Collection division method and system based on dynamic load balancing and intelligent weight calculation

Abstract

The invention provides a collection case division method and system based on dynamic load balancing and intelligent weight calculation. The method comprises the following steps: constructing a collection person ability portrait and a case feature library based on historical collection case records; extracting case features of a to-be-allocated case in a case feature library, inputting the case features into a pre-trained prediction model, and obtaining a standardized effective working time length of the to-be-allocated case as a case weight; obtaining a language threshold value of the to-be-allocated case under the corresponding language category, and taking the language threshold value as a language threshold value of the case; the method comprises the following steps: constructing a target function by taking a minimum weighted sum of a collection person overload degree and case distribution unevenness as an optimization target, solving the target function by using a greedy algorithm, and for each to-be-distributed case in the solving process, screening collection persons of which the language ability values are not less than a language threshold value of the case to form a candidate set, and obtaining a preliminary case allocation result based on the candidate set of each to-be-allocated case. According to the invention, the automatic case division method considering the language ability matching degree is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a collection case allocation method and system based on dynamic load balancing and intelligent weight calculation. Background Technology

[0002] Current debt collection systems typically consider the match between cases and collectors' overall capabilities and / or the overall workload balance of the collection team when allocating cases, so as to automatically execute the allocation of collection cases based on individual capabilities and individual task load.

[0003] However, the existing collection system's allocation schemes rarely consider the impact of collectors' language skills on collection efficiency. It is understandable that when collectors who only speak Mandarin are dealing with debtors from regions that only speak dialects such as Wu or Cantonese, their case collection efficiency will be significantly lower than normal, leading to a backlog of cases and exacerbating the uneven workload of the collection team, resulting in a vicious cycle of case allocation.

[0004] Therefore, how to reasonably allocate debt collectors with the corresponding language skills to cases has become an urgent technical problem to be solved. Summary of the Invention

[0005] This invention provides a debt collection case allocation method and system based on dynamic load balancing and intelligent weight calculation, which addresses the shortcomings of existing technologies that rarely consider the impact of debt collectors' language skills on collection efficiency. It realizes a debt collection case allocation method based on dynamic load balancing and intelligent weight calculation, and achieves automatic allocation of cases that take into account the influence of language.

[0006] This invention provides a collection case allocation method based on dynamic load balancing and intelligent weight calculation, comprising: A collection agent capability profile and a case feature database are constructed based on historical collection case records. The collection agent capability profile includes a collection agent's language ability matrix. Case features of cases to be assigned are extracted from the case feature database and input into a pre-trained prediction model. The shared feature encoder of the prediction model extracts shared features, which are then input into the decoding branch of Task 1 of the prediction model to obtain the standardized effective working time of the case to be assigned, which is used as the case weight. The shared features are then concatenated with the language category code of the case to be assigned and input into the decoding branch of Task 2 of the prediction model to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. An objective function is constructed with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be allocated, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be allocated, preliminary case allocation results are obtained.

[0007] The collection allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention further includes: If the objective function value corresponding to the preliminary case allocation result is greater than the preset threshold or is at a preset time node, a reallocation is triggered, and cases that can be released and currently unallocated cases are selected based on preset rules and entered into the allocation pool. The greedy algorithm is used again to solve the objective function for the cases in the allocation pool until the objective function value corresponding to the case allocation result is not greater than the preset threshold, or the preset number of iterations is reached; Output the results of the re-division of cases.

[0008] According to the debt collection allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention, before the step of constructing an objective function with the weighted sum of debt collector overload and allocation unevenness as the optimization objective, the method further includes: For each debt collector, the sum of the case weights of all cases in their process is calculated as the real-time weighted load of the debt collector. Calculate the standard deviation of the real-time weighted load of all collectors as the case allocation unevenness; Calculate the difference between the real-time weighted load of each collector and the preset rated load threshold. Sum the non-negative differences for each collector and divide by the number of collectors to obtain the collector overload level.

[0009] According to the collection case allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention, before the step of forming a candidate set by screening collection collectors whose language ability values ​​are not less than the language threshold of the case, the method further includes: The basic score for each collector in each language is determined based on the collector's language proficiency matrix. Calculate the difference between the collector's historical collection success rate in any language and the collection team's average collection success rate in that language, and then multiply it by a correction factor to obtain the collector's correction score in the corresponding language. The sum of the base score and correction score for each language used by the debt collector is calculated as the debt collector's language proficiency score.

[0010] The collection allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention further includes: Obtain case information from historical cases, calculate the weighted sum of case call duration, follow-up record writing time, and repayment negotiation time as a standardized effective working time, and normalize it to a preset range; Using the characteristics of historical cases as samples, their normalized and standardized effective working hours as the first label, and the language ability value of the corresponding debt collector in the language category of the historical case as the second label, a dataset is constructed for training the prediction model.

[0011] According to the present invention, a collection case allocation method based on dynamic load balancing and intelligent weight calculation is provided, wherein the total loss of the prediction model is composed of the weighted sum of the loss of task one and the loss of task two. The loss for Task 2 includes fitting loss, logical consistency loss, and range constraint loss. The fitting loss is used to constrain the predicted value of Task 2 to approach the target value. The logical consistency loss is used to make historical cases with longer standardization effective working time have higher language thresholds. The range constraint loss is used to constrain the predicted value of Task 2 to be within a pre-set threshold range.

[0012] This invention also provides a collection case allocation system based on dynamic load balancing and intelligent weight calculation, comprising: The construction module is used to build a debt collector capability profile and a case feature database based on historical debt collection case records. The debt collector capability profile includes a matrix of the debt collector's language ability. The prediction module is used to extract case features from the case feature library of cases to be assigned, and input them into the pre-trained prediction model. The shared feature encoder of the prediction model extracts shared features, which are input into the decoding branch of Task 1 of the prediction model to obtain the standardized effective working time of the case to be assigned, which is used as the case weight. The shared features are concatenated with the language category code of the case to be assigned and input into the decoding branch of Task 2 of the prediction model to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. The allocation module is used to construct an objective function with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be allocated, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be allocated, a preliminary case allocation result is obtained.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the collection assignment method based on dynamic load balancing and intelligent weight calculation as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the collection assignment method based on dynamic load balancing and intelligent weight calculation as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the collection case assignment method based on dynamic load balancing and intelligent weight calculation as described above.

[0016] The collection case allocation method and system based on dynamic load balancing and intelligent weight calculation provided by this invention uses a multi-task learning approach, with a clearly labeled Task 1 assisting in the training of Task 2, to accurately obtain the standardized effective working time and language threshold of each case to be allocated. This ensures that each case is assigned to a collector with the corresponding language ability during the case allocation process, and guarantees the overall load balance of the collection team, thus realizing an automated case allocation method. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts of the collection case allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention; Figure 2 This is the second flowchart of the debt collection case allocation method based on dynamic load balancing and intelligent weight calculation provided by the present invention. Figure 3 This is a schematic diagram of the debt collection and case allocation system based on dynamic load balancing and intelligent weight calculation provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] The following is combined Figure 1 and Figure 2 This invention introduces a debt collection case allocation method based on dynamic load balancing and intelligent weight calculation, such as... Figure 1 As shown, it includes: Step 101: Construct a debt collector capability profile and a case feature database based on historical debt collection case records. The debt collector capability profile includes a matrix of the debt collector's language ability. Historical collection case records are records of cases that have been completed, including the time the case was initiated, the time of completion, the collection record, and the collection agent.

[0021] Based on these historical records, a capability profile is constructed for each debt collector. Optionally, the capability profile of a debt collector includes a matrix of the debt collector's language skills, performance data, etc.

[0022] The language proficiency matrix records the types of languages / dialects that the corresponding debt collectors master and their proficiency ratings, for example, defined as: beginner = 1, intermediate = 2, native speaker / advanced = 3.

[0023] Collector performance data includes collection success rates and average case coarse-grained duration for collectors using different languages ​​over a certain historical period (e.g., within 12 months).

[0024] The case feature database records information on all cases, including historical collection cases, cases in progress, and collection cases awaiting allocation. Case information includes the number of overdue days, overdue amount, number of previous contacts, result of the last contact, customer credit score, geographic tags, and language preference tags.

[0025] Among them, the geographical label represents the customer's location, and the language preference label is determined according to the customer's non-Mandarin language category. For example, when a customer is proficient in both Mandarin and a dialect, the dialect is determined as the language preference label for the case.

[0026] Step 102: Extract case features from the case feature library of cases to be assigned, input them into the pre-trained prediction model, and extract shared features by the shared feature encoder of the prediction model. The shared features are input into the decoding branch of Task 1 of the prediction model to obtain the standardized effective working time of the case to be assigned, which is used as the case weight. The shared features are concatenated with the language category code of the case to be assigned and input into the decoding branch of Task 2 of the prediction model to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. When there are cases to be assigned in the case feature database, such as Figure 2 As shown, optionally, the case information of the cases to be assigned can be extracted into case features by a text encoder such as Transformer, and then the case features can be input into a pre-trained prediction model.

[0027] The prediction model is trained based on multi-task learning and includes two task branches. Task 1 is used to predict the standardized effective working time of the cases to be assigned as the case weights, and Task 2 is used to predict the language threshold of the language category corresponding to the language preference label of the cases to be assigned.

[0028] In this embodiment, the language preference labels of the cases to be assigned are encoded as language category codes, which are then concatenated with the shared features extracted by the shared feature encoder and input into the decoding branch of Task 2 to guide the decoding branch of Task 2 to output the language threshold corresponding to the language category. M it .

[0029] The standardized effective working time is the effective working time of cases mapped to a preset range, which is convenient for use as the weight of cases.

[0030] Optionally, when training the prediction model, Task 1 can directly use the standardized effective working time of historical collection cases as supervision labels. For Task 2, any existing unsupervised training method can be used, or corresponding language threshold labels can be added to historical collection cases by manual annotation before supervised training.

[0031] It should be noted that the prediction model proposed in this embodiment adopts a multi-task learning approach. By sharing a feature encoder, the feature representation of case processing difficulty learned by Task 1 when learning with its explicit duration label can be passed to the Task 2 branch through the sharing layer. This provides the Task 2 branch with effective indirect supervision signals and priors, which helps the decoding branch of Task 2 to learn more accurately the intrinsic relationship between case complexity and language threshold. Ultimately, this enables Task 2 to have relatively accurate prediction performance even when direct supervision signals are insufficient or labels are weak, thereby obtaining the language threshold for each case to be assigned.

[0032] Step 103: Construct an objective function with the weighted sum of collector overload and case allocation unevenness as the optimization objective. Solve the objective function using a greedy algorithm. During the solution process, for each case to be allocated, select collectors with language ability values ​​not less than the language threshold of the case to form a candidate set. Obtain preliminary case allocation results based on the candidate set of each case to be allocated.

[0033] Understandably, when each case is assigned, the prediction model predicts the standardized effective working time for each case and updates and stores it in the case feature database.

[0034] Optionally, the debt collector overload can be determined based on the standardized effective working time of all cases being processed by the debt collector, for example, by directly calculating the sum of the standardized effective working time of all cases being processed as the debt collector overload value.

[0035] Optionally, the case allocation unevenness can be determined by the standard deviation of the collector overload of all collectors in the collection team.

[0036] Furthermore, an objective function is constructed to minimize the weighted sum of collector overload and case allocation unevenness, aiming to ensure that collectors' workload is minimized while making the workload of the collection team more even. ; In the formula, and X represents the configurable weighting coefficients (ranging from [0.5, 1.5]), Y represents the case allocation unevenness, and Y represents the collector overload, in order to minimize... To optimize the objective.

[0037] At the same time, based on each collector's language ability matrix, the collection agent is determined. j In language type i Language proficiency score M ij .

[0038] In this implementation, a greedy algorithm is used to solve the objective function and complete the case allocation. During this process, for each case to be allocated… t ,filter The debt collectors are selected to form a candidate set. The candidate set is then iterated through to calculate case allocation. The increase in value ,choose The smallest collection agent allocation yields preliminary case assignment results after all pending cases have been allocated.

[0039] If the candidate set is empty, for example, if a special language case cannot be matched with a corresponding debt collector, then the rate is reduced. After reaching the preset value, the candidate set is re-filtered. If it is still empty, the collection agents with Mandarin proficiency level ≥ 2 are directly included in the candidate set to avoid cases being unable to be allocated.

[0040] By employing the above methods, the matching degree between each collector's language ability and the language requirements of the case is taken into account when allocating cases. This ensures that cases with language requirements are assigned to collectors with the corresponding language abilities, avoiding the long-term accumulation of cases among collectors without language skills, which would create a burden.

[0041] This invention employs a multi-task learning approach, using a clearly labeled Task 1 to assist in the training of Task 2. This allows for the accurate acquisition of standardized effective working time and language thresholds for each case to be assigned. As a result, during the case assignment process, each case is assigned to a collector with the corresponding language skills, ensuring overall load balancing for the collection team and achieving an automated case assignment method.

[0042] The collection case allocation method based on dynamic load balancing and intelligent weight calculation in this invention also includes: If the objective function value corresponding to the preliminary case allocation result is greater than the preset threshold or is at a preset time node, a reallocation is triggered, and cases that can be released and currently unallocated cases are selected based on preset rules and entered into the allocation pool. The greedy algorithm is used again to solve the objective function for the cases in the allocation pool until the objective function value corresponding to the case allocation result is not greater than the preset threshold, or the preset number of iterations is reached; Output the results of the re-division of cases.

[0043] It is understandable that, because the case allocation method in this implementation means that cases to be allocated are only assigned to collectors in the candidate set, it may lead to situations where a concentration of collection cases in certain regions arrives, resulting in... Even with the minimized initial case allocation, there is still a significant uneven distribution of cases, meaning that the cases to be allocated are concentrated among a portion of debt collectors with language skills.

[0044] Therefore, the preset threshold determined in this embodiment T If the preliminary case assignment results correspond to If the number of cases exceeds a preset threshold, it indicates that the initial case allocation results still place a significant burden on some collection agents. In this case, a reallocation is triggered. Optionally, T The value can be determined based on experience and the number of people in the collection team, giving larger collection teams greater potential. T value.

[0045] At the same time, the system will also trigger a redistribution at preset time nodes, such as 0:00 and 12:00 every day, to allocate cases to be assigned in the case feature database.

[0046] During the reallocation process, the case allocation system first filters out some releaseable cases based on the received preset rules and releases them together with the currently pending cases to form a new case allocation pool.

[0047] Optionally, preset rules are pre-configured in the case assignment system to ensure that overload is alleviated while avoiding interference with cases that are making progress. The preset rules in this embodiment are as follows: The first priority is status priority: prioritize the release of newly assigned cases that have not been contacted (i.e., the collector has not yet communicated with the customer, and release will not cause business interference) and cases that have been dormant due to multiple unsuccessful attempts to contact the customer (dormant time ≥ 7 days, indicating long-term lack of progress, and reassignment may bring new progress).

[0048] The second priority is weight priority: under the same conditions, cases with a weight lower than the preset weight threshold are released first. Low-weight cases require less effort and their release has a smaller impact on the overall collection progress. The third priority is time priority: when the status and weight are the same, the case with the earliest allocation time will be released first to avoid cases being backlogged in the hands of the same collector for a long time.

[0049] Protective exclusion: Prohibit the release of cases with agreed repayment plans, under negotiation, partially repaid, initiated by the customer, or marked as midpoints for follow-up by the collection agent, in order to ensure business continuity and collection effectiveness; prohibit the release of cases less than 24 hours after the last allocation, to avoid frequent allocation affecting the collection rhythm of the staff.

[0050] In this embodiment, each round of reallocation releases a portion of the already allocated cases according to the aforementioned preset rules, and these cases, along with new cases, enter the waiting-to-allocate pool for reallocation. The reallocation method is the same as the initial allocation method, that is, the aforementioned greedy algorithm is used to solve the problem. The smallest possible outcome.

[0051] During the refactoring phase, a preset number of iterations is set to avoid infinite loops. This iteration continues until... Less than T If the number of iterations reaches the preset number of iterations, the iteration stops, and the re-allocation result is generated as the final output of the re-allocation result.

[0052] In the collection case allocation method based on dynamic load balancing and intelligent weight calculation of the present invention, before the step of constructing an objective function with the weighted sum of collection agent overload and case allocation unevenness as the optimization objective, the method further includes: For each debt collector, the sum of the case weights of all cases in their process is calculated as the real-time weighted load of the debt collector. Calculate the standard deviation of the real-time weighted load of all collectors as the case allocation unevenness; Calculate the difference between the real-time weighted load of each collector and the preset rated load threshold. Sum the non-negative differences for each collector and divide by the number of collectors to obtain the collector overload level.

[0053] In this embodiment, the real-time weighted load of the collection agent is based on the collection agent's... j Currently allocated k The case weighting and calculation of each case reflect the current workload level of the debt collector: .

[0054] Then, the case allocation inequality is calculated based on the standard deviation of the real-time weighted workload of all collectors in the collection team: ; In the formula, n represents the total number of debt collectors in the collection team.

[0055] Based on this, the collector overload level Y is calculated to quantify the overall overload situation of collectors within the team: ; In the formula, debt collector j The preset rated load threshold is determined based on the employee's job level (e.g., junior / intermediate / senior debt collector).

[0056] In the debt collection case allocation method based on dynamic load balancing and intelligent weight calculation of the present invention, before the step of forming a candidate set by selecting debt collectors whose language ability value is not less than the language threshold of the case, the method further includes: The basic score for each collector in each language is determined based on the collector's language proficiency matrix. Calculate the difference between the collector's historical collection success rate in any language and the collection team's average collection success rate in that language, and then multiply it by a correction factor to obtain the collector's correction score in the corresponding language. The sum of the base score and correction score for each language used by the debt collector is calculated as the debt collector's language proficiency score.

[0057] Although it is possible to directly contact the debt collector j In language i The language level is used as a language ability value, but it does not mean that the debt collector does not have such language ability. For example, if a debt collector does not speak a certain dialect, he / she cannot carry out debt collection.

[0058] Therefore, in this embodiment, the language ability values ​​of the debt collectors used to select the candidate set are determined jointly based on the debt collector's language ability matrix and historical collection success rate. : ; In the formula, To determine debt collectors based on language proficiency levels in the language ability matrix j In language i The basic score below; For corrective scoring, Indicates the correction coefficient. debt collector j Use the target language i Historical debt collection success rate Use the target language within the teami The average collection success rate.

[0059] The collection case allocation method based on dynamic load balancing and intelligent weight calculation in this invention also includes: Obtain case information from historical cases, calculate the weighted sum of case call duration, follow-up record writing time, and repayment negotiation time as a standardized effective working time, and normalize it to a preset range; Using the characteristics of historical cases as samples, their normalized and standardized effective working hours as the first label, and the language ability value of the corresponding debt collector in the language category of the historical case as the second label, a dataset is constructed for training the prediction model.

[0060] In this implementation method, the standardized effective working time is defined as: Effective working hours = Time spent on processing × 0.6 + Time spent writing follow-up records × 0.3 + Time spent negotiating repayment × 0.1; The effective working time is then normalized to a preset range using Min-Max, which in this embodiment is the interval [0, 1], to obtain the standardized effective working time.

[0061] By using the above method, the standardized effective working time for each completed historical case can be determined based on the case information stored in the case feature database.

[0062] Furthermore, determine the language proficiency value of the collection agent corresponding to each historical case in the language category corresponding to that historical case.

[0063] Then, using the case characteristics of historical cases as samples, the first and second labels are determined in the manner described above, thus constructing a dataset for training the prediction model.

[0064] In the collection case allocation method based on dynamic load balancing and intelligent weight calculation of the present invention, the total loss of the prediction model is composed of the weighted sum of the loss of Task 1 and the loss of Task 2. The loss for Task 2 includes fitting loss, logical consistency loss, and range constraint loss. The fitting loss is used to constrain the predicted value of Task 2 to approach the target value. The logical consistency loss is used to make historical cases with longer standardization effective working time have higher language thresholds. The range constraint loss is used to constrain the predicted value of Task 2 to be within a pre-set threshold range.

[0065] It is understandable that even for completed historical cases, the language threshold is difficult to determine, and the language proficiency value of the debt collector handling the case cannot accurately define the language threshold of the case. Therefore, this implementation constructs the following loss for the prediction model to more accurately predict the language threshold of each case through multi-task learning.

[0066] Specifically, the total loss of the prediction model: ; In the formula, and The losses for Mission 1 and Mission 2 are respectively. and These are the corresponding weighting coefficients.

[0067] in: ; In the formula, m is the number of samples in the training batch, that is, the number of historical cases in the training batch; This represents the prediction result for the m-th case from the decoding branch of Task 1. This represents the actual standardized and effective working time for the m-th case.

[0068] Furthermore, the loss for Task 2 consists of the optimal fitting loss, the logical consistency loss, and the range constraint loss, where the fitting loss... Defined as: ; ; In the formula, The language threshold predicted by the decoding network of the second branch for the m-th case. Let m be the target threshold for the m-th case. Let m be the language proficiency score of the debt collector corresponding to the m-th case. To adjust the intensity hyperparameter, Let m be the actual case weight. The average case weight of all cases. The case weight standard deviation for all cases.

[0069] Using the above method, based on the language proficiency of the debt collector corresponding to the case, if the case weight is greater, that is, the standardized processing time is longer, the actual language threshold of the case will also increase accordingly.

[0070] Logical consistency loss Defined as: ; In the formula, This indicates all of this batch The mean, This indicates that a loss occurs only if the content within the parentheses is upright.

[0071] By using the above method, if the case weight is higher than the mean, the predicted language threshold will also be higher than the mean, thus ensuring that the language threshold and the case weight remain consistent in direction.

[0072] Range constraint loss Defined as: ; Among them, [y min ,y max [] indicates a pre-set threshold range, which can be determined based on the collectors with the highest and lowest language proficiency scores for each language.

[0073] Based on this, the loss for Task 2 is defined as: ; In the formula, , and These are the weight parameters corresponding to the fitting loss, logistic consistency loss, and range loss, respectively.

[0074] This invention, through the aforementioned loss function, enables Task 1 to better assist the training of Task 2, a task for which clear labels are difficult to obtain, thereby acquiring more accurate language thresholds for cases to be assigned in order to achieve automatic allocation of collection cases.

[0075] Among them, the shared feature encoder and the decoding networks for Task 1 and Task 2 can all be built based on the multilayer perceptron (MLP).

[0076] The collection allocation system based on dynamic load balancing and intelligent weight calculation provided by the present invention will be described below. The collection allocation system based on dynamic load balancing and intelligent weight calculation described below can be referred to in correspondence with the collection allocation method based on dynamic load balancing and intelligent weight calculation described above.

[0077] like Figure 3 As shown, the collection allocation system based on dynamic load balancing and intelligent weight calculation includes a construction module 301, a prediction module 302, and an allocation module 303. Module 301 is used to construct a debt collector capability profile and a case feature database based on historical debt collection case records. The debt collector capability profile includes a debt collector's language ability matrix. Historical collection case records are records of cases that have been completed, including the time the case was initiated, the time of completion, the collection record, and the collection agent.

[0078] Based on these historical records, a capability profile is constructed for each debt collector. Optionally, the capability profile of a debt collector includes a matrix of the debt collector's language skills, performance data, etc.

[0079] The language proficiency matrix records the types of languages / dialects that the corresponding debt collectors master and their proficiency ratings, for example, defined as: beginner = 1, intermediate = 2, native speaker / advanced = 3.

[0080] Collector performance data includes collection success rates and average case coarse-grained duration for collectors using different languages ​​over a certain historical period (e.g., within 12 months).

[0081] The case feature database records information on all cases, including historical collection cases, cases in progress, and collection cases awaiting allocation. Case information includes the number of overdue days, overdue amount, number of previous contacts, result of the last contact, customer credit score, geographic tags, and language preference tags.

[0082] Among them, the geographical label represents the customer's location, and the language preference label is determined according to the customer's non-Mandarin language category. For example, when a customer is proficient in both Mandarin and a dialect, the dialect is determined as the language preference label for the case.

[0083] Prediction module 302 is used to extract case features of cases to be assigned from the case feature library and input them into a pre-trained prediction model. The shared feature encoder of the prediction model extracts shared features, which are then input into the decoding branch of task one of the prediction model to obtain the standardized effective working time of the case to be assigned, which is used as the case weight. The shared features are then concatenated with the language category code of the case to be assigned and input into the decoding branch of task two of the prediction model to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. When there are cases to be assigned in the case feature database, such as Figure 2 As shown, optionally, the case information of the cases to be assigned can be extracted into case features by a text encoder such as Transformer, and then the case features can be input into a pre-trained prediction model.

[0084] The prediction model is trained based on multi-task learning and includes two task branches. Task 1 is used to predict the standardized effective working time of the cases to be assigned as the case weights, and Task 2 is used to predict the language threshold of the language category corresponding to the language preference label of the cases to be assigned.

[0085] In this embodiment, the language preference labels of the cases to be assigned are encoded as language category codes, which are then concatenated with the shared features extracted by the shared feature encoder and input into the decoding branch of Task 2 to guide the decoding branch of Task 2 to output the language threshold corresponding to the language category. Mit .

[0086] The standardized effective working time is the effective working time of cases mapped to a preset range, which is convenient for use as the weight of cases.

[0087] Optionally, when training the prediction model, Task 1 can directly use the standardized effective working time of historical collection cases as supervision labels. For Task 2, any existing unsupervised training method can be used, or corresponding language threshold labels can be added to historical collection cases by manual annotation before supervised training.

[0088] It should be noted that the prediction model proposed in this embodiment adopts a multi-task learning approach. By sharing a feature encoder, the feature representation of case processing difficulty learned by Task 1 when learning with its explicit duration label can be passed to the Task 2 branch through the sharing layer. This provides the Task 2 branch with effective indirect supervision signals and priors, which helps the decoding branch of Task 2 to learn more accurately the intrinsic relationship between case complexity and language threshold. Ultimately, this enables Task 2 to have relatively accurate prediction performance even when direct supervision signals are insufficient or labels are weak, thereby obtaining the language threshold for each case to be assigned.

[0089] The allocation module 303 is used to construct an objective function with the weighted sum of collector overload and case allocation unevenness as the optimization objective, solve the objective function using a greedy algorithm, and in the process of solving, for each case to be allocated, select collectors whose language ability value is not less than the language threshold of the case to form a candidate set, and obtain preliminary case allocation results based on the candidate set of each case to be allocated.

[0090] Understandably, when each case is assigned, the prediction model predicts the standardized effective working time for each case and updates and stores it in the case feature database.

[0091] Optionally, the debt collector overload can be determined based on the standardized effective working time of all cases being processed by the debt collector, for example, by directly calculating the sum of the standardized effective working time of all cases being processed as the debt collector overload value.

[0092] Optionally, the case allocation unevenness can be determined by the standard deviation of the collector overload of all collectors in the collection team.

[0093] Furthermore, an objective function is constructed to minimize the weighted sum of collector overload and case allocation unevenness, aiming to ensure that collectors' workload is minimized while making the workload of the collection team more even. ; In the formula, and X represents the configurable weighting coefficients (ranging from [0.5, 1.5]), Y represents the case allocation unevenness, and Y represents the collector overload, in order to minimize... To optimize the objective.

[0094] At the same time, based on each collector's language ability matrix, the collection agent is determined. j In language type i Language proficiency score M ij .

[0095] In this implementation, a greedy algorithm is used to solve the objective function and complete the case allocation. During this process, for each case to be allocated… t ,filter The debt collectors are selected to form a candidate set. The candidate set is then iterated through to calculate case allocation. The increase in value ,choose The smallest collection agent allocation yields preliminary case assignment results after all pending cases have been allocated.

[0096] If the candidate set is empty, for example, if a special language case cannot be matched with a corresponding debt collector, then the rate is reduced. After reaching the preset value, the candidate set is re-filtered. If it is still empty, the collection agents with Mandarin proficiency level ≥ 2 are directly included in the candidate set to avoid cases being unable to be allocated.

[0097] By employing the above methods, the matching degree between each collector's language ability and the language requirements of the case is taken into account when allocating cases. This ensures that cases with language requirements are assigned to collectors with the corresponding language abilities, avoiding the long-term accumulation of cases among collectors without language skills, which would create a burden.

[0098] This invention employs a multi-task learning approach, using a clearly labeled Task 1 to assist in the training of Task 2. This allows for the accurate acquisition of standardized effective working time and language thresholds for each case to be assigned. As a result, during the case assignment process, each case is assigned to a collector with the corresponding language skills, ensuring overall load balancing for the collection team and achieving an automated case assignment method.

[0099] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logic instructions in the memory 430 to execute a debt collection case allocation method based on dynamic load balancing and intelligent weight calculation. This method includes: constructing a debt collector capability profile and a case feature database based on historical debt collection case records, wherein the debt collector capability profile includes a language ability matrix of the debt collectors; extracting case features from the case feature database of cases to be allocated, inputting them into a pre-trained prediction model, where the shared feature encoder of the prediction model extracts shared features, which are then input into the decoding branch of Task 1 of the prediction model to obtain the standardized effective working time of the cases to be allocated, which serves as the case weight; concatenating the shared features with the language category encoding of the cases to be allocated and inputting the result into the decoding branch of Task 2 of the prediction model to obtain the language threshold of the cases to be allocated under the corresponding language category, which serves as the language threshold of the cases; constructing an objective function with the weighted sum of debt collector overload and case allocation unevenness as the optimization objective, solving the objective function using a greedy algorithm, and during the solution process, selecting debt collectors with language ability values ​​not less than the language threshold of each case to form a candidate set, and obtaining preliminary case allocation results based on the candidate set of each case to be allocated.

[0100] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the collection case allocation method based on dynamic load balancing and intelligent weight calculation provided by the above methods. This method includes: constructing a collection agent capability profile and a case feature database based on historical collection case records, wherein the collection agent capability profile includes a collection agent's language ability matrix; extracting case features of cases to be allocated from the case feature database, inputting them into a pre-trained prediction model, and extracting shared features by the shared feature encoder of the prediction model. The decoding branch of Task 1 of the prediction model obtains the standardized effective working time of the case to be assigned, which is used as the case weight. The shared features are concatenated with the language category code of the case to be assigned and input into the decoding branch of Task 2 of the prediction model to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. An objective function is constructed with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be assigned, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be assigned, a preliminary case allocation result is obtained.

[0102] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the debt collection case allocation method based on dynamic load balancing and intelligent weight calculation provided by the above methods. This method includes: constructing a debt collector competency profile and a case feature database based on historical debt collection case records, wherein the debt collector competency profile includes a language competency matrix of the debt collector; extracting case features of cases to be allocated from the case feature database, inputting them into a pre-trained prediction model, and having shared features extracted by the shared feature encoder of the prediction model, wherein the shared features are input into the decoding branch of Task 1 of the prediction model. The standardized effective working time of the cases to be assigned is obtained and used as the case weight. The shared features are concatenated with the language category codes of the cases to be assigned and input into the decoding branch of the prediction model task two to obtain the language threshold of the cases to be assigned under the corresponding language category, which is used as the language threshold of the cases. An objective function is constructed with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be assigned, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be assigned, a preliminary case allocation result is obtained.

[0103] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A debt collection case allocation method based on dynamic load balancing and intelligent weight calculation, characterized in that, include: A collection agent capability profile and a case feature database are constructed based on historical collection case records. The collection agent capability profile includes a collection agent's language ability matrix. Extract case features from the case feature database of cases to be assigned, input them into the pre-trained prediction model, extract shared features by the shared feature encoder of the prediction model, input the shared features into the decoding branch of Task 1 of the prediction model, and obtain the standardized effective working time of the cases to be assigned, which is used as the case weight; The shared features are concatenated with the language category code of the case to be assigned and then input into the decoding branch of the prediction model task two to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. An objective function is constructed with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be allocated, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be allocated, preliminary case allocation results are obtained.

2. The debt collection case allocation method based on dynamic load balancing and intelligent weight calculation according to claim 1, characterized in that, Also includes: If the objective function value corresponding to the preliminary case allocation result is greater than the preset threshold or is at a preset time node, a reallocation is triggered, and cases that can be released and currently unallocated cases are selected based on preset rules and entered into the allocation pool. The greedy algorithm is used again to solve the objective function for the cases in the allocation pool until the objective function value corresponding to the case allocation result is not greater than the preset threshold, or the preset number of iterations is reached; Output the results of the re-division of cases.

3. The collection case allocation method based on dynamic load balancing and intelligent weight calculation according to claim 1, characterized in that, Before the step of constructing the objective function with the weighted sum of collection agent overload and case allocation unevenness as the optimization objective, the following steps are also included: For each debt collector, the sum of the case weights of all cases in their process is calculated as the real-time weighted load of the debt collector. Calculate the standard deviation of the real-time weighted load of all collectors as the case allocation unevenness; Calculate the difference between the real-time weighted load of each collector and the preset rated load threshold. Sum the non-negative differences for each collector and divide by the number of collectors to obtain the collector overload level.

4. The collection case allocation method based on dynamic load balancing and intelligent weight calculation according to claim 1, characterized in that, Before the step of forming a candidate set of debt collectors whose language proficiency values ​​are not less than the language threshold for the case, the method further includes: The basic score for each collector in each language is determined based on the collector's language proficiency matrix. Calculate the difference between the collector's historical collection success rate in any language and the collection team's average collection success rate in that language, and then multiply it by a correction factor to obtain the collector's correction score in the corresponding language. The sum of the base score and correction score for each language used by the debt collector is calculated as the debt collector's language proficiency score.

5. The collection case allocation method based on dynamic load balancing and intelligent weight calculation according to any one of claims 1-4, characterized in that, Also includes: Obtain case information from historical cases, calculate the weighted sum of case call duration, follow-up record writing time, and repayment negotiation time as a standardized effective working time, and normalize it to a preset range; Using the characteristics of historical cases as samples, their normalized and standardized effective working hours as the first label, and the language ability value of the corresponding debt collector in the language category of the historical case as the second label, a dataset is constructed for training the prediction model.

6. The collection case allocation method based on dynamic load balancing and intelligent weight calculation according to any one of claims 1-4, characterized in that, The total loss of the prediction model is a weighted sum of the losses from Task 1 and Task 2. The loss for Task 2 includes fitting loss, logical consistency loss, and range constraint loss. The fitting loss is used to constrain the predicted value of Task 2 to approach the target value. The logical consistency loss is used to make historical cases with longer standardization effective working time have higher language thresholds. The range constraint loss is used to constrain the predicted value of Task 2 to be within a pre-set threshold range.

7. A debt collection case allocation system based on dynamic load balancing and intelligent weight calculation, characterized in that, include: The construction module is used to build a debt collector capability profile and a case feature database based on historical debt collection case records. The debt collector capability profile includes a matrix of the debt collector's language ability. The prediction module is used to extract case features from the case feature library of cases to be assigned, input them into the pre-trained prediction model, and extract shared features by the shared feature encoder of the prediction model. The shared features are input into the decoding branch of Task 1 of the prediction model to obtain the standardized effective working time of the cases to be assigned, which is used as the case weight. The shared features are concatenated with the language category code of the case to be assigned and then input into the decoding branch of the prediction model task two to obtain the language threshold of the case to be assigned under the corresponding language category, which is used as the language threshold of the case. The allocation module is used to construct an objective function with the weighted sum of collector overload and case allocation unevenness as the optimization objective. The objective function is solved using a greedy algorithm. During the solution process, for each case to be allocated, collectors with language ability values ​​not less than the language threshold of the case are selected to form a candidate set. Based on the candidate set of each case to be allocated, a preliminary case allocation result is obtained.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the collection case allocation method based on dynamic load balancing and intelligent weight calculation as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the collection case assignment method based on dynamic load balancing and intelligent weight calculation as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the collection case assignment method based on dynamic load balancing and intelligent weight calculation as described in any one of claims 1 to 6.

Images