Credit user intelligent identification method and system fusing multi-dimensional dynamic behavior characteristics

Abstract

The invention belongs to the technical field of financial risk control, and particularly relates to a credit user intelligent identification method and system fusing multi-dimensional dynamic behavior characteristics, and the method comprises the steps: firstly integrating multi-source heterogeneous data, and generating a loan application main body characteristic view through preprocessing and correlation analysis; through a dynamic time sequence model, dynamic risk features are extracted based on a time decay function, time window statistics and sequence pattern mining, and the dynamic risk features and the static features are spliced into a high-dimensional vector; outputting a risk prediction value and a quantitative score by using double modules of an improved GBDT model and an expert rule score card; and finally, determining a final risk level through consistency check and a weight voting mechanism. According to the method, the problem that single-dimension or static data analysis has relatively large limitation in the existing credit risk control technology can be solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of financial risk control, and particularly relates to a credit user intelligent identification method and system fusing multi-dimensional dynamic behavior characteristics. BACKGROUND

[0002] In the process of digital transformation of financial credit business, the convenience and accessibility of credit services continue to improve, but at the same time, some bad credit behaviors also derive, for example, first register or acquire a "clean" enterprise, complete the change of legal person in a short period of time, adjust the operating data to create a compliant appearance, and then initiate low-limit and high-frequency credit applications in the name of related individuals and fully perform them, artificially raise the individual credit score, and finally rely on joint application of enterprises and individuals for large credit, transfer after obtaining funds, and cause bad credit situations; such packaging behaviors not only cause financial institutions to suffer asset losses, but also seriously disrupt the normal order of the credit market and damage the healthy development environment of credit business.

[0003] At present, the existing credit risk control technology has significant limitations in dealing with such new risks: 1. Single recognition dimension: traditional models rely on static enterprise data or only focus on individual credit history, lacking the ability to analyze and penetrate the packaging behavior of enterprises and individuals in a coordinated manner; 2. Static recognition lag: the recognition method based on cross-sectional data is difficult to capture the dynamic evolution and clear time sequence characteristics of the packaging link of "shell cultivation-change-credit cultivation-loan application-run away"; 3. Weak interpretability: the risk results output by the "black box" model lack clear business logic support, making it difficult to meet the transparency requirements of risk control compliance and audit supervision.

[0004] Therefore, the above limitations result in insufficient recognition ability of financial institutions for such specialized and gang-like application behaviors, which in turn causes significant asset losses. SUMMARY

[0005] The technical problem solved by the present application is to provide a credit user intelligent identification method and system fusing multi-dimensional dynamic behavior characteristics, to solve the problem of significant limitations of existing credit risk control technology in single dimension or static data analysis.

[0006] The basic scheme provided by the present application is a credit user intelligent identification method fusing multi-dimensional dynamic behavior characteristics, comprising: S1: systematically integrating multi-dimensional data sources, collecting multi-source heterogeneous data, and pre-processing and correlation analyzing the collected multi-source heterogeneous data to generate a feature view representing the subject applying for credit; S2: The dynamic timing model receives the timing features in the feature view, respectively calls a preset time decay function to assign dynamic time weights to the timing features of different time nodes, and statistically calculates trend features through a preset time window, extracts dynamic risk features representing packaging behaviors based on a sequence pattern mining algorithm, splices the high-dimensional feature vector fused with static and dynamic risk features after being converted into a vector format, and generates the high-dimensional feature vector fused with static and dynamic risk features; S3: An intelligent identification and decision model is constructed based on a GBDT machine learning model and an expert rule score card, the high-dimensional feature vector fused is input into the intelligent identification and decision model, and a risk prediction value and a quantitative score result are respectively output; S4: The first risk level is obtained by judging the risk prediction value according to a preset risk threshold, the second risk level is extracted from the quantitative score result, and the first risk level and the second risk level are judged. If the judgment result is consistent, the risk level is directly used; if the judgment result is inconsistent, the final risk level is output based on a preset weight voting mechanism.

[0007] Further, the S1 comprises: S1-1: Based on a safe and controllable access protocol, a multi-channel data access gateway is built to collect multi-source heterogeneous data; S1-2: The collected multi-source heterogeneous data is preprocessed, and the preprocessing comprises format unification processing, abnormal data elimination processing and missing value completion processing; S1-3: Multi-dimensional entity matching rules are established to form subject data with a loan applicant as the core and an associated individual as the extension; S1-4: Based on the preprocessed multi-source heterogeneous data, a feature view comprising a basic atomic feature layer, a combined derived feature layer and an associated feature layer is constructed.

[0008] Further, in the S1-4, the basic atomic feature layer is used to store the original core features of each data source, the combined derived feature layer is used to cross-combine and calculate the original core features based on business risk control logic to generate and store high-differentiation derived features, and the associated feature layer is used to convert the association relationship in the subject data into quantitative features and store them.

[0009] Further, the S2 comprises: S2-1: The dynamic timing model comprises a data input layer, a decay weight calculation layer, a time window configuration layer, a timing feature mining layer and a feature output layer; S2-2: The data input layer receives the timing features with timestamps in the feature view, and performs data regularization operation to eliminate timing noise; S2-3: A preset time decay function is called on the time sequence features processed in S2-2 to give high weight to the features of recent time nodes and low weight to the features of distant time nodes; the expression is:

[0010] wherein, represents the time weight, represents the difference between the current time and the time when the feature occurs, if exceeds a preset maximum time threshold, the weight is set to 0, represents the decay coefficient, . S2-4: Based on the time window configuration layer, time windows are constructed according to the periodic characteristics of different risk behaviors, and the time window statistics are calculated by calling the constructed time window to extract features representing behavior trends; S2-5: Based on the PrefixSpan algorithm, typical sequences strongly related to packaging behaviors are mined, the features in the current time window are generated into behavior sequences of the loan subject according to the time stamp, and the behavior sequences are checked with the typical sequences to extract dynamic risk features representing packaging behaviors in the current behavior sequence; S2-6: The dynamic risk features are converted into vector format, and are spliced with the static features in the preprocessed multi-source heterogeneous data to generate a high-dimensional feature vector that fuses static features and dynamic risk features.

[0011] Further, the time window in S2-4 includes a sliding window, a rolling window, and an event-triggered window; the calculation of time sequence feature statistics by calling the constructed time window to extract features representing behavior trends is specifically: For continuous time sequence features, the mean, growth rate, and volatility are calculated within the time window; For features with time weight, the weighted sum and weighted mean are calculated within the time window.

[0012] Further, in S2-5, based on the PrefixSpan algorithm, typical sequences strongly related to packaging behaviors are mined, the features in the current time window are generated into behavior sequences of the loan subject according to the time stamp, and the behavior sequences are checked with the typical sequences to extract dynamic risk features representing packaging behaviors in the current behavior sequence, which is specifically: The key features of the loan enterprises and individuals are mapped into discrete behavior items; The behavior items are arranged into behavior sequences in chronological order for each loan subject; Based on a preset minimum support, the PrefixSpan algorithm is used to mine frequent behavior sequences, and combined with business experience, typical sequences strongly related to packaging behaviors are screened out; Based on sequence matching verification and weight threshold verification, the behavioral sequences generated within the time window are verified according to typical sequences. If both sequence matching verification and weight threshold verification are satisfied, the behavioral sequence is determined to be a packaging anomaly time series sequence. The features corresponding to the packaging anomaly time series sequence are extracted as dynamic risk features.

[0013] Furthermore, S3 includes: S3-1: Receive the fused high-dimensional feature vector, preprocess the high-dimensional feature vector, and output the optimized feature vector that adapts to the input of the model. S3-2: Input the optimized feature vector into the intelligent recognition and decision-making model constructed based on the improved GBDT machine learning model and expert rule scoring card, wherein the improved GBDT machine learning model outputs a risk probability score; In the expert rule scoring card, based on financial risk control business experience, a rule base is constructed, which includes trigger conditions, risk scores, and rule descriptions. The rule base is used to match the optimized feature vector information of the current loan applicant, output the risk score, and output the second risk level according to the preset scoring threshold. S3-3: Extracting the risk probability score from the improved GBDT model output The system outputs a first risk level based on a preset probability threshold. It then determines the consistency between the first and second risk levels. If they are consistent, the first risk level is directly adopted. If they are inconsistent, the system sets the weights of the improved GBDT model and the expert rule scoring card, calculates a weighted risk level for the first and second risk levels, and takes the risk level corresponding to the weighted score as the final risk level.

[0014] Furthermore, in S3-2, the improved GBDT machine learning model expression is as follows:

[0015] in, Indicates the risk forecast value. To improve the GBDT machine learning model Decision trees for feature vectors The output, The learning rate; Will Substituting into the Sigmoid function, the risk probability score is output. :

[0016] in, ; The GBDT model parameters are optimized by minimizing the loss function, which is the logarithmic loss function, expressed as follows:

[0017] in, This is a true label for the sample, representing a normal loan application. Abnormal loan applications ; In the improved GBDT model, the decision tree selects the optimal splitting features and splitting points based on the Gini coefficient, which is expressed as follows:

[0018] in, This represents the proportion of samples of class i in a tree node. This indicates two categories: normal and abnormal.

[0019] A credit user intelligent identification system integrating multi-dimensional dynamic behavioral features, applied to the aforementioned credit user intelligent identification method integrating multi-dimensional dynamic behavioral features, includes a cross-domain feature fusion module, a dynamic time-series analysis module, an intelligent identification decision module, and a risk scoring module, wherein: The cross-domain feature fusion module is used to systematically integrate multi-dimensional data sources, collect multi-source heterogeneous data, and preprocess and perform correlation analysis on the collected multi-source heterogeneous data to generate a feature view representing the subject applying for credit. The dynamic time series analysis module is used to construct a dynamic time series model. It receives time series features from the feature view, calls a preset time decay function to assign dynamic time weights to time series features at different time nodes, and statistically analyzes trend features through a preset time window. It also extracts dynamic risk features representing packaging behavior based on a sequence pattern mining algorithm, converts them into vector format, and concatenates them with static features to generate a high-dimensional feature vector that integrates static and dynamic risk features. The intelligent recognition and decision-making module is used to build an intelligent recognition and decision-making model based on the GBDT machine learning model and expert rule scoring card. The fused high-dimensional feature vector is input into the intelligent recognition and decision-making model, and the risk prediction value and quantitative scoring result are output respectively. The risk assessment module is used to determine the first risk level by judging the risk prediction value according to the preset risk threshold, extract the second risk level from the quantitative scoring result, and judge the first risk level and the second risk level. If the judgment results are consistent, the risk level is directly adopted; if the judgment results are inconsistent, the final risk level is output based on the preset weight voting mechanism.

[0020] The principles and advantages of this invention are as follows: In this application, a three-layer technical architecture of "multi-source data fusion - dynamic time-series feature mining - dual-model collaborative decision-making" is used to achieve accurate identification of credit risk. First, a multi-channel data access gateway is built to integrate multi-source heterogeneous data, such as corporate entities, business tax, personal credit information and related networks. After preprocessing and correlation analysis, a panoramic feature view of the loan applicant is constructed, breaking the limitations of traditional data silos. Second, relying on the dynamic time-series model, a time decay function is used to give higher weight to recent behaviors, and time windows are used to statistically analyze behavioral trends. Combined with the sequence pattern mining algorithm, dynamic risk features of packaging behaviors such as "shell company maintenance" and "credit maintenance" are extracted and spliced ​​with static features to form a high-dimensional feature vector, solving the problem of static recognition lag. Finally, based on the GBDT machine learning model, data-driven risk probability prediction is achieved. At the same time, expert rule scoring cards are used to complete business logic-oriented quantitative scoring. The results of the two models are integrated through consistency verification and weighted voting mechanism to output the final risk level, ensuring both identification accuracy and the interpretability and compliance of the decision.

[0021] The advantages are: 1. Multi-dimensional data penetration capability: For the first time, it realizes the systematic integration of cross-entity and cross-domain data of enterprises and individuals. By breaking down data silos through correlation analysis, it can capture the collaborative bad credit behavior chain of "enterprise packaging + personal credit beautification". 2. Dynamic and forward-looking identification capability: With the help of time decay function and sequence pattern mining, it can accurately capture the temporal evolution trajectory of bad credit behavior, shifting the risk identification point from after the fact to before and during the event, and effectively identifying hidden preparatory behaviors such as "shell company maintenance" and "credit rating maintenance". 3. Precise and interpretable decision-making capabilities: The GBDT model ensures the accuracy of risk identification, while the expert rule scoring card provides clear business logic support for decision-making. The dual-model fusion mechanism avoids the traceability problem of "black box" models and reduces the limitations of single rules, thus meeting the requirements of financial compliance audit. 4. Flexible decision-making adaptability: The weighted voting mechanism can balance the decision weights of data models and business rules, and can adjust thresholds and weights according to actual risk control needs, adapting to the balance between risk control and user experience in different credit scenarios. Attached Figure Description

[0022] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a functional block diagram of an embodiment of the present invention. Detailed Implementation

[0023] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1As shown: A method for intelligent identification of credit users that integrates multi-dimensional dynamic behavioral features, including: S1: Systematically integrate multi-dimensional data sources, collect heterogeneous data from multiple sources, and preprocess and perform correlation analysis on the collected heterogeneous data to generate a feature view representing the applicant for credit; wherein, S1 includes: S1-1: Based on a secure and controllable access protocol, build a multi-channel data access gateway to collect multi-source heterogeneous data; S1-2: Preprocess the collected multi-source heterogeneous data. The preprocessing includes format unification processing, abnormal data removal processing, and missing value completion processing. S1-3: Establish multi-dimensional entity matching rules to form subject data with loan applicants as the core and related individuals as extensions; S1-4: Based on the preprocessed multi-source heterogeneous data, construct a feature view including a basic atomic feature layer, a combined derived feature layer, and a related feature layer.

[0024] In this embodiment, addressing the pain point of fragmented data dimensions in traditional risk control, this application designs a multi-level, scalable cross-domain feature fusion architecture. Firstly, based on a secure and controllable access protocol, a multi-channel data access gateway is built to access enterprise entities, business tax, personal credit, and related networks, collecting multi-source heterogeneous data. Specifically: Enterprise entity data is obtained by connecting to the official interface of the Administration for Industry and Commerce to obtain basic identity data such as enterprise registration information, legal person and shareholder change records, enterprise existence status, and business scope; at the same time, supplementary information such as organizational structure and actual controller identification information from the enterprise's internal management system is collected to form an enterprise entity identity feature pool.

[0025] By accessing the compliant data interface of the tax administration system, the business tax data can obtain detailed information on the company's tax invoice issuance (including invoice amount, invoice recipient, and invoice frequency), tax declaration records, changes in tax burden rate, tax credit rating, and other business performance data. At the same time, it integrates supporting business data such as the company's bank statements and transaction vouchers from upstream and downstream of the supply chain to build a pool of characteristics of the company's business capabilities.

[0026] Personal credit data is collected by relying on the authorized interface of credit management agencies to collect credit history, repayment performance records, loan application frequency, debt level, credit inquiry records and other credit data of individuals (including corporate legal persons, actual controllers, core shareholders, etc.) associated with loan applications; supplemented by auxiliary data such as personal social security and housing provident fund contributions, occupational stability, etc., to form a personal credit qualification feature pool.

[0027] The associated network data uses graph data acquisition technology to uncover implicit connections between enterprises, between enterprises and individuals, and between individuals, including equity holding relationships, cross-appointment relationships of legal persons, fund transfer relationships, and address overlap relationships, to construct a relationship graph of loan applicants and form an associated network feature pool.

[0028] To address the issues of format heterogeneity, caliber differences, and inconsistent quality of the aforementioned multi-source heterogeneous data, a standardized end-to-end cleaning process was implemented. This process included format unification, anomaly removal, and missing value completion. For format unification, data format conversion rules were established to convert unstructured text data and semi-structured form data into structured key-value pair formats, and to unify basic formats such as data timestamps, monetary units, and encoding rules, eliminating data format barriers. For anomaly removal, multi-dimensional data validation rules were set up to identify and remove invalid data (null values, duplicate data), abnormal data (such as abnormal invoice records where the tax amount far exceeds the company's operating scale), and illegal data (such as unauthorized privacy data and data with untraceable sources), ensuring basic data quality. For missing value completion, a business rule-driven completion strategy was adopted for missing data in key fields. For fields that can be derived, such as the company's years of operation, the completion was automatically calculated using the registration date and the current date. For missing fields that cannot be directly derived, such as tax credit rating, the completion was reasonably filled based on the average or median of companies in the same industry and of similar size, ensuring feature completeness.

[0029] After the aforementioned multi-source heterogeneous data preprocessing, multi-dimensional entity matching rules are established. First, the main entity is anchored, using the enterprise's unified social credit code and the individual's ID number as the core unique identifiers to establish a main identity index for the loan applicant. This binds the same entity data scattered across different data sources, forming a data cluster of entities with the loan applicant enterprise as the core and related individuals as extensions. Next, the related network data is deeply analyzed to identify implicit connections between loan applicant entities. For example, hidden connections such as "multiple corporate legal representatives sharing the same communication address," "the same actual controller controlling multiple shell companies," and "frequent fund transfers between individuals and multiple companies" are discovered. The characteristics such as connection strength and connection type are quantified and stored to complete the dimensions of entity connection features.

[0030] Finally, based on the preprocessed multi-source heterogeneous data, a three-layer cross-domain fusion feature set is constructed to form a unified feature view of loan applicants, including a basic atomic feature layer, a combined derived feature layer, and a relational network feature layer. The basic atomic feature layer retains the original core features of each data source, such as the company's establishment year, the legal representative's tenure, the number of personal loan delinquencies, and the highest monthly invoice amount, serving as the basic data units for the feature view. The combined derived feature layer, based on business risk control logic, performs cross-combination calculations on the atomic features to generate highly discriminative derived features, such as "the legal representative's tenure..." The system incorporates features such as the ratio of "length of business operations / number of years of business operation", the month-on-month growth rate of "business invoice amount in the past 3 months", and the "frequency of personal loan applications in the past 6 months / total number of historical loans". The association network feature layer transforms implicit associations in the association graph into quantitative features, such as "number of shell companies associated with a business", "number of companies with cross-appointments by an individual", and "average credit default rate of related entities", thus making the associated risks explicit. Finally, through the feature fusion engine, the three levels of features are integrated into a unified feature vector of loan applicants, providing a complete data foundation for subsequent dynamic risk identification.

[0031] S2: Construct a dynamic time-series model to receive time-series features from the feature view. It then applies a preset time decay function to assign dynamic time weights to the time-series features at different time points and statistically analyzes trend features through a preset time window. Based on a sequence pattern mining algorithm, it extracts dynamic risk features representing packaging behavior, converts them into vector format, and concatenates them with static features to generate a high-dimensional feature vector that integrates static and dynamic risk features. S2 includes: S2-1: Construct a dynamic time series model, including a data input layer, a decay weight calculation layer, a time window configuration layer, a time series feature mining layer, and a feature output layer; S2-2: Based on the data input layer, receive the time-stamped temporal features in the feature view and perform data warping to eliminate temporal noise; S2-3: Apply a preset time decay function to the time-series features processed in S2-2, assigning high weights to features at recent time points and low weights to features at distant time points; the expression is:

[0032] in, Indicates time weight, This represents the difference between the current time and the time when the feature occurred. If the preset maximum time threshold is exceeded, the weight is set to 0. Indicates the attenuation coefficient. ; S2-4: Based on the time window configuration layer, a time window is constructed according to the periodic characteristics of different risk behaviors, and the constructed time window is called to calculate the time series feature statistics and extract the features that represent the trend of behavior. S2-5: Based on the PrefixSpan algorithm, typical sequences strongly correlated with packaging behavior are mined. The features within the current time window are used to generate loan applicant behavior sequences according to timestamps. The behavior sequences are then verified with typical sequences to extract dynamic risk features representing packaging behavior from the current behavior sequences. S2-6: Convert dynamic risk features into vector format and concatenate them with static features in preprocessed multi-source heterogeneous data to generate a high-dimensional feature vector that integrates static and dynamic risk features.

[0033] In this embodiment, to address the shortcomings of traditional static identification in capturing the dynamic link of "shell company building - change of ownership - credit rating building - loan application", a configurable and interpretable dynamic time series model is constructed. By assigning higher risk weights to recent behaviors through a time decay function and combining multi-granularity time window mining behavior sequence patterns, dynamic tracking of the entire life cycle of credit packaging behavior can be achieved. The dynamic time series model architecture includes a data input layer, a decay weight calculation layer, a time window configuration layer, a time series feature mining layer, and a feature output layer. The data input layer receives the feature set in the feature view and extracts time series features with timestamps, such as legal person change records, tax invoice issuance sequences, and equity change records for enterprises, and credit application time series, repayment record sequences, and credit inquiry time series for individuals. The decay weight calculation layer uses an exponential decay function. To calculate the temporal weights of features, first extract the timestamp of each temporal feature, and then calculate the time interval between it and the current time. ,when (Default maximum time threshold) days, then the weight ;when Then, substitute the values ​​into the decay function to calculate the weights. The default value is 0.05. Finally, the feature value is multiplied by the weight to obtain the feature value with time weight. At the same time, the purpose of the above decay weight calculation layer is to achieve risk focus with high weight for recent behavior and low weight for long-term behavior, so as to provide a differentiated weight basis for subsequent time series feature mining.

[0034] The time window configuration layer provides a multi-level time window parameter configuration interface, allowing administrators to customize the window type and length based on the periodic characteristics of different risk behaviors (such as "shell company maintenance" behavior usually lasting 3-6 months, and "credit maintenance" behavior usually lasting 1-3 months). Specifically, it includes sliding windows, scrolling windows, and event-triggered windows. Sliding windows are used to track recent behaviors in real time, while scrolling windows are used to accumulate long-term behavior sequences, such as automatically scrolling by quarter or semi-annually. Event-triggered windows start from key events and extend for a custom duration, focusing on abnormal behaviors after the event, such as starting from a change in the legal representative of a company and extending for 60 days.

[0035] This application uses a sliding window as an example. After receiving time-series feature data, it calculates time-series statistics and extracts features that reflect behavioral trends. For example, for continuous time-series data such as enterprise invoice amount and individual loan application frequency, it calculates the average daily invoice amount of enterprises within a 60-day sliding window, the month-on-month growth rate of invoice amount within a 60-day sliding window, and the standard deviation of invoice amount within a 60-day sliding window. For time-weighted feature values, it calculates the weighted number of individual loan applications within a 60-day sliding window, which is obtained by multiplying the weight of each loan application by the number of loan applications. It also calculates the weighted average daily invoice amount of enterprises within 60 days, which is obtained by dividing the daily weighted invoice value by 60, thereby highlighting the impact of recent behavior on trends.

[0036] Next, typical sequence mining processing is carried out. First, the key features of loan applicants and individuals in the past experience data are mapped into discrete behavioral items, such as "change of legal person of enterprise (A)", "invoice amount of enterprise increases by more than 50% (B)", "individual concentrated loan application (C)", and "the frequency of personal credit inquiry surges (D)". Then, for each loan applicant, the behavioral items are arranged into a behavioral sequence according to the time sequence, for example, the sequence of a certain applicant is ABCD; then, based on the preset minimum support, which is 5% by default, the PrefixSpan algorithm is used to mine frequent behavioral sequences, and combined with business experience, typical sequences that are strongly related to packaging behavior are selected, that is, sequences that appear in at least 5% of loan applicants are considered typical patterns, and the generated typical sequences are used as reference templates for subsequent anomaly identification. Finally, based on sequence matching verification and weighted threshold verification, the behavioral sequence of the current loan applicant within the time window is verified against typical sequences. The behavioral sequence is generated by a series of behavioral features of the current loan applicant according to the time sequence of the time window. Sequence matching verification involves comparing the current behavioral sequence with the mined typical sequences based on their similarity. If the matching degree... If the current behavior sequence is ABC and the typical sequence is ABCD, the matching degree is 75%, then the sequence matching check is not satisfied. If the current behavior sequence is ABCD, the matching degree is 100%, then the sequence matching check is satisfied, and it is initially determined to be an abnormal sequence. The weighted threshold verification calculates the weighted sum of key behaviors in the current behavior sequence, such as the weighted sum of a surge in invoice issuance or a concentrated loan application. If it exceeds a preset threshold, the anomaly is further confirmed.

[0037] If both sequence matching verification and weight threshold verification are satisfied, it is determined to be a packaging anomaly time series sequence. The features corresponding to the packaging anomaly time series sequence are extracted as dynamic risk features.

[0038] Finally, the dynamic risk features are transformed into a vector format and concatenated with the static features in the preprocessed multi-source heterogeneous data to generate a high-dimensional feature vector that integrates static and dynamic risk features. The static features are those of enterprises and individuals that are relatively stable under specific dimensions and do not change significantly over time, such as basic enterprise attribute features, inherent personal qualification features, and fixed relationship features. In the application of this application, the static features only provide a basic reference for credit risk identification and still need to complement the dynamic time-series features.

[0039] S3: Construct an intelligent recognition and decision-making model based on the GBDT machine learning model and expert rule scoring cards. Input the fused high-dimensional feature vector into the intelligent recognition and decision-making model, and output risk prediction values ​​and quantitative scoring results respectively; wherein, S3 includes: S3-1: Receive the fused high-dimensional feature vector, preprocess the high-dimensional feature vector, and output the optimized feature vector that adapts to the input of the model. S3-2: Input the optimized feature vector into the intelligent recognition and decision-making model constructed based on the improved GBDT machine learning model and expert rule scoring card, wherein the improved GBDT machine learning model outputs a risk probability score; In the expert rule scoring card, based on financial risk control business experience, a rule base is constructed, which includes trigger conditions, risk scores, and rule descriptions. The rule base is used to match the optimized feature vector information of the current loan applicant, output the risk score, and output the second risk level according to the preset scoring threshold. S3-3: Extracting the risk probability score from the improved GBDT model output The system outputs a first risk level based on a preset probability threshold. It then determines the consistency between the first and second risk levels. If they are consistent, the first risk level is directly adopted. If they are inconsistent, the system sets the weights of the improved GBDT model and the expert rule scoring card, calculates a weighted risk level for the first and second risk levels, and takes the risk level corresponding to the weighted score as the final risk level.

[0040] In this embodiment, to address the issues of weak interpretability and lack of business logic support for risk decision-making in traditional "black box" models, the intelligent recognition and decision-making model designed in this application integrates an improved Gradient Boosting Decision Tree (GBDT) model with an expert rule scoring card, achieving the dual goals of accurate recognition and transparent decision-making. Specifically, the intelligent recognition and decision-making model includes a feature access layer, a feature preprocessing layer, a dual-model inference layer, a risk decision-making layer, and a result output layer, wherein: The feature access layer accesses the fused high-dimensional feature vector, which includes a static feature set and a dynamic feature set. The static feature set includes, for example, the number of years an enterprise has been established and an individual's credit rating. The static feature set and the dynamic feature set together form an original feature vector containing N features. The feature preprocessing layer is used to perform standardization, normalization, and feature selection operations on the original feature vector. Standardization and normalization operations can significantly improve feature quality. For feature selection, in order to avoid redundant features increasing model complexity, the feature importance score of the GBDT model is used to select core features. In the selection process, the basic GBDT model is first trained using historical labeled data (including normal / abnormal loan application samples) to output the importance score of each feature. Then, a feature importance threshold is set, for example, the top 80%, and features above the threshold are retained. Finally, multicollinearity features are extracted through variance inflation factor, and the optimized feature vector is finally output. The above feature selection uses existing technologies, which will not be elaborated on in this paper.

[0041] The optimized feature vector output after preprocessing is input into the dual-model inference layer. This layer includes an improved GBDT machine learning model and an expert rule scoring card. The improved GBDT machine learning model has the following structure: the default number of decision trees is 100, configurable from 80 to 150; the default learning rate is 0.1, configurable from 0.05 to 0.2; the default tree depth is 6 layers, configurable from 4 to 8 layers; and the minimum number of samples per leaf node is 20 to ensure statistical significance. The core computational logic of the improved GBDT machine learning model outputs the following expression:

[0042] in, Indicates the risk forecast value. To improve the GBDT machine learning model Decision trees for feature vectors The output, The learning rate; Will Substituting into the Sigmoid function, the risk probability score is output. :

[0043] in, ; The loss function used is the logarithmic loss function, expressed as:

[0044] in, This is a true label for the sample, representing a normal loan application. Abnormal loan applications ;Optimize model parameters by minimizing the loss function; The tree splitting criterion uses the Gini coefficient to select the optimal splitting features and splitting points. The expression for the Gini coefficient is:

[0045] in, This represents the proportion of samples of class i in a tree node. , representing two categories: normal and abnormal; ensuring maximum purity of nodes after splitting.

[0046] Finally, based on the improved GBDT machine learning model's structure and core computational logic described above, the following training process is executed: Dataset partitioning: The historical labeled samples were partitioned into the training set in a 7:2:1 ratio. (used for model parameter learning), validation set ( (used for parameter tuning), test set ( (used for model performance evaluation) Initial model training: Train the improved GBDT machine learning model using the training set, and record the validation set AUC and KS values; Parameter tuning: Optimize parameters such as the number of decision trees, learning rate, and tree depth using grid search (e.g., learning rate traversal of 0.05 / 0.1 / 0.2, tree depth traversal of 4 / 5 / 6), and select the parameter combination with the largest AUC on the validation set; Model evaluation: Evaluate the performance of the final model using a test set, requiring... , Otherwise, readjust the settings.

[0047] Finally, the improved GBDT machine learning model, after training, outputs a risk probability score.

[0048] The expert rule scoring card, based on financial risk control business experience, constructs a rule base that includes triggering conditions, risk scores, and rule descriptions. It then matches the optimized feature vector information of the current loan applicant using this rule base to output a risk score. An example of the rule base is shown in Table 1 below. Table 1. Example of a rule base

[0049] When calculating the risk score, the optimized feature vector of the current loan applicant is first matched against expert rules one by one. If a trigger condition is met, the corresponding risk score is accumulated, the total risk score S is calculated, and then compared with a set scoring threshold to output a second risk level, for example, low risk. Medium risk High risk .

[0050] Finally, the risk decision-making layer extracts the risk probability score output by the improved GBDT machine learning model, compares it with a set probability threshold, and outputs the first risk level, such as low risk. Medium risk High risk The system extracts the second risk level from the expert rule scoring card output. It then checks if the first and second risk levels are consistent; if they are, the risk level is output. If they are inconsistent, a weighted voting process is initiated. The weight of the improved GBDT machine learning model is set to 0.6, and the weight of the expert rule scoring card is set to 0.4. A weighted contribution level is calculated, for example, low risk = 1 point, medium risk = 2 points, and high risk = 3 points. The level corresponding to the weighted score is taken as the risk level. The level corresponding to the weighted score includes: score A score between 1.8 and 2.2 indicates a high risk level; a score between 1.8 and 2.2 indicates a medium risk level. Low risk.

[0051] The output layer outputs a structured risk report, including the following core information: Basic information: Name of the company applying for the loan, Unified Social Credit Code, Loan amount applied for, Loan application time, and associated personal information; Risk assessment results: final risk level, GBDT risk probability score, and expert rule total risk score; Key risk characteristics: Top 3 risk characteristics ranked by contribution (e.g., "legal representative's tenure for 2 months" and "6 loan applications in the past 30 days"); Recommended course of action: Automatic approval / manual review / automatic blocking. If review is required, specify the key points of review (e.g., "the authenticity of the surge in corporate tax invoices needs to be verified").

[0052] Therefore, compared with existing technologies, the technical solution of this application firstly has high identification accuracy and strong targeting, and can directly target the core behavioral patterns of data packaging. The technical solution originates from the in-depth deconstruction of the real packaging chain, and the feature design directly targets the key points, realizing the leap from "general screening" to "precision strike". Secondly, by analyzing the dynamic behavioral sequence features, it can accurately identify packaging behaviors such as "shell company maintenance" and "credit maintenance", shifting the risk identification point from ex-post to ex-ante and in-process. Furthermore, it is the first to systematically integrate corporate entity behavior, business tax data, personal credit trajectory and complex correlation network. Through cross-validation of multi-dimensional information, the confidence and robustness of the identification results are greatly improved. Finally, the selected features all have clear business meaning and risk logic, and the model output results can be traced back to specific abnormal behaviors, which highly meets the strict requirements of financial risk control for model interpretability and compliance audit.

[0053] like Figure 2 As shown, in another embodiment of this example, a credit user intelligent identification system integrating multi-dimensional dynamic behavioral features is also included, applied to the above-mentioned credit user intelligent identification method integrating multi-dimensional dynamic behavioral features. This system includes a cross-domain feature fusion module, a dynamic time-series analysis module, an intelligent identification decision module, and a risk scoring module, wherein: The cross-domain feature fusion module is used to systematically integrate multi-dimensional data sources, collect multi-source heterogeneous data, and preprocess and perform correlation analysis on the collected multi-source heterogeneous data to generate a feature view representing the subject applying for credit. The dynamic time series analysis module is used to construct a dynamic time series model. It receives time series features from the feature view, calls a preset time decay function to assign dynamic time weights to time series features at different time nodes, and statistically analyzes trend features through a preset time window. It also extracts dynamic risk features representing packaging behavior based on a sequence pattern mining algorithm, converts them into vector format, and concatenates them with static features to generate a high-dimensional feature vector that integrates static and dynamic risk features. The intelligent recognition and decision-making module is used to build an intelligent recognition and decision-making model based on the GBDT machine learning model and expert rule scoring card. The fused high-dimensional feature vector is input into the intelligent recognition and decision-making model, and the risk prediction value and quantitative scoring result are output respectively. The risk assessment module is used to determine the first risk level by judging the risk prediction value according to the preset risk threshold, extract the second risk level from the quantitative scoring result, and judge the first risk level and the second risk level. If the judgment results are consistent, the risk level is directly adopted; if the judgment results are inconsistent, the final risk level is output based on the preset weight voting mechanism.

[0054] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A credit user intelligent identification method fusing multi-dimensional dynamic behavior characteristics, characterized in that: Comprise: S1: Systematic integration of multi-dimensional data sources, collect multi-source heterogeneous data, and pre-process and correlate the collected multi-source heterogeneous data to generate a feature view representing the subject applying for credit; S2: Construct a dynamic time series model to receive time series features in the feature view, respectively call a preset time decay function to assign dynamic time weights to time series features at different time nodes, and statistically analyze trend features through a preset time window, and extract dynamic risk features representing packaging behavior based on a sequence pattern mining algorithm, convert to vector format and splice with static features to generate a high-dimensional feature vector that integrates static and dynamic risk features; S3: Based on the GBDT machine learning model and the expert rule scoring card, an intelligent identification and decision model is constructed, and the fused high-dimensional feature vector is input into the intelligent identification and decision model to output risk prediction values and quantitative scoring results respectively; S4: The risk prediction value is judged according to the preset risk threshold to obtain a first risk level, and the second risk level in the quantitative scoring result is extracted, and the first risk level and the second risk level are judged, if the judgment result is consistent, the risk level is directly adopted; If the judgment result is inconsistent, the final risk level is output based on a preset weight voting mechanism. 2.The credit user intelligent identification method of fusing multi-dimensional dynamic behavior characteristics according to claim 1, characterized in that: The S1 comprises: S1-1: Based on a safe and controllable access protocol, a multi-channel data access gateway is built to collect multi-source heterogeneous data; S1-2: Pre-process the collected multi-source heterogeneous data, which includes format unification processing, abnormal data elimination processing and missing value completion processing; S1-3: Establish multi-dimensional entity matching rules to form a subject data with the applicant enterprise as the core and the associated individual as the extension; S1-4: Based on the pre-processed multi-source heterogeneous data, a feature view including a basic atomic feature layer, a combined derived feature layer and a correlation feature layer is constructed. 3.The credit user intelligent identification method of fusing multi-dimensional dynamic behavior characteristics according to claim 2, characterized in that: In S1-4, the basic atomic feature layer is used to store the original core features of each data source, the combined derived feature layer is used to cross-combine and calculate the original core features based on business risk control logic to generate high-discrimination derived features and store them, and the correlation feature layer is used to convert the correlation in the subject data into quantitative features and store them. 4.The credit user intelligent identification method of fusing multi-dimensional dynamic behavior characteristics according to claim 1, characterized in that: The S2 comprises: S2-1: Construct a dynamic time series model, including a data input layer, a decay weight calculation layer, a time window configuration layer, a time series feature mining layer and a feature output layer; S2-2: Based on the data input layer, receive the time series features with timestamps in the feature view, and perform data regularization operations to eliminate time series noise; S2-3: Call a preset time decay function on the time series features processed in S2-2 to assign high weights to features at recent time nodes and low weights to features at distant time nodes; the expression is: wherein, denotes a time weight, denotes a difference between the current time and the time of occurrence of the feature, and if the difference exceeds a preset maximum time threshold, the weight is set to 0, denotes a decay coefficient, ; S2-4: Based on the time window configuration layer, construct a time window according to the periodic characteristics of different risk behaviors, and call the constructed time window to calculate time series feature statistics and extract features representing behavior trends; S2-5: Mining typical sequences strongly related to packaging behavior based on the PrefixSpan algorithm, generating the behavior sequence of the loan application subject according to the timestamp in the current time window, and checking the behavior sequence with the typical sequence to extract the dynamic risk features representing the packaging behavior in the current behavior sequence; S2-6: Convert the dynamic risk features into a vector format, and splice them with the static features in the preprocessed multi-source heterogeneous data to generate a high-dimensional feature vector that integrates static features and dynamic risk features.

5. The credit user intelligent recognition method of fusing multi-dimensional dynamic behavior characteristics according to claim 4, characterized in that: The time window in S2-4 includes sliding window, rolling window and event triggered window; the calling of the constructed time window calculates the timing feature statistics, and the extraction of the features representing the behavior trend is specifically: For continuous timing features, calculate the mean, growth rate and volatility within the time window; For features with time weight, calculate the weighted sum and weighted mean within the time window. 6.The credit user intelligent identification method of fusing multi-dimensional dynamic behavior characteristics according to claim 5, characterized in that: In S2-5, based on the PrefixSpan algorithm, typical sequences strongly related to packaging behavior are mined, and the behavior sequence of the loan application subject is generated according to the timestamp in the current time window, and the behavior sequence is checked with the typical sequence to extract the dynamic risk features representing the packaging behavior in the current behavior sequence, which is specifically: Map the key features of the loan application enterprise and individual to discrete behavior items; Arrange the behavior items in time sequence into behavior sequences for each loan application subject; Based on the preset minimum support, the PrefixSpan algorithm is used to mine frequent behavior sequences, and combined with business experience, typical sequences strongly related to packaging behavior are selected; Based on sequence matching verification and weight threshold verification, the behavior sequence generated within the time window is verified according to the typical sequence, and both the sequence matching verification and the weight threshold verification are satisfied, then the behavior sequence is determined as a packaging abnormal timing sequence, and the features corresponding to the packaging abnormal timing sequence are extracted as dynamic risk features.

7. The credit user intelligent recognition method of fusing multi-dimensional dynamic behavior characteristics according to claim 1, characterized in that: S3 includes: S3-1: Receive the fused high-dimensional feature vector, pre-process the high-dimensional feature vector, and output the optimized feature vector suitable for model input; S3-2: Input the optimized feature vector into the intelligent identification and decision-making model based on the improved GBDT machine learning model and the expert rule score card, wherein the improved GBDT machine learning model outputs a risk probability score; In the expert rule score card, based on the financial risk control business experience, a rule library including trigger conditions, risk scores and rule explanations is constructed, the optimized feature vector information of the current loan application subject is matched through the rule library, a risk score is output, and a second risk level is output according to the preset scoring threshold; S3-3: Extract the risk probability score output by the improved GBDT model , output a first risk level based on a preset probability threshold; and perform consistency determination on the first risk level and the second risk level, if consistent, directly adopt the risk level; if inconsistent, set the improved GBDT model weight and the expert rule score card weight, calculate the weighted risk level of the first risk level and the second risk level, and take the risk level corresponding to the weighted score as the final risk level. 8.The credit user intelligent identification method of fusing multi-dimensional dynamic behavior characteristics according to claim 7, characterized in that: In S3-2, the expression of the improved GBDT machine learning model is: wherein, represents a risk prediction value, is a learning rate; and is an output of a decision tree to a feature vector in a GBDT machine learning model, is a learning rate; and Will Substituting into the Sigmoid function, the risk probability score is output. : wherein ; Optimize the parameters of the improved GBDT model by minimizing the loss function, and the loss function uses the logarithmic loss function, the expression is: wherein, is the true label of the sample, normal loan application , abnormal loan application ; The decision tree in the improved GBDT model selects the best split feature and split point based on the Gini coefficient, and the expression of the Gini coefficient is: wherein, represents the proportion of the i-th type of samples in the tree node, represents normal and abnormal classes.

9. A credit user intelligent identification system fusing multi-dimensional dynamic behavior characteristics, applied to the credit user intelligent identification method fusing multi-dimensional dynamic behavior characteristics according to any one of claims 1-8. It includes a cross-domain feature fusion module, a dynamic timing analysis module, an intelligent identification and decision-making module, and a risk scoring module, wherein: The cross-domain feature fusion module is used for systematically integrating multi-dimensional data sources, collecting multi-source heterogeneous data, and pre-processing and correlation analysis of the collected multi-source heterogeneous data, to generate a feature view representing the subject applying for credit; The dynamic time series analysis module is used for constructing a dynamic time series model to receive time series features in the feature view, respectively calling a preset time decay function to give dynamic time weights to time series features of different time nodes, and statistically analyzing trend features through a preset time window, and extracting dynamic risk features representing packaging behaviors based on a sequence pattern mining algorithm, and converting the dynamic risk features into a vector format to splice with static features, to generate a high-dimensional feature vector fusing static and dynamic risk features; The intelligent identification and decision module is used for constructing an intelligent identification and decision model based on a GBDT machine learning model and an expert rule score card, inputting the fused high-dimensional feature vector into the intelligent identification and decision model, and respectively outputting a risk prediction value and a quantitative scoring result; The risk assessment module is used for judging the risk prediction value according to a preset risk threshold to obtain a first risk level, extracting a second risk level in the quantitative scoring result, and determining the first risk level and the second risk level, if the determination results are consistent, directly using the risk level, if the determination results are inconsistent, outputting a final risk level based on a preset weight voting mechanism.

Images