A knowledge graph-based intelligent debt collection method, system, and storage medium

Abstract

The application relates to an intelligent arrears prompting method and system based on a knowledge graph, and a storage medium, the method comprising the following steps: constructing a knowledge graph; obtaining backup contact information based on the knowledge graph when the original contact information of a user to be prompted is invalid; constructing a user portrait of the user to be prompted; determining a prompting incentive scheme; matching at least one prompting strategy from a plurality of preset prompting strategies; using an artificial intelligence external call system to communicate with the user to be prompted according to the backup contact information and the matched prompting strategy; the application effectively solves the problem of low reach rate caused by user disconnection in traditional prompting; and based on the user portrait, the prompting incentive scheme is determined, and the accuracy and robustness of incentive recommendation are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and specifically to an intelligent debt collection method, system, and storage medium based on knowledge graphs. Background Technology

[0002] In industries such as telecommunications, finance, and public services, efficient and compliant debt collection from overdue customers is a core operational aspect of ensuring cash flow and reducing bad debt risk. With the development of artificial intelligence and big data technologies, existing intelligent debt collection systems are beginning to replace traditional linear operations such as manual outbound calls and mass SMS messaging. Current technologies primarily employ AI-based risk level quantification models to output a percentage-based risk score. Users are then categorized into high, medium, and low risk levels based on these scores, and corresponding collection scripts and strategies are automatically matched. The model predicts the probability of a user defaulting on their monthly payment, and then matches a suitable collection strategy from a pre-set strategy library based on this probability.

[0003] While using artificial intelligence to classify overdue users by risk level and predicting the probability of delinquency through models can improve the effectiveness of debt collection to some extent, the over-reliance on pre-registered phone numbers means that once a user cancels their account or their service is suspended, they become unreachable. The lack of effective contact recovery and multi-dimensional clue mining capabilities results in a large number of overdue users being unable to be contacted, ultimately leading to bad debts. This demonstrates the insufficient reach of existing collection methods for unreachable users. Furthermore, collection strategies generally employ the same model, failing to differentiate based on user type. Incentive programs lack targeting, making it difficult to match different strategies to different user types, resulting in low user acceptance rates. Additionally, most models are trained only once, unable to dynamically optimize strategies based on real-time user feedback, lacking a mechanism to discover potentially better solutions, ultimately resulting in a lack of refined, segmented collection strategies. Summary of the Invention

[0004] To address the technical problems in existing technologies, such as insufficient reach for lost users and a lack of refined stratification in debt collection strategies, this invention provides an intelligent debt collection method, system, and storage medium based on knowledge graphs.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A knowledge graph-based intelligent debt collection method includes the following steps: Obtain user data of users who have outstanding payments, perform feature engineering, and construct multi-dimensional feature vectors; Based on the user data, a knowledge graph containing the associated entities and relationships of the users to be urged to pay are constructed. When the original contact information of the users to be urged to pay becomes invalid, backup contact information associated with the users to be urged to pay is obtained based on the knowledge graph. Based on the user data and the multi-dimensional feature vector, a user profile is constructed for the user to be urged to pay; Based on the user profile, at least one collection incentive scheme is determined to be recommended to the user whose payment is pending. Based on the user profile and the recommended collection incentive scheme, at least one collection strategy is matched for the current user to be collected from a variety of preset collection strategies. Using an AI-powered outbound calling system, communication is conducted with the user to be contacted based on the backup contact information and the matched collection strategy.

[0006] The beneficial effects of this invention are as follows: when the original contact information of the user to be contacted becomes invalid, backup contact information associated with the user to be contacted is obtained based on a knowledge graph, effectively solving the problem of low reach rate caused by user loss of contact in traditional debt collection; and based on the user profile, a debt collection incentive scheme is determined, improving the accuracy and robustness of incentive recommendations; based on the user profile and the recommended incentive scheme, at least one debt collection strategy is matched for the current user to be contacted from a variety of preset debt collection strategies, ultimately improving the debt collection success rate.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, the knowledge graph is constructed using the associated entities of the users to be urged to pay as nodes and the relationships between the associated entities as edges.

[0009] Furthermore, obtaining backup contact information associated with the user to be urged to pay based on the knowledge graph includes the following steps: The Louvain algorithm is used to divide the knowledge graph into communities to identify close user groups that are closely connected with the users whose debts are to be collected. When searching for multiple candidate contact information within a preset number of jumps in the knowledge graph using a breadth-first search algorithm, priority is given to selecting multiple candidate contact information that belong to the same close user group as the user to be urged to pay. The PageRank algorithm was used to calculate credibility scores for the preferred candidate contact information. The candidate contact information is sorted in descending order of credibility score to form a candidate contact information list, so as to obtain the backup contact information.

[0010] Furthermore, based on the user data and the multi-dimensional feature vector, a user profile is constructed for the user to be urged to pay, and the specific steps are as follows: Based on the user data and the multi-dimensional feature vectors, a clustering algorithm is used to group the multi-dimensional feature vectors into multiple user group labels with different behavioral characteristics.

[0011] Furthermore, based on the user profile, at least one collection incentive scheme is determined to be recommended to the user whose payment is pending, including the following steps: Based on the user profile and user history behavior data, predict and rank the probability of the user to be urged to pay the debt for accepting different collection incentive plans, and output a ranked list of recommended incentive plans. At least one of the collection incentive schemes is identified from the sorted list of recommended incentive schemes.

[0012] Furthermore, combining the user profiles and historical user behavior data, the probability of acceptance of different collection incentive schemes by the users to be contacted is predicted and ranked, and a ranked list of recommended incentive schemes is output, including the following steps: The acceptance score of each collection incentive scheme for the user to be urged to pay is calculated by using a gradient boosting tree-based prediction algorithm, a collaborative filtering algorithm, and a matrix factorization algorithm, respectively. The acceptance scores output by different algorithms are weighted and fused to obtain the comprehensive acceptance score for each collection incentive plan; All collection incentive schemes are sorted from highest to lowest according to the comprehensive acceptance score, and a list of collection incentive schemes with a preset number of positions before sorting is output as the recommended incentive scheme list.

[0013] Furthermore, based on the user profile and the recommended collection incentive scheme, at least one collection strategy is matched for the current user to be contacted from a variety of preset collection strategies. The specific steps are as follows: Using the user context features, which include the user profile and the recommended collection incentive scheme, as input, the multi-armed slot machine algorithm is used to match the current optimal collection strategy for the user to be collected from a variety of preset collection strategies. The multi-armed slot machine algorithm is updated based on user feedback after the optimal collection strategy is executed.

[0014] Furthermore, updating the multi-armed slot machine algorithm based on user feedback after the execution of the matched optimal collection strategy includes the following steps: After the user to be reminded to pay the fee within the preset time window, the payment result is used as a reward signal. This signal is then combined with the user context feature vector when the optimal collection strategy is executed and the selected optimal collection strategy, and then updated to the multi-armed slot machine algorithm.

[0015] To address the aforementioned technical problems, this invention also provides an intelligent debt collection system based on a knowledge graph, the specific technical content of which is as follows: A knowledge graph-based intelligent debt collection system includes: The data access and processing module is used to acquire user data of users who need to pay their debts and perform feature engineering to construct multi-dimensional feature vectors. The knowledge graph construction and application module is used to construct a knowledge graph containing the associated entities and relationships of the user to be urged to pay the debt based on the user data. When the original contact information of the user to be urged to pay the debt becomes invalid, the module obtains the backup contact information associated with the user to be urged to pay the debt based on the knowledge graph. The user demand identification module is used to construct a user profile for the user to be urged to pay based on the user data and the multi-dimensional feature vector. The user incentive program recommendation module is used to determine at least one collection incentive program to recommend to the user to be collected, based on the user profile. The collection strategy matching and optimization module is used to match at least one collection strategy for the current user to be collected from a variety of preset collection strategies, based on the user profile and the recommended collection incentive scheme. The AI ​​robot outbound calling module is used to communicate with the user to be urged to pay using an artificial intelligence outbound calling system, based on the backup contact information and the matched collection strategy.

[0016] To address the aforementioned technical problems, this invention also provides an intelligent debt collection system based on a knowledge graph, the specific technical content of which is as follows: A storage medium storing a computer program or computer instructions, which, when executed by a computer's processor, implement the steps of the aforementioned knowledge graph-based intelligent debt collection method. Attached Figure Description

[0017] Figure 1 This is a flowchart of a knowledge graph-based intelligent debt collection method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a knowledge graph-based intelligent debt collection system in an embodiment of the present invention. Detailed Implementation

[0018] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0019] PageRank, also known as the PageRank algorithm or PageRank algorithm, is a link analysis algorithm proposed by Page and Brin. PageRank is a method used to identify the rank and importance of a webpage, and is an important metric for measuring a webpage's significance. In addition to considering the number of inbound links, the PageRank algorithm also takes into account webpage quality factors. By combining these two metrics, the evaluation of a webpage's importance is more accurate.

[0020] The LinUCB algorithm is a contextual bandit algorithm that incorporates contextual information and is widely used in recommender systems to balance the conflict between "exploration" and "exploitation." This algorithm models the expected click-through rate or revenue for each recommendation candidate by introducing contextual features of users and items (such as user age, device, historical behavior, and item content), and simultaneously calculates an upper confidence bound (UCB). Finally, it selects the item with the largest sum of expected revenue and the upper confidence bound for recommendation, thus ensuring recommendation effectiveness while giving new or low-exposure items a reasonable opportunity to be displayed.

[0021] The Louvain algorithm is a classic community detection algorithm based on modularity optimization, which aims to identify subgroups (communities) with tight internal connections and sparse external connections in complex networks.

[0022] Example 1 like Figure 1 As shown, this embodiment provides an intelligent debt collection method based on knowledge graphs, including the following steps: S1. Obtain user data of users to be urged to pay, and perform feature engineering to construct multi-dimensional feature vectors; Data extraction process: Comprehensively and accurately extract raw data related to debt collection business from multiple sources within the enterprise.

[0023] Data source integration involves connecting to customer relationship management systems, accounting systems, billing systems, customer service systems, historical overdue payment work order systems, and user online behavior log systems via API interfaces, direct database connections, and file imports.

[0024] Data types include basic attribute data such as user name, ID number, contact information, address, occupation information, etc.; billing and aging data such as overdue amount, overdue duration, aging structure, average payment cycle, commonly used payment channels, historical overdue records, etc.; and interaction behavior data such as historical outbound call time, connection status, communication duration, promised repayment status, customer service fault reporting records, and clickstream data from online channels.

[0025] Data processing: By cleaning, standardizing, performing derivative calculations and feature engineering on the raw data, it is transformed into a unified format suitable for knowledge graph construction and model algorithm analysis.

[0026] Data cleaning: Data deduplication is performed to remove duplicate data based on a unique user identifier (such as a user ID) to ensure the uniqueness of user profiles.

[0027] Data error correction processing includes correcting incorrectly formatted phone numbers and invalid addresses, and handling missing values ​​(e.g., by filling in missing values ​​with the mean or using interpolation).

[0028] Noise filtering removes obviously illogical abnormal data (such as payment records from future times).

[0029] Data standardization: The format is standardized, unifying date, time, amount, contact information and other data from different sources into a standard format (e.g., unified date format YYYY-MM-DD, unified amount unit "yuan").

[0030] Value range mapping process maps status codes from different business systems (such as "settled", "normal settlement", "overdue for 1-30 days") to a unified standard status code for debt collection.

[0031] Feature engineering: Calculation of derived indicators: Repayment ability characteristics are calculated by determining the user's average consumption / payment amount and bill aging status over the past 6 months.

[0032] The repayment willingness characteristic is used to construct a "user cooperation" score based on historical connection rate, commitment fulfillment rate, and complaint frequency.

[0033] Time characteristics, based on historical connection times to determine the best contact time periods and segmentation of overdue bill days.

[0034] Data discretization and normalization processes are performed, and continuous variables (such as the amount of arrears) are binned (e.g., divided into "0-100 yuan", "100-300 yuan", "300-500 yuan", etc.); data of different dimensions are normalized to meet the requirements of subsequent model input.

[0035] Data loading involves loading standardized data, which has undergone data cleaning, data transformation, and feature engineering, into a specific target storage system for use by upper-layer modules.

[0036] The data is loaded into a knowledge graph database, where entities such as users, identification documents, phone numbers, and addresses, along with their relationships, are stored as nodes and edges. These nodes serve as the raw material for constructing the knowledge graph. Neo4j can be used as the graph database.

[0037] The feature-engineered wide table data (i.e., user feature vectors) is loaded into a high-performance relational database for training and inference of the user requirement identification module. Oracle is used as the relational database.

[0038] S2. Based on the user data, construct a knowledge graph that includes the associated entities and relationships of the user to be urged to pay. When the original contact information of the user to be urged to pay becomes invalid, obtain backup contact information associated with the user to be urged to pay based on the knowledge graph. Specifically, the knowledge graph is constructed using the associated entities of the users to be urged to pay as nodes and the relationships between the associated entities as edges. Obtaining backup contact information associated with the user to be urged to pay based on the knowledge graph includes the following steps: The Louvain algorithm is used to divide the knowledge graph into communities to identify close user groups that are closely connected with the users whose debts are to be collected. When searching for multiple candidate contact information within a preset number of jumps in the knowledge graph using a breadth-first search algorithm, priority is given to selecting multiple candidate contact information that belong to the same close user group as the user to be urged to pay. The PageRank algorithm was used to calculate credibility scores for the preferred candidate contact information. The candidate contact information is sorted in descending order of credibility score to form a candidate contact information list, so as to obtain the backup contact information.

[0039] S201. Multi-source data integration and entity relationship definition; Data is acquired from the data access and processing module, defining the entity types of the knowledge graph and the types of relationships between them. Specifically, entity types include users, identification documents, phone numbers, fusion groups, addresses, etc., as detailed in Table 1. Table 1 Entity Type Configuration Table Relationship types include: possessing identification documents, having a reserved phone number, belonging to a group, and sharing an address.

[0040] S201. Graph Database Construction; The extracted entity types are used as nodes, and the relationship types are used as edges. These are stored in a graph database to construct a large-scale user relationship knowledge graph.

[0041] S203. Contact Repair and Trustworthiness Assessment In graph databases, hidden user connections are found through path search and community discovery. For example, when it is necessary to contact a target user in arrears, i.e., the aforementioned user (User A) whose payment is pending, the system uses its known information (such as ID number) as a starting point and executes a contact repair algorithm in the knowledge graph, specifically including: The path search employs a breadth-first search algorithm to find associated active phone numbers within N1 hops; N1 is a positive integer. For example, a typical 3-hop query path is: (User A), [possessing ID] to (ID X), [owner is] to (User B), [possessing phone number] to (phone number Y). Through this path, the phone number Y of user B, who uses the same ID as user A who has outstanding charges, can be found.

[0042] Community Detection: The Louvain algorithm is used to divide the graph into communities, identifying close-knit communities such as families and businesses. The Louvain algorithm measures the strength of the edge density within a community relative to a random network by optimizing modularity; a value closer to 1 indicates a more pronounced community structure. Its objective function is: in, This refers to the modularity of Louvain's algorithm. It represents the total number of edges in the knowledge graph. These are adjacency matrix elements (representing nodes) and (Is there an edge between them?) and These are nodes and The degree, and These are nodes and The community to which it belongs It is the Kronecker delta function; when hour, =1, otherwise It is 0.

[0043] For example, for user A who is in arrears, the edges in the graph are: (User A), [possessing ID] to (ID X), (ID X), [possessing ID] to (User B); (User A), [Shared address] to (User C); (User A), [call] to (User C); (User D), [call] to (User E); (User C) [calls] to (User F).

[0044] Louvain's algorithm partitioning process: First, each node is treated as an independent community. Then, attempts are made to move nodes to maximize the modularity increment. After iterative calculations, the following community partitioning is finally obtained: Community 1 (Family Group): Includes A, B, and C. They are closely connected through identification documents, addresses, and phone calls (multiple internal boundaries).

[0045] Community 2 (Colleagues Group): Includes D and E (D has spoken with E and has a weak relationship with A).

[0046] Community 3 (Other): Contains F, which may be alone or weakly associated with C, but C has been assigned to Community 1, so F is classified as Other.

[0047] In practice, due to the strong correlation between A and B / C, they are grouped into the same community, while F, which is only unilaterally connected to C, may be grouped into other communities or used as a boundary node. When user A, who is in arrears, becomes unreachable, attempts can be made to contact other active users within their community 1. Users outside the community (such as D, E, and F) are not prioritized for contact due to their weaker connections, to avoid harassing irrelevant individuals.

[0048] Credibility assessment uses the PageRank algorithm to calculate credibility scores for the searched candidate phone numbers. The core idea of ​​the PageRank algorithm is that a node's ranking depends on the number and ranking of other nodes pointing to it. Its core formula is: in, It is a node PageRank value (importance metric). It is the damping coefficient (usually taken as 0.85). This represents the total number of nodes in the graph. It points to a node The set of nodes, It is a node The degree of credibility. A number that is repeatedly linked to by multiple high-value users has a higher credibility score. For example: finding three candidate contact phone numbers through path search. , , , and It has a border. and It has a border, and and No borders. Therefore, the total number of nodes N = 3. According to the PageRank formula: Therefore, the confidence scores for the three candidate numbers are 0.05, 0.05, and 0.135, respectively.

[0049] S204. Output the list of candidate contact numbers; All found candidate phone numbers are sorted from highest to lowest credibility score to form a candidate contact number list for use by the subsequent AI robot outbound calling module. This list contains multiple candidate numbers, and the AI ​​robot will prioritize dialing the number with the highest credibility score when making outbound calls. For example, in the case above, the contact number will be dialed first. .

[0050] S3. Based on the user data and the multi-dimensional feature vector, construct a user profile for the user to be urged to pay; Specifically, based on the user data and the multi-dimensional feature vectors, a clustering algorithm is used to group the multi-dimensional feature vectors, forming multiple user group labels with different behavioral characteristics. This invention constructs diversified collection strategy recommendation models for different types of users, ensuring that the models can accurately match the collection needs of different users.

[0051] S301. Multi-dimensional Feature Engineering Based on the acquired user data, a multi-dimensional feature vector is constructed. Specifically, this includes: RFM features include: the time since the last payment, the number of payments within a certain time window (such as the past 12 months), and the cumulative payment amount within a preset time window.

[0052] Behavioral characteristics include: average monthly call duration, average monthly data usage, package type (prepaid or postpaid, package tier), network duration, number of historical overdue payments, and amount owed.

[0053] The strategy acceptance characteristics include: the payment conversion rate in response to historical collection methods (such as SMS, telephone, and APP push).

[0054] These features together form a high-dimensional feature vector, which can comprehensively characterize the user's behavior patterns and preferences.

[0055] S302. User Segmentation for Unsupervised Learning An unsupervised clustering algorithm is used to cluster the high-dimensional feature vectors of the users mentioned above. Specifically, the K-Means clustering algorithm is used, which divides users into a predetermined k groups by minimizing the Within-Cluster Sum of Squares (WCSS). Its objective function is: in, For the number of clusters, It is the first The number of samples in each cluster It is the first The th cluster One sample, For the first The centroid of a cluster, express The Euclidean distance.

[0056] The algorithm steps are as follows: S11. Random Initialization S12. Assign each sample to the nearest centroid, forming a centroid; S12. Count the number of clusters; S13. Recalculate the centroid of each cluster; S14. Repeat steps S12 to S13 until the centroid no longer changes. The number of clusters is... The contour coefficient is used to determine the contour coefficient in this embodiment. =6.

[0057] S303. Generate User Profile Feature analysis was performed on the six user groups obtained from clustering. The mean of typical indicators for each group was extracted and compared with the mean of all users. Significantly deviating features were identified as group identifiers. Based on this, semantic tags were assigned to each group to form standardized user profiles, as follows: User profile 1 consists of high-value, low-risk users, whose characteristics are shown in Table 2. This group of users has high consumption levels, long online time, very few historical arrears, strong regularity in payment behavior, and is particularly positive about APP push notifications.

[0058] Table 2 Characteristics of High-Value, Low-Risk Users User profile 2 is a price-sensitive user, and its characteristics are shown in Table 3. This group of users is highly sensitive to prices, has a low level of consumption, a long and unstable payment cycle, and a relatively average response time at payment touchpoints.

[0059] Table 3. Characteristics of Price-Sensitive Users User profile 3 represents users heavily reliant on data usage, as shown in Table 4. This group is primarily composed of young people and consumes a significant amount of data. They have a high response rate to payment reminders via app and phone.

[0060] Table 4. Characteristics of Users Heavily Dependent on Traffic User profile 4 consists of users who heavily rely on phone calls, and their characteristics are shown in Table 5. This group is mainly composed of middle-aged and elderly people who spend a lot of time online, have a strong demand for phone calls, and have a much higher response rate to telephone channels than other groups.

[0061] Table 5. Characteristics of Users Heavily Dependent on Calls User profile 5 represents users at potential churn risk, and their characteristics are shown in Table 6. This group has recently experienced a significant decline in activity, with all indicators at low levels, extremely long payment cycles, high frequency and amount of overdue payments, and generally low response rates across various collection channels.

[0062] Table 6. Characteristics of Users at Potential Churn Risk User profile 6 represents users with long-term arrears, whose characteristics are shown in Table 7. This group has multiple arrears records, with substantial outstanding amounts, raising questions about their willingness and ability to pay. Their activity level is extremely low, and various collection channels have yielded unsatisfactory responses.

[0063] Table 7 Characteristics of Users with Long-Term Overdue Payments S4. Based on the user profile, determine at least one collection incentive scheme to recommend to the user to be urged to pay; Specifically, it includes: Based on the user profile and user history behavior data, predict and rank the probability of the user to be urged to pay the debt for accepting different collection incentive plans, and output a ranked list of recommended incentive plans. The acceptance score of each collection incentive scheme for the user to be urged to pay is calculated by using a gradient boosting tree-based prediction algorithm, a collaborative filtering algorithm, and a matrix factorization algorithm, respectively. The acceptance scores output by different algorithms are weighted and fused to obtain the comprehensive acceptance score for each collection incentive plan; All collection incentive schemes are sorted from highest to lowest according to the comprehensive acceptance score, and a list of collection incentive schemes with a preset number of positions before sorting is output as the recommended incentive scheme list.

[0064] At least one of the collection incentive schemes is identified from the sorted list of recommended incentive schemes.

[0065] Based on the profile characteristics of users in arrears, the collection incentive plan with the highest acceptance rate is selected from a variety of incentive programs to improve user response rate and satisfaction during the collection process.

[0066] S401. Data Preparation: Construct a historical data matrix with users, incentive plans, responses, and conversion times as elements. The response refers to whether the incentive plan was accepted (1 if accepted, 0 otherwise). The conversion time records the number of hours elapsed from the push notification to the user's acceptance. An example data structure is as follows: User ID Incentive Plan ID Response (0 / 1) Conversion Time (hours) U001P001 (Data Package) 12 U001P002 (Phone Bill Discount) 0- U002P001 (Data Package) 0- U002P003 (Installment Payment) 1242.

[0067] S402. Model Training: The recommendation model is trained using gradient boosting tree algorithms (such as GBDT, XGBoost, LightGBM, CatBoost, GBRT, MART, etc.) with user and solution features as input and predicted user acceptance probability as output.

[0068] Model Input and Output: For any user-solution pair (a, b), the model input feature vector Output a continuous value ∈ [0,1], representing the probability that user accepts plan , i.e. For example, by inputting the feature vectors of user U001 and scheme P001 into a trained XGBoost model, the acceptance probability p=0.85 is calculated, indicating that the user is highly likely to accept this data package scheme.

[0069] Training process: The objective function is to minimize the log loss function. ,in, For real labels, To predict probabilities.

[0070] The core hyperparameters of XGBoost were tuned using 5-fold cross-validation. For example, a typical set of parameters tuned in a practical application is as follows: `params = { 'max_depth': 6, 'eta': 0.1, 'n_estimators': 200, 'min_child_weight': 1, 'subsample': 0.8, 'colsample_bytree': 0.8, 'gamma': 0, 'objective':'binary:logistic', 'eval_metric':'logloss', 'seed': 42}` This parameter configuration is represented as follows: Basic control parameters `max_depth=6`: The maximum depth of a single decision tree is 6, which controls model complexity and avoids overfitting. eta=0.1: Learning rate, controlling the weight of each tree's contribution to the final result; the smaller the value, the smoother the model. n_estimators=200: The number of decision trees trained is 200. More trees can improve the model's generalization ability. seed=42: Random seed, ensuring that the model training results are reproducible.

[0071] Regularization and sampling parameters min_child_weight=1: The minimum sum of the weights of the child nodes, used to control the minimum number of samples in the leaf nodes and prevent overfitting. subsample=0.8: Randomly sample 80% of the samples when training each tree to reduce model variance. colsample_bytree=0.8: Randomly sample 80% of the features when training each tree to improve the model's generalization ability. gamma=0: The minimum loss reduction for node splitting; the larger the value, the more conservative the model. Tasks and evaluation parameters objective='binary:logistic': Specifies a binary classification task, outputting probability prediction results. `eval_metric='logloss'`: Uses the log loss function as an evaluation metric to measure the difference between the model's predicted probabilities and the true labels.

[0072] S403. Collaborative Filtering: User-based collaborative filtering is employed as an auxiliary strategy for incentive recommendation, aiming to predict the acceptance level of target users for unfamiliar incentive schemes by leveraging the preferences of similar user groups. User similarity is calculated using cosine similarity. in, and There are two users. User and A set of incentive schemes that have been jointly evaluated (i.e., schemes in which both individuals have historical interaction records). User Incentive scheme In this scenario, the rating is directly based on the user's response to the solution: if the user accepts the solution, then... If a user was pushed to a notification but did not accept it, then For schemes that have not been pushed, they will not be included in the similarity calculation.

[0073] For example: Suppose there are 3 users (U001, U002, U003) and 4 incentive plans (P001 data package, P002 phone bill discount, P003 installment payment, P004 credit activation). The user plan rating matrix composed of historical interaction data is shown in Table 8. Table 8 User Solution Evaluation Matrix In Table 8, "?" indicates that user U001 has not yet been exposed to plan P004, and it is necessary to predict the probability of their acceptance of the plan. "1" indicates that the user accepted the plan, and "0" indicates that the plan was pushed but not accepted.

[0074] Step 1: Calculate the similarity between users. Taking users U001 and U002 as an example, calculate their common rating schemes. The corresponding rating vector is , The cosine similarity is calculated as follows: Similarly, the similarity between U001 and U003 is calculated as follows: Similarity between U002 and U003 Step 2: Select nearest neighbors. Choose the K users with the highest similarity to target user U001 as neighbors. Let K=2, then the neighbors are U003 (similarity 0.707) and U002 (similarity 0.5). Neighbor set .

[0075] Step 3: Predict the target user's rating for untouched solutions. For solution P004, which the target user U001 has not encountered, use the neighbor ratings for that solution for weighted prediction: The predicted score, ranging from 0 to 1, can be considered an estimate of the probability that user U001 will accept option P004. .

[0076] S404. Matrix Decomposition: Using singular value decomposition, the user-incentive scheme rating matrix is ​​transformed. Decomposed into a low-dimensional user latent vector matrix and the implicit vector matrix of the scheme Then the user Regarding the plan The predicted score can be calculated using the following formula: in, It is the global average score. and It is a bias term for users and solutions. and These are latent vectors. The optimal parameters are found by minimizing the loss function using optimization algorithms (such as stochastic gradient descent).

[0077] For example, in the above example, assuming that after training with SVD (let the hidden factor dimension k=2), the learned parameters are as follows: Global average score: =0.5 (Based on the average of all known ratings: (1+0+1+0+1+1+0+1+0+0+1) / 11=0.545, taken as 0.5 for simplicity) User bias: (U001 has a slightly higher acceptance rate than average) (U002's acceptance rate is slightly below average) (U003 acceptance rate is close to average) Scheme bias: (P001 data package is more popular) (P002 has low acceptance of phone bill discounts) (P003 installment repayment is slightly more acceptable) 2 (P004 has low acceptance rate for credit-based account activation) User latent vector (two-dimensional): =[0.8,0.3] =[0.2,0.9] =[0.7,0.4] Implicit vectors of the scheme (two-dimensional): =[0.9,0.2] =[0.3,0.8] =[0.6,0.5] =[0.5,0.6] Predict the probability (rating) of user U001 accepting option P004, i.e., calculate... Calculate the implicit vector inner product: =0.8×0.5+0.3×0.6=0.4+0.18=0.58 Substitute into the prediction formula: =0.5+0.1+(−0.2)+0.58=0.98; The predicted score is 0.98, higher than the 0.58 obtained from collaborative filtering mentioned earlier. This difference stems from the matrix factorization's utilization of the global structure: the latent vectors of U001 and P004 have a high matching degree (inner product 0.58), and U001 itself has a positive bias, resulting in a high overall score despite the negative bias of the alternative. This suggests that U001 may have a potential interest in P004. Similarly, it can predict U001's scores for other untouched alternatives (in this case, only P004 is missing), or help fill in missing values ​​for other users.

[0078] S405. Recommended Solution: In the incentive scheme recommendation module, the predicted score is output by matrix factorization. XGBoost probability Collaborative filtering score Weighted fusion is performed. For target users in arrears, a Top-N incentive list (such as data packages, call discounts, credit-based activation, installment payments, etc.) ranked by score is calculated and returned for use by the strategy matching module. For example, for the prediction of P004 by U001 in the previous example, the weighted fusion comprehensive score is: Score1 = 0.3 × 0.85 (assuming XGBoost output) + 0.3 × 0.58 + 0.4 × 0.98 ≈ 0.82. The prediction of other schemes is similar. Finally, the Top-N schemes are selected based on the comprehensive score and output to the strategy matching module. Top-N represents the top N items.

[0079] S5. Based on the user profile and the recommended collection incentive scheme, match at least one collection strategy for the current user to be collected from a variety of preset collection strategies; Specifically, based on the user profile and the recommended collection incentive scheme, at least one collection strategy is matched for the current user to be collected from a variety of preset collection strategies. The specific steps are as follows: Using the user context features, which include the user profile and the recommended collection incentive scheme, as input, a multi-armed slot machine algorithm is used to dynamically match the current optimal collection strategy for the user to be collected from a variety of preset collection strategies.

[0080] The multi-armed slot machine algorithm is updated based on user feedback after the execution of the dynamically matched collection strategy. The specific method for updating the multi-armed slot machine algorithm is as follows: After the user to be reminded to pay the fee within the preset time window, the payment result is used as a reward signal. This signal is then combined with the user context feature vector when the optimal collection strategy is executed and the selected optimal collection strategy, and then updated to the multi-armed slot machine algorithm.

[0081] The goal is to match each user with the optimal combination of debt collection strategies (including collection methods, timing, wording, and incentives) and continuously optimize them based on user feedback. The core of this approach is the use of a contextual multi-armed slot machine (MAB) algorithm to resolve the balance dilemma between "explore" and "exploit" in debt collection strategies, enabling real-time dynamic optimization of the strategies.

[0082] S501. Obtain user context information; Generate a rich context vector for each user to be contacted for debt collection. The context vector This includes: user profile tags (such as "high value, low risk", "price sensitive", etc.), real-time overdue payment information (overdue amount, overdue days), time characteristics (current time, day of the week), and Top-N incentive schemes output by the user incentive scheme recommendation module.

[0083] S502. Strategy selection based on the LinUCB algorithm; The "arms" of the LinUCB algorithm represent a complete, executable combination of debt collection strategies (e.g., "3 PM + AI call + standard script + 20GB data package incentive"). The LinUCB algorithm calculates a score for each "arm" using the following formula and selects the arm with the highest score to execute: Part One This represents "exploitation," which means selecting the strategy with the highest expected return based on current model predictions. Part Two It stands for "exploration," which means giving strategies whose models are not yet certain a chance to try them out in order to discover new strategies that may be better. The exploration coefficient (usually set to 0.2-0.5) enables the system to automatically find a balance between the known optimal strategy and exploring new strategies.

[0084] S503. Implement debt collection strategies; Based on the selected strategy The AI ​​robot outbound call module is invoked to execute the corresponding debt collection actions.

[0085] Observe user feedback and calculate rewards. Within a preset time window (e.g., 24 hours) after the strategy is implemented, observe whether users make payments. If they do, a reward is given. ,otherwise For partial payments, a continuous reward value can be set, such as... .

[0086] Real-time updates of MAB model parameters, including (context) The chosen strategy ,award The triplet is fed back to the LinUCB model, updating its parameters in real time. The specific update formula is as follows: For example, suppose there are two candidate policy arms: Arm A: [3 PM, AI call, standard script, 20GB data package] Arm B: [10 AM, AI call, gentle tone, 10% discount on call charges] User context (Two-dimensional; the first dimension is the user value score, and the second dimension is the normalized value of the number of days of overdue payment). Exploration Coefficient Initially, the parameters of both arms are... (Unit Array) Therefore .

[0087] Step 1: Initial state Arm A: , , Arm B: Similarly, , , During the first interaction, for user U001, calculate the score for arm A as: Utilization item + Exploration item. Arm B's score is exactly the same as Arm A's (because they started the same), at 0.1886. The system randomly selects Arm A. The Arm A strategy is executed; user U001 pays the full amount within 2 hours, and receives a reward. =1.

[0088] Step 2: Update the parameters of arm A. The parameters of arm B remain unchanged from the initial state.

[0089] Step 3: During the second interaction, for user U002, the context... Similar to the process above, calculate the score for arm A as 0.4625 + 0.1356 = 0.5981, and the score for arm B as 0 + 0.186 = 0.186. The system selects arm A with the higher score and executes the policy. If U002 fails to pay, the reward r = 0. As before, continue updating the parameters of arm A to allow the policy recommendation model to continuously evolve and become increasingly accurate through the closed-loop learning process described above.

[0090] S6. Using an AI-powered outbound calling system, communicate with the user to be contacted based on the backup contact information and the matching collection strategy.

[0091] By deploying AI-powered voice robots to replace human debt collectors, automated, large-scale, and humanized voice-based debt collection can be achieved.

[0092] The entire dialogue process is as follows: user voice input, Automatic Speech Recognition (ASR), Natural Language Understanding (NLU), Dialogue Management (DM), collection strategy matching and optimization module, Natural Language Generation (NLG), Text-to-Speech (TTS), and AI voice output. At any stage, if human intervention is detected, the system can immediately transfer the call to a human agent.

[0093] Example 2 like Figure 2 As shown, this embodiment provides an intelligent debt collection system based on knowledge graphs, including: The data access and processing module is used to acquire user data of users who need to pay their debts and perform feature engineering to construct multi-dimensional feature vectors. The knowledge graph construction and application module is used to construct a knowledge graph containing the associated entities and relationships of the user to be urged to pay the debt based on the user data. When the original contact information of the user to be urged to pay the debt becomes invalid, the module obtains the backup contact information associated with the user to be urged to pay the debt based on the knowledge graph. The user demand identification module is used to construct a user profile for the user to be urged to pay based on the user data and the multi-dimensional feature vector. The user incentive program recommendation module is used to determine at least one collection incentive program to recommend to the user to be collected, based on the user profile. The collection strategy matching and optimization module is used to match at least one collection strategy for the current user to be collected from a variety of preset collection strategies, based on the user profile and the recommended collection incentive scheme. The AI ​​robot outbound calling module is used to communicate with the user to be urged to pay using an artificial intelligence outbound calling system, based on the backup contact information and the dynamically matched collection strategy.

[0094] The data access and processing module, knowledge graph construction and application module, user demand identification module, user incentive scheme recommendation module, collection strategy matching and optimization module, and AI robot outbound call module can be program modules or computing devices with corresponding functions.

[0095] Example 3 Based on Embodiments 1 and 2, this embodiment addresses a user named "Zhang San" whose mobile phone number has been cancelled and who owes 200 yuan. The method described in Embodiment 1 is used to initiate a debt collection task for Zhang San. The specific process is as follows: S701. Contact Repair: The AI ​​robot's outbound calling module initially attempted to dial Zhang San's original number, 156********. When the attempt failed, the task flowed to the knowledge graph construction and application module. Starting with Zhang San's ID number, the system performed a Breadth-First Search (BFS) search within the knowledge graph, discovering an active number, 139********, under the same ID. This number belongs to user "Li Si". Using the Louvain community discovery algorithm, the system identified that Zhang San and Li Si belong to the same family community. The PageRank algorithm calculated the number's credibility to be 0.85, the highest score. The system then selected 139******** as its primary contact number.

[0096] S702. User Profile and Strategy Generation: Before initiating the outbound call, the user demand identification module analyzed Zhang San's historical data and found that his average monthly data usage was 15GB, classifying him as a heavy data user; his historical payment records showed that he actively responded to promotional activities, indicating that he is a price-sensitive user. The system tagged Zhang San with profiles such as "heavy data user" and "price-sensitive." The user incentive program recommendation module, based on an algorithm, provided the top 3 incentive programs with the highest scores.

[0097] S703. Strategy Matching: The LinUCB model in the debt collection strategy matching and optimization module combines the current context (3 PM, weekday, user profile, and recommendation incentives) to calculate the score of each strategy. The strategy "AI outbound calling + 3 PM + standard collection script + 20GB data package incentive" was found to have the highest score (0.65 for utilization item + 0.12 for exploration item = 0.77) and was selected for execution.

[0098] Intelligent outbound calling and negotiation: The AI ​​robot outbound calling module dials number 139******** and has a conversation with the person who answered, Li Si.

[0099] S704. Results Feedback and Model Optimization: Two hours later, the system detected that Zhang San's outstanding payment had been settled. The system recorded this collection as successful and awarded a bonus. The context of this interaction ,Strategy ,award The data was used to update the parameters of the LinUCB model. and This strengthens the model's confidence in the effectiveness of the "contact relatives + traffic incentives + afternoon time slot" strategy on similar users. The probability of this strategy being selected will increase the next time a similar situation arises.

[0100] Example 4 Based on Embodiment 1, this embodiment provides a storage medium storing a computer program or computer instructions. When the computer program or computer instructions are executed by the processor of a computer, the steps of the above-described intelligent debt collection method based on knowledge graph are implemented.

[0101] The storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0102] Example 5 Based on Embodiment 1, this embodiment provides a computer, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they implement the steps of the above-described intelligent debt collection method based on knowledge graphs.

[0103] The memory can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or RAM. The memory can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the memory can include both internal storage units and external storage devices of any data processing device. The memory is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the concept and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge graph-based intelligent debt collection method, characterized in that, Includes the following steps: Obtain user data of users who have outstanding payments, perform feature engineering, and construct multi-dimensional feature vectors; Based on the user data, a knowledge graph containing the associated entities and relationships of the users to be urged to pay are constructed. When the original contact information of the users to be urged to pay becomes invalid, backup contact information associated with the users to be urged to pay is obtained based on the knowledge graph. Based on the user data and the multi-dimensional feature vector, a user profile is constructed for the user to be urged to pay; Based on the user profile, at least one collection incentive scheme is determined to be recommended to the user whose payment is pending. Based on the user profile and the recommended collection incentive scheme, at least one collection strategy is matched for the current user to be collected from a variety of preset collection strategies. Using an AI-powered outbound calling system, communication is conducted with the user to be contacted based on the backup contact information and the matched collection strategy.

2. The intelligent debt collection method based on knowledge graphs according to claim 1, characterized in that, The knowledge graph is constructed using the associated entities of the users to be urged to pay as nodes and the relationships between the associated entities as edges.

3. The intelligent debt collection method based on knowledge graphs according to claim 1 or 2, characterized in that, Obtaining backup contact information associated with the user to be urged to pay based on the knowledge graph includes the following steps: The Louvain algorithm is used to divide the knowledge graph into communities to identify close user groups that are closely connected with the users whose debts are to be collected. When searching for multiple candidate contact information within a preset number of jumps in the knowledge graph using a breadth-first search algorithm, priority is given to selecting multiple candidate contact information that belong to the same close user group as the user to be urged to pay. The PageRank algorithm was used to calculate credibility scores for the preferred candidate contact information. The candidate contact information is sorted in descending order of credibility score to form a candidate contact information list, so as to obtain the backup contact information.

4. The intelligent debt collection method based on knowledge graphs according to claim 1, characterized in that, Based on the user data and the multi-dimensional feature vector, a user profile is constructed for the user to be urged to pay, and the specific steps are as follows: Based on the user data and the multi-dimensional feature vectors, a clustering algorithm is used to group the multi-dimensional feature vectors into multiple user group labels with different behavioral characteristics.

5. The intelligent debt collection method based on knowledge graphs according to claim 1, characterized in that, Based on the user profile, determine at least one collection incentive program to recommend to the user whose payment is pending, including the following steps: Based on the user profile and user history behavior data, predict and rank the probability of the user to be urged to pay the debt for accepting different collection incentive plans, and output a ranked list of recommended incentive plans. At least one of the collection incentive schemes is identified from the sorted list of recommended incentive schemes.

6. The intelligent debt collection method based on knowledge graphs according to claim 5, characterized in that, Based on the user profiles and historical user behavior data, the probability of acceptance of different collection incentive schemes by the users to be contacted for payment is predicted and ranked, and a ranked list of recommended incentive schemes is output, including the following steps: The acceptance score of each collection incentive scheme for the user to be urged to pay is calculated by using a gradient boosting tree-based prediction algorithm, a collaborative filtering algorithm, and a matrix factorization algorithm, respectively. The acceptance scores output by different algorithms are weighted and fused to obtain the comprehensive acceptance score for each collection incentive plan; All collection incentive schemes are sorted from highest to lowest according to the comprehensive acceptance score, and a list of collection incentive schemes with a preset number of positions before sorting is output as the recommended incentive scheme list.

7. The intelligent debt collection method based on knowledge graphs according to claim 5, characterized in that, Based on the user profile and the recommended collection incentive scheme, at least one collection strategy is matched for the current user to be contacted from a variety of preset collection strategies. The specific steps are as follows: Using the user context features, which include the user profile and the recommended collection incentive scheme, as input, a multi-armed slot machine algorithm is used to match the current optimal collection strategy for the user to be collected from a variety of preset collection strategies. The multi-armed slot machine algorithm is updated based on user feedback after the optimal collection strategy is executed.

8. The intelligent debt collection method based on knowledge graphs according to claim 7, characterized in that, The multi-armed slot machine algorithm is updated based on user feedback after the optimal collection strategy is executed, including the following steps: After the user to be reminded to pay the fee within the preset time window, the payment result is used as a reward signal. This signal is then combined with the user context feature vector when the optimal collection strategy is executed and the selected optimal collection strategy, and then updated to the multi-armed slot machine algorithm.

9. A system employing the knowledge graph-based intelligent debt collection method as described in any one of claims 1 to 8, characterized in that, include: The data access and processing module is used to acquire user data of users who need to pay their debts and perform feature engineering to construct multi-dimensional feature vectors. The knowledge graph construction and application module is used to construct a knowledge graph containing the associated entities and relationships of the user to be urged to pay the debt based on the user data. When the original contact information of the user to be urged to pay the debt becomes invalid, the module obtains the backup contact information associated with the user to be urged to pay the debt based on the knowledge graph. The user demand identification module is used to construct a user profile for the user to be urged to pay based on the user data and the multi-dimensional feature vector. The user incentive program recommendation module is used to determine at least one collection incentive program to recommend to the user to be collected, based on the user profile. The collection strategy matching and optimization module is used to match at least one collection strategy for the current user to be collected from a variety of preset collection strategies, based on the user profile and the recommended collection incentive scheme. The AI ​​robot outbound calling module is used to communicate with the user to be urged to pay using an artificial intelligence outbound calling system, based on the backup contact information and the matched collection strategy.

10. A storage medium, characterized in that, The storage medium stores a computer program or computer instructions, which, when executed by the computer's processor, implement the steps of the knowledge graph-based intelligent debt collection method as described in claims 1 to 8.

Images