A risk control system and method for bank loan business based on big data analysis

Abstract

This invention discloses a risk control system and method for bank loan business based on big data analysis, belonging to the field of financial risk control technology. The method includes: collecting and preprocessing real-time behavioral data to obtain a user behavior feature set; based on the user behavior feature set, calling a behavior graph modeling engine to perform structured graph construction, matching it with a risk anchor rule base, identifying potential risk patterns, and labeling initial anchor risk tags; using a value function optimization method to correct the deviation between the initial risk anchor tags and actual default records, and combining user historical behavior sample data and loan feedback data to predict the risk level score of each user's current behavior; predicting the probability distribution of future migration to a default state through user risk level scores and historical state evolution data; assessing the current loan business risk by combining the potential losses under each behavioral path, and generating risk control strategies through risk level mapping rules and a strategy decision engine.

Description

Technical Field

[0001] This invention relates to the field of financial risk control technology, and in particular to a risk control system and method for bank loan business based on big data analysis. Background Technology

[0002] In banking lending, loan risk control is a crucial part of financial risk management. Traditional methods primarily rely on financial statements and credit scores to build credit assessment models. However, with the development of big data and algorithmic capabilities, the industry is gradually adopting structured information, scoring card systems, and machine learning models to enhance risk identification capabilities. Existing risk control systems mostly use static statistical characteristics such as historical default records and credit ratings as the basis for risk judgment, combining fixed rules or predictive models to assess credit risk. These systems are widely used in pre-loan approval, loan monitoring, and post-loan management to meet the balance between regulatory requirements and business stability.

[0003] Traditional modeling methods have two limitations in the dynamic risk identification and scoring update of loan users: First, they rely on static variables or offline models, making it difficult to perform multi-dimensional modeling and graph fusion of real-time user behavior characteristics; second, state prediction is mostly based on linear migration settings of a single path, which makes it difficult to reflect the diversity and stages of risk evolution, affecting the accuracy and timeliness of strategy generation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a risk control method for bank loan business based on big data analysis to solve the problems of insufficient real-time behavioral risk identification and single risk state transition path.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] Firstly, this invention provides a risk control method for bank loan business based on big data analysis, comprising: collecting and preprocessing real-time behavioral data to obtain a user behavior feature set; based on the user behavior feature set, calling a behavior graph modeling engine to perform structured processing on the real-time behavioral data, constructing a multi-dimensional behavior graph and matching it with a risk anchor rule base to obtain potential risk behavior patterns, while labeling initial risk anchor labels; based on the initial risk anchor labels, using a value function optimization method in reinforcement learning to correct the deviation between the initial risk anchor and actual default records, and combining user historical behavior sample data and loan feedback data to predict the risk level score of each user's current behavior; combining historical state evolution data and the risk level score of the user's current behavior to predict the probability distribution of future migration to a default state; based on the probability distribution of users migrating to a default state and the potential losses under each behavioral path, obtaining the user's loan business risk at the current time, and generating a risk control strategy through risk level mapping rules and a strategy decision engine.

[0008] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the real-time behavioral data includes online behavioral data generated by users in bank loan services.

[0009] The preprocessing includes data cleaning, timestamp correction, redundant field removal, and behavior normalization.

[0010] Based on preprocessed real-time behavior data, user behavior features are extracted through aggregated statistical analysis, and a user behavior feature set is generated.

[0011] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the steps of calling the behavior graph modeling engine to perform structured processing on real-time behavior data, constructing a multi-dimensional behavior graph and matching it with the risk anchor point rule base are as follows:

[0012] Based on user behavior feature sets and combined with a behavior graph modeling engine, user operation behavior data at different time points are marked and organized in a graph data structure in the form of event nodes. Edges are constructed on the logical and temporal relationships between nodes to obtain a basic behavior graph based on nodes and edges.

[0013] Extract substructure patterns related to risky behaviors from the basic behavior graph, perform pattern matching with the risk anchor rule base, and identify high-risk structures in the basic behavior graph.

[0014] By using a pattern fitting algorithm, risk anchor points are marked on the behavioral paths and nodes of high-risk structures to obtain a multi-dimensional behavioral map containing preliminary risk labels.

[0015] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the steps for predicting the risk level score of each user's current behavior are as follows:

[0016] Based on the multidimensional behavioral graph, the risk scoring processing logic is invoked to perform statistical analysis on the correlation between the initial risk labels and the user's historical loan performance, and to predict the historical default probability of each type of risk anchor label in the historical loan sample.

[0017] The value function optimization method in reinforcement learning algorithm is adopted to form a state-action-feedback triplet for the user's credit decision actions and subsequent actual credit results under high-risk structure.

[0018] By constructing a value function that takes user behavior status as input and actual credit results as feedback, the performance deviation of the current risk scoring processing logic in long-term risk prediction is evaluated.

[0019] If the performance deviation exceeds the risk scoring tolerance threshold, the initial risk label will be corrected by adjusting the risk scoring processing logic.

[0020] By integrating the revised risk anchor labels and temporal sequences in the behavioral graph with high-risk structures, a comprehensive behavioral feature vector characterizing the user's current behavioral risk status is generated. A temporal learning method based on a long short-term memory network model is then used to predict the user's risk level score at the current behavioral risk level.

[0021] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the steps for predicting the probability distribution of future migration to a default state by combining historical state evolution data and risk level scores of current user behavior are as follows:

[0022] Based on the risk level score and the business attributes of the current loan account, the credit status of the current loan account is obtained;

[0023] Using a multi-state Markov modeling method, based on historical account state evolution data and the current credit status of loan accounts, a transition matrix for different loan risk states is established, and the path distribution and probability distribution of users migrating from the current credit status to various target risk states in future time periods are dynamically calculated.

[0024] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the steps for obtaining the user's loan business risk at the current point in time based on the probability distribution of users migrating to a default state and the potential losses under each behavioral path are as follows:

[0025] Based on the probability distribution, obtain the cumulative probability of migrating to a high-risk structural state;

[0026] By combining the potential credit loss value corresponding to each migration path, a risk assessment calculation is performed to obtain the expected credit loss value of the loan account;

[0027] By fusing expected credit loss values ​​with probability distributions, a comprehensive loan risk score for the user at the current point in time is generated.

[0028] As a preferred embodiment of the risk control method for bank loan business based on big data analysis described in this invention, the steps for generating a risk control strategy through risk level mapping rules and a strategy decision engine are as follows:

[0029] The risk level mapping rules and strategy decision-making logic are used to map the risk score of loan business to a risk level range;

[0030] Based on the mapping results, corresponding strategy templates are matched according to different risk levels to generate an initial risk control strategy. This strategy is then dynamically modified by combining user behavior feature sets to generate a personalized risk control strategy.

[0031] Secondly, the present invention provides a risk control system for bank loan business based on big data analysis, including a behavior extraction module, a graph modeling module, a scoring correction module, a risk prediction module, and a strategy generation module;

[0032] The behavior extraction module is used to collect real-time behavior data and preprocess it to obtain a set of user behavior features;

[0033] The graph modeling module is used to call the behavior graph modeling engine to perform structured processing on real-time behavior data based on user behavior feature sets, construct multi-dimensional behavior graphs and match them with the risk anchor rule base to obtain potential risk behavior patterns, and label the initial risk anchor labels.

[0034] The scoring correction module is used to correct the deviation between the initial risk anchor point and the actual default record based on the initial risk anchor point label and the value function optimization method in reinforcement learning. It also combines user historical behavior sample data and loan feedback data to predict the risk level score of each user's current behavior.

[0035] The risk prediction module is used to combine historical state evolution data and the risk level score of the user's current behavior to predict the probability distribution of the future migration to the default state;

[0036] The strategy generation module is used to obtain the user's loan business risk at the current point in time based on the probability distribution of the user's migration to the default state and the potential loss under each behavioral path, and to generate risk control strategies through risk level mapping rules and strategy decision engine.

[0037] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the risk control method for bank loan business based on big data analysis as described in the first aspect of the present invention.

[0038] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the risk control method for bank loan business based on big data analysis as described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are as follows: Through the graph modeling step, a structured representation of user behavior data is achieved. By matching a risk anchor rule base, high-risk behavior patterns are identified and initial risk anchor labels are assigned, enhancing the targeting and accuracy of risk identification. Through the scoring correction step, a state-action-feedback triplet is constructed, and a value function is used to optimize the deviation between the initial risk labels and the loan results, achieving dynamic adjustment of the risk scoring logic. Ultimately, this improves the accuracy and adaptability of risk scoring, enhancing the intelligent perception and control capabilities for loan behavior risks. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of a risk control method for bank loan business based on big data analysis.

[0042] Figure 2 This is a schematic diagram of a risk control system for bank loan business based on big data analysis.

[0043] Figure 3 A flowchart for behavioral graph modeling and risk anchor point labeling.

[0044] Figure 4 A flowchart for risk level scoring correction and prediction. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-4 This is one embodiment of the present invention, which provides a risk control method for bank loan business based on big data analysis, including the following steps:

[0049] S1. Collect real-time behavioral data and preprocess it to obtain a user behavior feature set;

[0050] Furthermore, real-time behavioral data includes online behavioral data generated by users in bank loan services, including login frequency and time period distribution, number of times and dwell time when browsing loan products, operation paths (such as clicking, swiping, filling out forms), loan simulation calculation behavior, timing of application submission, and exit behavior during the submission process.

[0051] Real-time behavioral data is collected in real time through user interaction tracking components embedded in bank clients, apps, and web pages, and uploaded through an encrypted channel.

[0052] Preprocessing includes data cleaning, timestamp correction, redundant field removal, and behavior normalization.

[0053] Furthermore, the collected real-time behavioral data suffers from redundancy, multi-source heterogeneity, and temporal inconsistencies. The following preprocessing steps are employed:

[0054] Data cleaning: Remove records with format errors, missing fields, or logical inconsistencies, including behavior logs where the login time is later than the operation time;

[0055] Timestamp correction: The timestamps of user behavior records are uniformly converted to the bank's standard time zone GMT+8 and stored based on the unified format ISO 8601;

[0056] Redundant field removal: Remove fields that are not directly related to risk control, including APP version number and device brand;

[0057] Behavior normalization and standardization: Performing maximum and minimum normalization on behavior frequency, time, and duration;

[0058] Z-score standardization is used to ensure the uniformity of feature distribution and eliminate the influence of dimensions between user behavior dimensions.

[0059] Based on preprocessed real-time behavior data, user behavior features are extracted through aggregated statistical analysis, and a user behavior feature set is generated.

[0060] Furthermore, based on the preprocessed real-time user behavior data, aggregated statistical analysis is performed using a big data analytics framework to extract multi-dimensional user behavior features with discriminative and risk-predictive capabilities, and a standardized user behavior feature set is constructed. The entire process includes the following key steps:

[0061] The preprocessed user behavior data is aggregated and statistically analyzed. It is grouped and aggregated by user ID, and combined with sliding time windows and behavior type tags to extract various user behavior characteristics, including: behavior density (daily / weekly login frequency, monthly loan product browsing frequency, and total operation frequency in the past 7 days); behavior intensity characteristics (maximum number of clicks in a single session, total number of pages visited daily and volatility, and number of steps taken before submitting an application); dwell time characteristics (average page dwell time, dwell time on the loan contract description page, and browsing depth, referring to the number of page levels browsed by the user after entering from the entry page); activity time characteristics (main periods of active behavior, presence of late-night operations, e.g., 11 PM to 5 AM the next day); operation concentration (whether behavior is highly concentrated in a certain period); and path behavior characteristics (common operation path statistics, including "homepage → loan page → credit assessment → exit"), operation path jump frequency and interruption point analysis (indicating at which steps the user interrupts the behavior process).

[0062] Various user behavior features are transformed and encoded to obtain multi-dimensional behavioral features. Feature transformation and encoding include one-hot encoding and label encoding for discrete value features; continuous numerical features retain their original values ​​or undergo normalization and scaling to ensure that all features are on a uniform numerical scale when input into the model.

[0063] Multidimensional behavioral features are combined in the form of structured vectors to generate user behavior feature sets. The features of user behavior feature sets include: using user ID as the primary key index; unified naming conventions for all feature dimensions; support for adding new features or dynamically updating existing feature values ​​to meet real-time risk scoring requirements; and horizontal integration with user profiles, credit scores, and account history features.

[0064] S2. Based on the user behavior feature set, call the behavior graph modeling engine to perform structured processing on real-time behavior data, construct a multi-dimensional behavior graph and match it with the risk anchor rule base to obtain potential risk behavior patterns, and label the initial risk anchor labels at the same time.

[0065] Based on user behavior feature sets and combined with a behavior graph modeling engine, user operation behavior data at different time points are marked and organized in a graph data structure in the form of event nodes. Edges are constructed on the logical and temporal relationships between nodes to obtain a basic behavior graph based on nodes and edges.

[0066] Furthermore, key user behaviors at different points in time are modeled to obtain user operation behavior data, which is marked as an event node. The user operation behavior data includes login, browsing, clicking, loan application, redirection, and logout. Each event node contains information fields including behavior type, behavior occurrence time, device, network, geographical location, and behavior risk weight.

[0067] It should be noted that modeling refers to abstracting user actions into event nodes with multi-dimensional attributes such as behavior type, time, and device in chronological order, and representing the behavior sequence in a standardized structure.

[0068] Each user's sequence of actions within a business process is abstracted into several time-series event nodes, which serve as the basic units of the graph.

[0069] Two edge types are constructed between nodes: temporal edges represent the chronological order of user actions, and logical edges represent the business logic dependencies and path associations between actions.

[0070] Edge structure attributes include edge type, trigger interval (i.e., the time difference between actions), and path number;

[0071] Based on the construction results of event nodes, logical edges, and temporal edges, a basic behavior graph is formed, which reflects the node-edge structure relationship of operation behavior on a user-by-user basis.

[0072] Extract substructure patterns related to risky behaviors from the basic behavior graph, perform pattern matching with the risk anchor rule base, and identify high-risk structures in the basic behavior graph.

[0073] Furthermore, a risk anchor rule base is defined. Based on behavioral trajectory analysis of a large number of historical risk samples, the risk anchor rule base uses structured rule templates to summarize risky behaviors. It combines this with a behavioral graph modeling engine to match and identify graph structure levels, including the following risky behaviors: high-frequency operation anchors, such as a user continuously querying loan limits three times within one minute; path reversal anchors, i.e., a user repeatedly performing back operations or frequently jumping between pages; late-night operation anchors, such as a user performing critical loan operations between 00:00 and 04:00; abnormal path anchors, i.e., skipping key information pages and directly submitting a loan application; and abnormal time difference anchors, i.e., a user repeatedly changing their IP address and geolocation and quickly completing the application process.

[0074] In the basic behavior graph, the subgraph isomorphic matching algorithm is used to search for substructures that are equivalent to the risk anchor structure in the risk anchor rule base;

[0075] If a substructure meets the risk behavior criteria, it is determined to be a high-risk structure, and the user ID, behavior path, and risk type are recorded.

[0076] By using a pattern fitting algorithm, risk anchor points are marked on the behavioral paths and nodes of high-risk structures to obtain a multi-dimensional behavioral map containing preliminary risk labels.

[0077] Furthermore, for the identified high-risk substructures, a pattern fitting algorithm is used for fitting and identification. The behavioral paths and behavioral nodes corresponding to the high-risk substructures are located in the basic behavioral map. Risk anchor labels are assigned to the risky behavioral nodes in the path, including "suspicious clicks", "avoidance information", and "late-night operations".

[0078] By retaining the basic behavior graph structure and only adding risk anchor point annotations, a multidimensional behavior graph containing preliminary risk labels is generated. The graph structure of the multidimensional behavior graph includes original behavior nodes, basic edge relationships, and preliminary risk labels annotated with risk anchor points.

[0079] S3. Based on the initial risk anchor label, the value function optimization method in reinforcement learning is used to correct the deviation between the initial risk anchor and the actual default record. Combined with user historical behavior sample data and loan feedback data, the risk level score of each user's current behavior is predicted.

[0080] Furthermore, based on the multidimensional behavioral graph, the risk scoring processing logic is invoked to perform correlation statistical analysis between the initial risk labels and the user's historical loan performance, and to predict the historical default probability of each type of risk anchor label in the historical loan sample.

[0081] It should be noted that the risk scoring processing logic includes: Risk scoring rule definition: setting initial scoring weights for each type of risk anchor tag; for example, the weight of the "skip information disclosure" tag is 8 points, "late-night operations" is 5 points, and "frequent credit limit inquiries" is 2 points; the position, repetition frequency, and co-occurrence relationship of the risk anchor tag in the behavioral path are also used as scoring factors. Historical statistical modeling: based on historical loan samples, the default probability of each tag is statistically analyzed to correct the initial weights, making the risk score closer to the actual risk level; Scoring calculation mechanism: scoring each user graph structure in the multi-dimensional behavioral graph, converting individual and combined risk anchor tags into numerical scores, comprehensively calculating the risk level, and outputting credit recommendations. Risk score correction and feedback optimization:

[0082] A reinforcement learning method is used to evaluate the deviation between the scoring results and the actual default results. If the scoring deviation exceeds the tolerance threshold, the scoring rules or model parameters are adjusted.

[0083] The risk scoring processing logic is invoked to perform a scoring operation on the initial risk anchor point labels marked in the structure of each user behavior node in the multi-dimensional behavior graph, resulting in a risk anchor point score. The scoring dimensions include: the severity of the initial risk anchor point label (e.g., "frequent credit limit inquiries" has a lower weight than "skipping information disclosure"); the position of the initial risk anchor point in the path (e.g., the beginning / end / key turning point of the path); and the number of times similar risk anchor points appear repeatedly.

[0084] It obtains users' historical loan performance data from existing bank credit approval databases and post-loan management databases, possesses complete user loan records and result tags, and establishes a correspondence between user IDs and user behavior feature sets in multi-dimensional behavior graphs.

[0085] The frequency of each type of risk anchor tag in historical credit behavior sample data is statistically analyzed, and the historical default probability corresponding to each type of risk anchor tag in historical credit behavior sample data is obtained. The risk anchor scores in the user's current multidimensional behavior graph are weighted and summarized as a preliminary risk score.

[0086] The value function optimization method in reinforcement learning algorithm is adopted to form a state-action-feedback triplet for the user's credit decision actions and subsequent actual credit results under high-risk structure.

[0087] By constructing a value function that takes user behavior status as input and actual credit results as feedback, the performance deviation of the current risk scoring processing logic in long-term risk prediction is evaluated.

[0088] If the performance deviation exceeds the risk scoring tolerance threshold, the initial risk label will be corrected by adjusting the risk scoring processing logic.

[0089] Furthermore, a triplet is formed, consisting of a high-risk structure as the state, an initial risk score generated by the risk scoring processing logic as the credit decision-making action, and subsequent actual credit results as feedback. The subsequent actual credit results refer to the user's repayment performance in subsequent actions, including default and on-time repayment.

[0090] The value function model from reinforcement learning is used to evaluate the performance bias of the current risk scoring logic in long-term risk prediction. The performance bias formula is as follows:

[0091]

[0092] Wherein, ΔV represents the performance deviation; This represents the historical default risk return predicted based on the risk scoring processing logic; This represents the actual risk and return calculated based on real feedback; Indicates the user's behavior status. Indicates based on the current state The actions performed Indicates the execution of an action Then, feedback is returned based on the actual results;

[0093] It should be noted that the risk scoring tolerance threshold is set according to the risk management strategy. For example, the probability of non-risk users being misjudged as high-risk is set to ≤5%, and the acceptable scoring deviation is ≤±10 points.

[0094] The risk score tolerance is derived by working backward from the business's tolerance level. For example, a risk score tolerance greater than 15 points is considered unacceptable.

[0095] If the performance deviations calculated in multiple triplets exceed the risk score tolerance threshold, the risk score processing logic correction process will be automatically triggered, including adjusting the risk level score interval boundaries, for example, re-dividing the risk level score interval (0–100).

[0096] By integrating the revised risk anchor labels with the temporal and high-risk structures in the multidimensional behavior graph, a comprehensive behavioral feature vector is generated that characterizes the user's current behavioral risk status. A risk scoring method based on a weighted sliding window is then used to predict the user's risk level score at the current behavioral risk level.

[0097] Furthermore, the user's operational behavior information in the multidimensional behavior graph structure is vectorized, including risk anchor tag encoding, multidimensional behavior graph structure feature extraction, and temporal behavior attribute fusion.

[0098] Among them, the risk anchor label encoding includes structuring the risk anchor labels marked in the multidimensional behavior map, including label type, risk weight, and positional features in the path, and forming a vector representation by using one-hot encoding and weight normalization.

[0099] Multidimensional behavioral graph structural feature extraction includes statistical analysis of structural indicators in the user's current graph, including graph attribute features such as operation path length, number of behavioral nodes, jump frequency, number of loop structures, and average time interval between behavioral nodes.

[0100] The risk anchor tag encoding information, the information extracted from the multi-dimensional behavior graph structure features, and the temporal behavior attribute fusion information are fused to generate a unified format comprehensive behavior feature vector, which serves as a numerical representation of the user's behavior state at the current time point.

[0101] It should be noted that the weighted sliding window scoring method obtains a trend score of behavioral status changes by weighting and averaging user behavior characteristics across multiple consecutive time segments, reflecting the degree of potential credit risk at the current point in time.

[0102] S4. Combining historical state evolution data and risk level scores of current user behavior, predict the probability distribution of future migration to a default state;

[0103] Based on the risk level score and the business attributes of the current loan account, the credit status of the current loan account is obtained;

[0104] Furthermore, based on risk level scoring, combined with the loan account's overdue days, payment period, and settlement status, the user's current risk level is transformed into a corresponding credit status probability distribution quantification of the user's credit risk level at the current point in time through rule mapping; according to the current loan account's business attribute data, including payment period, overdue days, and settlement status, the account status is divided into multiple discrete states, including normal, watchlist, minor overdue, major overdue, default, and settlement;

[0105] Using a multi-state Markov modeling method, based on historical account state evolution data and the current credit status of loan accounts, a transition matrix for different loan risk states is established, and the path distribution and probability distribution of users migrating from the current credit status to various target risk states in future time periods are dynamically calculated.

[0106] Furthermore, historical migration path statistics are performed. Based on users' historical credit behavior records, the state label sequence of each loan account at different points in time is extracted in chronological order to construct the state transition path. By aggregating the state sequences of different users, the migration frequency between different loan states is counted, and the natural trend of state change is analyzed.

[0107] The state transition probability is calculated by taking each loan state as the initial state and calculating the conditional probability of transitioning to each target state, as shown in the following formula:

[0108]

[0109] in, Indicates the user's current state Migrate to state in the next time period The conditional probability; Represents a probability function; Indicates the user's time The quantitative value of the loan status at any given moment; Indicates the user's time The specific value of the loan status at any given moment; This indicates a specific value representing the current period's loan status; A quantitative value representing the loan status for the next cycle;

[0110] It should be noted that the status and All are determined by the loan account's payment period, overdue days, and settlement status at two consecutive points in time, and are categorized into six discrete credit statuses based on the loan account's historical loan performance data. For example:

[0111] Normal (payments made on time, no overdue payments)

[0112] Pay attention to (those with a tendency to be overdue or overdue by 1-7 days).

[0113] Mild overdue (8–30 days overdue)

[0114] Severe delinquency (31–90 days overdue)

[0115] : Default (overdue for more than 90 days or written off)

[0116] : Settled (the loan has been fully settled either ahead of schedule or on schedule)

[0117] Historical account status evolution data refers to the status sequence data of a loan account recorded in chronological order throughout the entire credit cycle, consisting of payment terms, overdue status, and settlement status. It is used to describe the entire process of account status evolution over time.

[0118] Construct a state transition matrix and count the frequency of historical transitions between state pairs. For example, for each pair of states... From the statistical historical data, from the state Migration to state The number of times;

[0119] Calculate conditional probabilities, for example: divide the frequency of each state pair by the number of states. The total number of transitions to all states is obtained. ;

[0120] Fill the transition matrix, for example: all Fill the matrix according to the state combination order to obtain a The state transition matrix, where each row represents a starting state and each column represents a target state;

[0121] Verify the validity of the matrix, for example, by ensuring that the sum of all transition probabilities in each row is 1, satisfying the properties of a Markov chain.

[0122] The probability distribution of future transitions can be obtained from the state transition matrix. For example, when a user is currently in a certain state... When, obtain the first state in the state transition matrix. The data from the row is directly used as the probability distribution of the user's migration to various states in the next time period.

[0123] The probability distribution describes the likelihood that a user will migrate to a "mildly overdue", "severely overdue", or "default" risk state in the future, starting from the current state. It is an important basis for assessing default risk.

[0124] S5. Based on the probability distribution of users migrating to a default state and the potential losses under each path, obtain the user's loan business risk at the current point in time, and generate risk control strategies through risk level mapping rules and strategy decision engine.

[0125] Based on the probability distribution, obtain the cumulative probability of migrating to a high-risk structural state;

[0126] Furthermore, paths leading to high-risk structural states are selected from the state transition paths, and the state transition probabilities of the selected paths are accumulated to obtain the cumulative probability of a user migrating to a high-risk state in the future.

[0127] It should be noted that the state transition path is a recursive sequence of state evolution based on the user's current loan status and the conditional probabilities given in the state transition matrix.

[0128] By combining the potential credit loss value corresponding to each state transition path, a risk assessment calculation is performed to obtain the expected credit loss value of the loan account;

[0129] Furthermore, a corresponding potential credit loss value is associated with each state transition path. The potential credit loss value is obtained by modeling based on historical data and reflects the principal loss, interest loss and disposal costs that occur under each transition path.

[0130] It should be noted that the potential credit loss value is obtained by statistically analyzing the average principal loss corresponding to each state transition path in the historical sample when it occurs, thereby establishing a one-to-one quantitative relationship with each transition path.

[0131] The probability of a path occurring refers to the historical probability of a user's current state transitioning to a target state within a future time period. It is obtained based on statistics of the state evolution trajectories of a large number of loan accounts and reflects the likelihood of different state transitions occurring in actual business.

[0132] To comprehensively assess a user's credit risk at the current point in time, the expected credit loss value of the loan account is calculated by combining the probability of each path occurring and the potential credit loss value corresponding to each path, using the following formula:

[0133]

[0134] in, This represents the expected credit loss value; Indicates the first The cumulative probability of a state transition path is the product of the conditional probabilities of the user passing through each state in the state transition path from the current state, reflecting the actual likelihood of this path occurring. Indicates the first The potential credit loss value corresponding to each state transition path; This represents the number of all reachable paths.

[0135] By fusing expected credit loss values ​​with probability distributions, a comprehensive loan risk score for the user at the current point in time is generated.

[0136] The risk level mapping rule is based on the comprehensive loan business risk score, and according to the preset score range, it maps the user's current credit risk level to multiple discrete risk levels.

[0137] The strategy decision-making logic is to match the risk response strategies defined in the corresponding strategy template library according to the user's risk level, and to execute automatic decisions in combination with business objectives (such as controlling the default rate and maintaining the loan disbursement rate).

[0138] The risk level mapping rules and strategy decision-making logic are used to map the risk score of comprehensive loan business to a risk level range;

[0139] Based on the mapping results, corresponding strategy templates are matched according to different risk levels to generate an initial risk control strategy. This strategy is then dynamically modified by combining user behavior feature sets to generate a personalized risk control strategy.

[0140] Furthermore, based on the expected credit loss value, and integrating core risk factors, a comprehensive loan risk score for the user at the current point in time is generated. Core risk factors include the cumulative probability of the user migrating to a high-risk state, the strength and number of risk anchor points in the behavioral graph, and the user's recent risk level score. The expected credit loss value refers to the weighted result of the potential credit loss under different state migration paths and the probability of that path occurring.

[0141] In the process of generating a comprehensive loan risk score, a linear weighted function is used, with the expected credit loss value as the main axis. Core risk elements are numerically normalized and then merged into a unified risk score input vector, generating a standardized, continuous numerical comprehensive loan risk score. The output of the comprehensive loan risk score is a standardized continuous value used for subsequent risk level classification. The higher the comprehensive loan risk score, the greater the loan risk under the user's current behavioral state.

[0142] Using the comprehensive loan business risk score as input, risk level mapping rules are set to divide the comprehensive business risk score into fixed risk level ranges, for example:

[0143] Rating range Risk level 0–30 Low risk 31–60 Medium risk 61–80 Higher risk 81–100 High risk

[0144] Based on the specific range of the risk level score, the mapping logic function in the strategy decision engine is invoked to determine the user's risk level label.

[0145] Based on risk level tags, the system retrieves the corresponding initial strategy template from the strategy rule base. Each initial strategy template contains risk control strategies set for different risk levels, including low-risk, medium-risk, higher-risk, and high-risk templates. The low-risk template includes automatic approval, normal credit limit, and no manual intervention required; the medium-risk template includes reducing the credit limit and restricting the loan term; the higher-risk template includes suspending automatic approval and transferring to manual review; and the high-risk template includes rejecting credit, adding the applicant to the post-loan monitoring list, and providing early warnings.

[0146] It should be noted that the strategy rule base is comprehensively formulated by the bank's risk control department based on long-term credit business experience, analysis results of historical default sample data, policy compliance requirements, and risk tolerance modeling rules. It can be continuously optimized through a combination of manual rule setting and machine learning model recommendation.

[0147] Based on the template content in the user risk level tag matching strategy rule base, and combined with the current user behavior characteristics and account status, the initial risk control strategy is generated by dynamically adjusting the parameters of each strategy, including credit review rules, credit limit control, post-loan management, and collection priority.

[0148] Based on the initial risk control strategy, the strategy parameters are dynamically adjusted by combining user behavior feature sets to achieve personalized strategies. The adjustment logic includes: if the user's recent behavior is stable and their credit performance is good, the control intensity is appropriately relaxed and part of the credit limit is restored; if there are abnormal structures in the behavior graph, such as skipping key information or making multiple attempts in a short period of time, the risk control level is increased or the restriction policy is triggered in advance; if the user's device fingerprint changes frequently or the operation time is abnormal, additional verification steps are added or the loan cycle is shortened.

[0149] We provide personalized risk control strategies, including multi-dimensional risk control rules covering approval, credit limits, loan terms, and post-loan monitoring, to ensure that risk control responses are highly aligned with user risk dynamics.

[0150] This embodiment also provides a risk control system for bank loan business based on big data analysis, including:

[0151] The module includes a behavior extraction module, a graph modeling module, a scoring correction module, a risk prediction module, and a strategy generation module.

[0152] The behavior extraction module is used to collect real-time behavior data and preprocess it to obtain a set of user behavior features;

[0153] The graph modeling module is used to call the behavior graph modeling engine to perform structured processing on real-time behavior data based on user behavior feature sets, construct multi-dimensional behavior graphs and match them with the risk anchor rule base to obtain potential risk behavior patterns, and label the initial risk anchor labels.

[0154] The scoring correction module is used to correct the deviation between the initial risk anchor point and the actual default record based on the initial risk anchor point label and the value function optimization method in reinforcement learning. It also combines user historical behavior sample data and loan feedback data to predict the risk level score of each user's current behavior.

[0155] The risk prediction module is used to combine historical state evolution data and the risk level score of the user's current behavior to predict the probability distribution of the future migration to the default state;

[0156] The strategy generation module is used to obtain the user's loan business risk at the current point in time based on the probability distribution of the user's migration to the default state and the potential loss under each path, and to generate risk control strategies through risk level mapping rules and strategy decision engine.

[0157] This embodiment also provides a computer device applicable to the risk control method for bank loan business based on big data analysis, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the risk control method for bank loan business based on big data analysis as proposed in the above embodiment.

[0158] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0159] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the risk control method for bank loan business based on big data analysis as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0160] In summary, this invention achieves a structured representation of user behavior data through a graph modeling step, and enhances the targeting and accuracy of risk identification by matching a risk anchor rule base to identify high-risk behavior patterns and labeling them with initial risk anchor tags. Through a scoring correction step, a state-action-feedback triplet is constructed, and a value function is used to optimize the deviation between the initial risk tags and the loan results, achieving dynamic adjustment of the risk scoring logic. Ultimately, this improves the accuracy and adaptability of risk scoring, enhancing the intelligent perception and control capabilities for loan behavior risks.

[0161] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A big data analysis-based bank loan business risk control method, characterized in that: Comprising, Collect real-time behavior data and preprocess to obtain user behavior feature set; Based on the user behavior feature set, call the behavior graph modeling engine to structure the real-time behavior data, build a multi-dimensional behavior graph, and match it with the risk anchor point rule library to obtain potential risk behavior patterns, while labeling the initial risk anchor point label; Based on the initial risk anchor point label, use the value function optimization method in reinforcement learning to correct the deviation between the initial risk anchor point and the actual default record, and combine user historical behavior sample data and loan feedback data to predict the risk level score of each user's current behavior; Combine historical state evolution data and the risk level score of the user's current behavior to predict the probability distribution of migrating to the default state in the future; According to the probability distribution of the user migrating to the default state and the potential loss under each behavior path, obtain the loan business risk of the user at the current time point, and generate a risk control strategy through the risk level mapping rule and the strategy decision engine; The steps of predicting the risk level score of each user's current behavior are as follows, Based on the multi-dimensional behavior graph, statistically analyze the association between the initial risk anchor point label and the user's historical loan performance to predict the historical default probability of each type of initial risk anchor point label in the historical loan sample; Use the value function optimization method in reinforcement learning to form a state-action-feedback triple of the user's credit decision action and subsequent actual credit results under the high-risk structure; By constructing the value function, the user behavior state is taken as the input and the actual credit result is taken as the feedback to evaluate the deviation of the current risk score processing logic in long-term risk prediction; If the deviation is greater than the risk score tolerance threshold, adjust the risk score processing logic to correct the preliminary risk label; Fuse the corrected initial risk anchor point label and the time sequence and high-risk structure in the multi-dimensional behavior graph to generate a comprehensive behavior feature vector that characterizes the user's current behavior risk situation, and predict the risk level score of the user's current behavior risk level; The steps of predicting the probability distribution of migrating to the default state in the future by combining the historical state evolution data and the risk level score of the user's current behavior are as follows, Based on the risk level score, combine the business attributes of the current loan account to obtain the credit state of the current loan account; Use the multi-state Markov modeling method to establish a different loan risk state transition matrix based on the historical account state evolution data and the credit state of the current loan account, and dynamically calculate the path distribution and probability distribution of the user migrating from the current credit state to each type of target risk state in the future time period; The steps of obtaining the loan business risk of the user at the current time point according to the probability distribution of the user migrating to the default state and the potential loss under each behavior path are as follows, Based on the probability distribution, obtain the cumulative probability of migrating to the high-risk structure state; Combine the potential credit loss value corresponding to each migration path to perform risk assessment calculation and obtain the expected credit loss value of the loan account; Fuse the expected credit loss value and the probability distribution to generate a comprehensive loan business risk score of the user at the current time point.

2. The big data analysis based bank loan business risk control method according to claim 1, characterized in that: The real-time behavior data includes online behavior data generated by the user in the bank loan service; The preprocessing includes standardization processing of data cleaning, timestamp correction, redundant field elimination, and behavior normalization. 3.The big data analysis based bank loan business risk control method according to claim 2, characterized in that: The behavior graph modeling engine is called to perform structured processing on the real-time behavior data, construct a multi-dimensional behavior graph, and match with a risk anchor point rule library, and the steps are as follows, Based on the user behavior feature set, the operation behavior data of the user at different time points is marked in combination with the behavior graph modeling engine, and is organized in a graph data structure in the form of event nodes, the logical and time sequence relationship between the event nodes is constructed by edges, and a basic behavior graph based on nodes-edges is obtained; The substructure mode related to the risk behavior in the basic behavior graph is extracted, and pattern matching is performed with the risk anchor point rule library to identify high-risk structures in the basic behavior graph; Through a pattern fitting algorithm, the risk anchor points of each behavior path and event node of the high-risk structure are labeled to obtain a multi-dimensional behavior graph containing initial risk anchor point labels.

4. The big data analysis based bank loan business risk control method according to claim 1, characterized in that: The risk control strategy is generated by the risk level mapping rule and the strategy decision engine, and the steps are as follows, The comprehensive loan business risk score is mapped to a risk level interval by using a risk level mapping rule; Based on the mapping result, the corresponding strategy template is matched according to different risk levels to generate an initial risk control strategy, and the personalized risk control strategy is generated by dynamically correcting the user behavior feature set.

5. A bank loan business risk control system based on big data analysis, based on the bank loan business risk control method based on big data analysis of any one of claims 1-4, characterized in that: It includes a behavior extraction module, a graph modeling module, a score correction module, a risk prediction module, and a strategy generation module; The behavior extraction module is used for collecting real-time behavior data and performing preprocessing to obtain a user behavior feature set; The graph modeling module is used for calling the behavior graph modeling engine to perform structured processing on the real-time behavior data based on the user behavior feature set, constructing a multi-dimensional behavior graph, and matching with a risk anchor point rule library to obtain potential risk behavior patterns, while labeling initial risk anchor point labels; The score correction module is used for correcting the deviation between the initial risk anchor points and the actual default records by using the value function optimization method in reinforcement learning based on the initial risk anchor point labels, and predicting the risk level score of each user's current behavior in combination with historical behavior sample data and loan feedback data of the user; The risk prediction module is used for predicting the probability distribution of migrating to a default state in the future in combination with historical state evolution data and the risk level score of the user's current behavior; The strategy generation module is used for obtaining the loan business risk of the user at the current time point in combination with the potential loss under each behavior path according to the probability distribution of the user migrating to a default state, and generating a risk control strategy by the risk level mapping rule and the strategy decision engine. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the bank loan business risk control method based on big data analysis in any one of claims 1-4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the bank loan business risk control method based on big data analysis in any one of claims 1-4.

Images