous users, all from first-tier cities / age 25-35 / female. We can intervene in one group of users and not intervene in the other group. Then we count the difference in conversion rate between the two groups. This difference can be approximately regarded as the possible average causal effect of people with the same characteristics. Therefore, the essence of the gain model is to learn the average causal effect of the conditions from the training samples. At the same time, because the model has a certain generalization ability, it can also predict samples that have not been seen.

[0078] ROI: Return On Investment refers to the value that should be returned through investment and the economic return obtained from an investment activity.

[0079] SOP: Standard Operation Procedure, describes the standard operating steps and requirements of an event in a unified format to guide and standardize daily work.

[0080] Text tendency analysis: Text tendency analysis is to mine the user's opinions or comments on a certain thing (such as products, news, etc.), so as to obtain whether the opinion or comment is a positive or negative opinion on the thing. Generally speaking, text sentiment analysis is a three-category problem: positive, negative, and other. Moreover, for a sentence, the sentiment is also specific to a certain aspect (such as willingness to communicate, willingness to repay or ability to repay, etc.). In other words, different aspects of a certain thing are evaluated in one sentence. Existing commonly used analysis methods include sentiment dictionaries, deep learning, sentiment knowledge graphs, sentiment neural networks, etc., and large prediction models can also be used for text tendency analysis.

[0081] Response rate: refers to the response of the collection target during the collection process. For example, when the collection target is the repayment amount, the repayment rate can be used as the response rate; when the collection target is the refund amount, the refund rate can be used as the response rate.

[0082] The above implementation modes are not limitations of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the scope of the technical solution of the present invention also belong to the protection scope of the present invention.

Claims

1. An intelligent allocation method for collection cases based on gain model, Features: The following steps are involved: Step 1: Cases to be assigned are entered into the collection database: Eligible risk accounts enter the collection database and wait for assignment; Step 2: Process the account data in the database through text tendency analysis and feature engineering; Step 3: Construct a gain model: Based on the differential response model of the two models, obtain the gain effect; Step 4: Allocate collection channels based on the results of the gain model; Step 5: Reassign cases based on case / personnel portraits: After completing the initial assignment in step 4, cases need to be reassigned based on the operator's portrait; Step 6: Match the operators with the cases to be assigned to complete case allocation: prioritize high-value cases and assign them to appropriate operators.

2. According to claim 1, the intelligent allocation method for debt collection cases based on the gain model, Features: In the step 2, text tendency analysis refers to using speech recognition to convert the call process into text, and then using text tendency analysis to identify the customer's current communication intention, repayment ability and repayment willingness, and adding the recognition results to the construction features of the gain model.

3. The method for intelligently allocating collection cases based on a gain model according to claim 1, Features: In the step three, one model G is used to estimate the user's response in the presence of intervention, and another model G' is used to learn the user's response in the absence of intervention. The outputs of the two models are then subtracted to obtain a gain effect.

4. The method for intelligently allocating collection cases based on a gain model according to claim 1, Features: In step 4, the costs of different collection methods are considered as weights, and the weights are multiplied by the gain effect values ​​and then sorted. The collection method corresponding to the maximum weighted gain effect is selected, that is, the collection method with the maximum return on investment (ROI) is selected. The calculation formula for ROI is: ROI = benefit effect / collection cost.

5. The method for intelligently allocating collection cases based on a gain model according to claim 1, Features: Step 5 The following steps are involved: First, the repayment factors of the collection cases are profiled based on the case information; Secondly, we will create a collection profiling for the operators based on the repayment factor profile records of historical collection cases; Finally, the optimal allocation of cases is achieved by matching the repayment factor profile with the collection expertise profile.

6. The method for intelligently allocating collection cases based on the gain model according to claim 5, Features: In step 5, when profiling repayment factors, the repayment factors of the case should be profiled in multiple dimensions based on customer information, account information, collection tendency information, and operator annotations, and the m labels with the highest impact on repayment should be retained, denoted as A i ,i∈1,…,m; Then, based on the operator's historical performance, the operator's response rate for each repayment factor label is calculated, and the n labels with the highest response rate are selected as the collection expert labels, denoted as B j ,j∈1,…,n, the response rate corresponding to the tag, that is, the historical response rate of the operator to the tag, is recorded as C j ,j∈1,…,n; then the matching degree between the case and the operator can be defined as the sum of the response rates of the intersection labels of the case's repayment factor label set and the operator's collection expertise label set. The response rate formula of case A and operator B is as follows: Among them, ind(B j ∈A) means that if the operator is good at label B j If it is included in the case repayment label A, the value is 1, otherwise the value is 0.

7. The method for intelligently allocating collection cases based on a gain model according to claim 1, Features: In step 6, the matching process between operators and cases to be assigned is as follows: ① Traverse and sort the collection cases in descending order according to the case value; ② Select the top-ranked cases to be assigned; ③ Calculate the matching degree of all operators to the case, that is, the response rate; ④ Assign the case to the operator with the highest matching degree; ⑤ Determine whether there are cases to be assigned. If there are cases to be assigned, remove the operators who have reached the case assignment quota and repeat the process ②-⑤. Otherwise, all cases have been assigned.

8. The method for intelligently allocating collection cases based on a gain model according to claim 7, Features: In step ①, the case value is evaluated using the repayment expectation of the case, that is, the predicted repayment rate under the current collection method multiplied by the amount of debt owed.