Credit overdue prediction analysis method and system based on knowledge graph

Abstract

The invention discloses a credit overdue prediction analysis method and system based on a knowledge graph, and the method comprises the steps: extracting entity network structure and business association two-dimensional features based on the credit knowledge graph, forming a numerical feature set, matching overdue labels, and removing invalid samples; through double-index two-stage screening of core features, feature discretization is completed in combination with quartile and a credit business threshold value; effective association rules are mined by focusing on overdue labels to form a rule base, and a linkage judgment mechanism of'rule matching priority + GNN model understanding 'is adopted. According to the method, the problems of incomplete feature coverage, single screening, poor adaptability and the like in traditional prediction are solved, the prediction accuracy and the coverage are considered, the rule mining efficiency and the feature quality are improved, and reliable technical support is provided for credit decision-making of financial institutions.

Description

Technical Field

[0001] This invention relates to the field of predictive analytics, specifically to a method and system for predicting and analyzing credit delinquency based on knowledge graphs. Background Technology

[0002] As a core component of the financial system, credit business's risk management capabilities directly impact the asset security and sustainable operation of financial institutions. Accurate identification and prediction of overdue risk are crucial aspects of credit risk management. With the expansion of credit business and the increasing complexity of transaction scenarios, the relationships between entities are becoming increasingly diverse, revealing significant limitations in traditional credit risk assessment methods. Firstly, existing solutions often rely on single-dimensional financial indicators or isolated customer information, failing to fully explore the network structure relationships (such as the density of relationships) and business relationship characteristics (such as credit business interactions) of entities within the credit knowledge graph, resulting in insufficient comprehensiveness in feature extraction. Secondly, the feature selection process often uses single-indicator judgments, lacking dual-dimensional collaborative verification of relationship strength and statistical significance, easily introducing invalid features with weak discriminative power, affecting the accuracy of risk prediction. Furthermore, feature discretization often relies solely on statistical distributions or single business thresholds, failing to effectively integrate statistical regularities with credit risk awareness, resulting in poor adaptability. Additionally, the feature combination mining process does not selectively retain high-risk single feature combinations, leading to a large amount of invalid computation, and the risk assessment system lacks effective linkage between the rule base and the model, making it difficult to balance assessment efficiency and coverage.

[0003] Patent document CN119886312A discloses a method for generating strategy rules based on decision trees, including: acquiring historical credit performance data from credit institutions; processing and sampling the historical credit performance data according to business scenarios to obtain a raw dataset; establishing a decision tree model; mining rules from all nodes of the decision tree model; selecting calculation indicators based on credit business scenarios to obtain rules and their corresponding indicator information; and filtering and sorting the rules output by the decision tree model in conjunction with credit business indicators to obtain a combination of strategy rules. The process of acquiring the raw dataset includes the following steps: acquiring historical customer credit performance data from credit institutions; adjusting the variable data types according to actual business scenarios; performing stratified sampling of data samples according to the proportion of each type of sample in the credit performance data; calculating the weight variables of the sampled data; and dividing the sampled data into training and test sets to obtain a usable raw dataset.

[0004] The aforementioned problems in existing technologies lead financial institutions to face challenges in credit risk prediction, such as low feature quality, insufficient rule effectiveness, and low accuracy in risk identification, failing to meet the needs of refined and intelligent risk management. To address these pain points, there is an urgent need to construct a comprehensive feature processing and risk assessment technology system based on credit knowledge graphs. This system should comprehensively extract multi-dimensional features of entities, scientifically quantify feature differentiation capabilities, optimize feature discretization and combination mining logic, and establish a judgment mechanism that links rules and models. This will improve the scientific rigor, accuracy, and reliability of credit delinquency risk prediction, providing strong support for financial institutions' credit decisions.

[0005] Therefore, optimizing the existing credit delinquency prediction and analysis system is a problem worth studying. Summary of the Invention

[0006] To address the shortcomings of the existing technology, the present invention aims to provide a knowledge graph-based method for predicting and analyzing credit delinquency, and a knowledge graph-based system for predicting and analyzing credit delinquency, in order to solve the problems mentioned in the background.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] The first aspect is a knowledge graph-based method for predicting and analyzing credit delinquency, which includes the following steps:

[0009] Based on the credit knowledge graph, the network structure features and business association features of entities are extracted and transformed into a numerical raw feature set;

[0010] Match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples;

[0011] The ability to distinguish features is quantified from the dimensions of correlation strength and statistical significance. The features are selected by priority. The first priority retains features with high correlation strength or strong statistical significance, the second priority retains features with medium correlation strength and significant statistical correlation, and features that meet only one indicator but not both indicators are eliminated.

[0012] The core features are discretized based on the quartiles and credit business thresholds. Then, overdue labels are used as constraints to extract feature combinations that meet business requirements as candidate sets. Combinations with high-risk single features are retained to prune invalid calculations. Finally, effective association rules are selected based on confidence, lift, and coverage that meet risk prediction requirements.

[0013] The rules are validated using a test set, and the compliant rules are retained to form a rule base. For the entity to be predicted, the risk is determined by matching it against the rule base. If there is no match, the GNN model is called to output the result.

[0014] Furthermore, the step of extracting network structure features and business association features of entities based on the credit knowledge graph and transforming them into a numerical raw feature set involves, specifically, querying statements based on the credit knowledge graph and extracting core features of entities, including network structure features reflecting the density of entity graph associations and business association features reflecting the credit business associations of entities. All extracted features are then transformed into identifiable numerical data and associated with the corresponding entity IDs to form a raw feature set containing unique entity identifiers and various numerical features.

[0015] Furthermore, the process of matching the original feature set with overdue records from the past three years, adding overdue binary classification labels, and removing invalid samples involves using the entity ID in the original feature set as the association key to match customer overdue records from the credit business system from the past three years, adding overdue binary classification labels to the entities, assigning a label of 1 if an overdue record exists, and 0 if no record exists, cleaning the data after adding labels, and removing invalid samples that do not have corresponding overdue labels or are missing key features, thus forming a feature label dataset.

[0016] Furthermore, the method of quantifying the feature differentiation ability from the dimensions of association strength and statistical significance specifically employs a dual indicator system of mutual information value and chi-square test. The mutual information value is used to quantify the dependency relationship between the feature and the overdue label, and a mutual information value of not less than 0.3 indicates that the association degree meets the standard. The chi-square test is used to construct a contingency table of feature values ​​and overdue labels. By calculating the chi-square statistic and querying the corresponding p-value, the association significance is judged. A p-value less than 0.05 indicates that there is a significant association between the feature and the label, and a p-value less than 0.01 indicates a strong significant association.

[0017] Furthermore, the first priority retains features with high correlation strength or strong statistical significance; specifically, the first priority retains features with both high correlation strength and strong statistical significance. During screening, feature indicators are verified, and features that meet the criteria of high correlation strength or strong statistical significance are included in the core feature set. Features that meet both criteria are marked with priority levels.

[0018] Furthermore, the second priority retains features with moderate correlation strength and significant statistical correlation. Specifically, for the features remaining after the first priority screening, a second screening is conducted based on the dual-indicator compliance as the core condition. During the screening, the original data and verification records of the dual indicators of the feature are retrieved first, and the indicator range is checked to avoid overlap with the first priority standard. For compliant features, the integrity of the calculation process needs to be verified. If the indicator is critical or questionable, it is recalculated. The compliant features are marked with the corresponding level, and the indicator values ​​and ranges are recorded simultaneously.

[0019] Furthermore, the removal of features that meet only one indicator but not both indicators simultaneously involves using simultaneous compliance of both indicators as the core criterion. After priority screening, features that meet only one indicator are investigated, including those that meet the correlation strength requirement but have no significant statistical correlation, and those that have a significant statistical correlation but have low correlation strength. During screening, an indicator ledger is created for each feature, marking the compliance type and value. Indicators that are at the critical point need to be recalculated and verified. Once it is confirmed that there are no errors, the feature is removed. At the same time, the information of the removed features is recorded to form a traceability file, ensuring that all features in the core feature set meet the dual requirements of correlation strength and statistical significance.

[0020] Furthermore, the discretization of core features using quartiles and credit business thresholds involves the following steps: First, the core feature set and corresponding feature label dataset, filtered by dual indicators, are retrieved. The distribution of feature values ​​is analyzed, and quartiles are calculated to clarify the natural distribution range. Business risk thresholds are determined for different types of core features. The boundaries are adjusted by integrating the quartile ranges and business thresholds, taking into account both statistical regularity and credit risk perception. After discretization, the numerical features are converted into a "feature name-range level" classification format. The range level is defined as low, medium, and high according to business risk. The relevant calculations and basis are recorded to form a traceability document.

[0021] Furthermore, the step of extracting feature combinations that meet business requirements as candidate item sets, constrained by overdue labels, involves retrieving discretized core features and corresponding feature label data, focusing on samples with an overdue label of 1, generating a 1-itemset for a single discretized feature, calculating label support and removing substandard item sets as the base item set, generating k-itemsets based on this, and retaining only combinations containing high-risk 1-itemsets. High-risk 1-itemsets must meet the dual-indicator strong correlation standard and a high business risk level. Business rules are verified to exclude logically contradictory combinations, ultimately forming a candidate item set that meets the requirements and recording relevant information.

[0022] Secondly, a knowledge graph-based credit delinquency prediction and analysis system includes:

[0023] The feature extraction module is used to extract the network structure features and business association features of entities based on the credit knowledge graph and transform them into a numerical raw feature set.

[0024] The data preprocessing module is used to match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples;

[0025] The feature filtering module is used to quantify the distinguishing ability of features from the dimensions of correlation strength and statistical significance, filter features according to a specified priority, and remove features that meet only one indicator but do not meet both indicators at the same time.

[0026] The association rule mining module is used to discretize core features based on credit business thresholds using quartiles, extract a candidate set of feature combinations with overdue labels as constraints, and filter effective association rules according to confidence, lift, and coverage.

[0027] The rule base building module is used to validate rules with a test set and retain the qualifying rules to form a rule base.

[0028] The risk assessment module is used to determine the risk of the entity to be predicted by matching it with the rule base. If there is no match, the GNN model is called to output the result.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This application effectively addresses the pain points of traditional credit delinquency prediction, such as incomplete feature extraction, simplistic feature selection, poor adaptability of discrete features, redundant and ineffective computations, and insufficient linkage in the judgment system. It possesses significant technical advantages: First, it extracts dual-dimensional features—entity network structure and business association—based on a credit knowledge graph, overcoming the limitations of traditional solutions that rely on single indicators. Combined with entity ID binding, it forms a complete numerical feature set, improving the comprehensiveness and relevance of feature coverage. Second, it employs a dual-indicator system to quantify feature discrimination capabilities. Through a two-level priority selection and single-indicator criterion feature elimination mechanism, it ensures that the core feature set possesses both high correlation strength and statistical significance. The system significantly improves feature quality; thirdly, it integrates quartile statistical intervals and credit business thresholds for feature discretization, taking into account both data patterns and business risk awareness, thus enhancing feature adaptability; fourthly, it focuses on generating feature combinations from overdue label samples, retains high-risk single feature combinations to prune invalid calculations, and combines confidence, lift, and coverage to screen effective rules, thereby improving the efficiency and effectiveness of rule mining; fifthly, it constructs a risk judgment mechanism that links the rule base and the GNN model, prioritizing rule matching to ensure judgment efficiency, while the model provides fallback coverage for special scenarios, balancing the accuracy and coverage of risk prediction, and providing more reliable technical support for financial institutions' credit decisions. Attached Figure Description

[0031] Figure 1 This is a flowchart of the knowledge graph-based credit delinquency prediction and analysis method of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] This application discloses a knowledge graph-based method for predicting and analyzing credit delinquency, such as... Figure 1 The steps include:

[0034] S1. Based on the credit knowledge graph, extract the network structure features and business association features of entities and transform them into a numerical raw feature set;

[0035] S2. Match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples;

[0036] S3. Quantify the ability to distinguish features from the dimensions of association strength and statistical significance, and filter them according to priority. The first priority retains features with high association strength or strong statistical significance, the second priority retains features with medium association strength and significant statistical association, and eliminates features that meet the standard only for one indicator but not for both indicators.

[0037] S4. Discretize the core features based on the quartiles and credit business thresholds, then extract feature combinations that meet the business requirements as candidate sets using overdue labels as constraints. Retain combinations containing high-risk single features to prune invalid calculations. Finally, select effective association rules based on confidence, lift, and coverage that meet the risk prediction requirements.

[0038] S5. Validate the rules using the test set, retain the compliant rules to form a rule base; for the entity to be predicted, match the rule base to determine the risk, and if there is no match, call the GNN model to output the result.

[0039] In step S1, based on the credit knowledge graph, network structure features and business association features of entities are extracted and transformed into a numerical raw feature set. Specifically, based on the constructed credit knowledge graph, Neo4j graph query statements are used to accurately locate and extract two types of core features of entities. One type is network structure features reflecting the density of associations within the graph, such as degree centrality and betweenness centrality; these features are quantified using the NetworkX toolkit. The other type is business association features reflecting the credit business relationships of entities, such as the number of guarantee relationships, shareholding relationships, and loan relationships; these features are obtained by statistically analyzing the associated edges corresponding to the entity. All extracted features are transformed into numerical data that can be recognized by subsequent processing stages and associated with the corresponding entity ID, ultimately forming a raw feature set containing unique entity identifiers and various numerical features.

[0040] In step S2, the original feature set is matched with overdue records from the past three years, and overdue binary classification labels are added to remove invalid samples. Specifically, the entity ID in the original feature set is used as the unique association key to establish an association mapping with the customer overdue records from the past three years stored in the credit business system. The credit performance of each entity is matched one by one, and an overdue binary classification label is added to the entity. That is, if the entity has credit overdue records in the past three years, the label is assigned a value of 1, and if there are no credit overdue records in the past three years, the label is assigned a value of 0. After the label is added, the "feature-label" data is cleaned to remove two types of invalid samples: one is entity data without corresponding overdue labels, and the other is entity data with missing values ​​in key features. Finally, a standardized "feature-label" dataset with a sample size of no less than 5,000 and an overdue sample ratio controlled between 10% and 30% is formed, providing a high-quality data foundation for the subsequent feature selection process.

[0041] The S3 section quantifies the distinguishing ability of features from the dimensions of association strength and statistical significance. In practice, it employs a dual-indicator system of mutual information value and chi-square test to quantify the distinguishing ability of features against overdue labels from the two dimensions of association strength and statistical significance.

[0042] The mutual information value is calculated based on the information entropy theory, which quantifies the dependency relationship between features and labels. The formula is: MI(X,Y)=H(Y)-H(Y|X); where X is the feature and Y is the overdue label.

[0043] First, calculate the marginal entropy of Y, with the formula: H(Y)=-ΣP(y)log2P(y); then calculate the conditional entropy, with the formula: H(Y|X)=-ΣP(x)ΣP(y|x)log2P(y|x); where P(y) is the label probability, P(x) is the feature value probability, and P(y|x) is the conditional probability.

[0044] The calculation results are mapped to the [0,1] interval using Min-Max normalization. The normalization formula is: MI norm =(MI-MI) min ) / (MI max -MI min ); MI norm A value of ≥0.3 indicates that the correlation meets the standard.

[0045] The chi-square test requires constructing a contingency table of feature values ​​and overdue labels, with the number of rows equal to the number of feature bins and the number of columns equal to 2. The chi-square statistic is calculated using the formula: χ² 2 =Σ[(A ij -E ij ) 2 / E ij ]; where A ij E represents the actual number of observations. ij= (Total rows × Total columns) / Total number of samples is the theoretical expected value. Degrees of freedom df = (Number of rows - 1) × (Number of columns - 1), calculated using χ². 2 The distribution table shows the corresponding p-value. A p-value < 0.05 indicates a significant association between the feature and the label, while a p-value < 0.01 indicates a strong significant association.

[0046] This application further proposes to verify statistical significance through "kernel density estimation and permutation test," combining kernel density distribution fitting with permutation test to overcome the dependence of the chi-square test on contingency tables, making it applicable to both continuous and discrete features. Specifically:

[0047] Input data, including feature X i numerical sequences ( m is the number of samples), and the corresponding overdue label sequence (y1, y2, ..., y). m , y∈{0,1}), permutation number K (default K=1000, used to construct the null hypothesis distribution).

[0048] Divide the samples into subsets, and classify the samples into two classes based on the label Y: S0 = {x i |y i =0} (no overdue sample set), S1={x i |y i =1} (overdue sample set), let the sample sizes be m0 and m1 (m0+m1=m). Use a Gaussian kernel function to fit the probability density distributions of S0 and S1 respectively, the kernel density function is:

[0049]

[0050] The standard Gaussian kernel function φ(t) is:

[0051]

[0052] The bandwidth h is selected using the Silverman rule:

[0053] h = 1.06·σ·m -1 / 5 ;

[0054] In the formula, σ is the standard deviation of the corresponding sample set, m is the size of the sample set; f0(x) and f1(x) are the probability density functions of feature X in the samples without overdue payments and overdue payments, respectively.

[0055] Calculate the original distribution difference statistic T0, and use the Kullback-Leibler (KL) divergence to measure the difference between f0 and f1. The larger the KL divergence, the more significant the difference in the characteristic distributions of the two classes of samples. The formula is:

[0056]

[0057] The integral range is the range of values ​​for feature X. T0 ≥ 0 and T0 = 0 indicate that the two distributions are completely identical. The larger T0 is, the more significant the difference.

[0058] A null hypothesis distribution is constructed using a permutation test, with the null hypothesis H0 being "characteristic X". i "Irrelevant to label Y", meaning that the feature distributions of the two types of samples are indistinguishable.

[0059] The permutation statistics are generated using the following steps:

[0060] During each permutation, the label sequence (y1,…,y_{m}) is randomly shuffled to obtain a new label sequence y'1,…,y'{m};

[0061] Divide the sample sets S'0 and S'1 according to the new labels, and repeatedly calculate the corresponding distribution difference statistic T. k (k = 1, 2, ..., K);

[0062] Repeat the permutation K times to obtain K permutation statistics {T1,T2,…,T}. k} constitutes the statistical distribution under the null hypothesis.

[0063] Then calculate the P-value, which is the proportion of the original statistic T0 in the null hypothesis distribution. The formula is:

[0064]

[0065] Where I(·) is the indicator function, when T k When T ≥ T0, I(·) = 1; otherwise, I(·) = 0. The smaller the P-value, the lower the probability that the null hypothesis is true, and the more significant the association between the feature and the label. P < 0.05 indicates a significant association, and P < 0.01 indicates a strong significant association.

[0066] The output is the p-value (statistical significance quantification) for each feature.

[0067] In S3, the first priority is to retain features with high correlation strength or strong statistical significance. Specifically, in implementation, a feature-indicator list is first compiled based on the dual-indicator quantification results, and the normalized mutual information value (MI) corresponding to each feature is clarified. norm () and p-value, and calculate traceability information. During screening, each feature index is verified individually; if feature MI... norm A p-value ≥0.5 indicates a high level of association with the overdue label, and it can be directly identified as a core feature; if the feature p-value <0.01, it indicates a strong statistical significance in its association with the overdue label, and it is also included in the core feature set. For features that simultaneously satisfy MI... norm Features with a p-value ≥0.5 and p-value <0.01 are marked as "strong association - high reliability" and given priority for subsequent rule mining.

[0068] During the screening process, the criteria type of each feature is recorded simultaneously, i.e., it belongs to "high correlation strength" or "strong statistical significance". If a feature with abnormal indicator calculation is found, the dual indicator quantification process needs to be traced back to re-verify it to ensure that the features included in the first priority meet the indicator standards and the data is reliable. Finally, a subset of the first priority core features is formed and features that do not meet the criteria are excluded.

[0069] In S3, the second priority retains features with moderate correlation strength and significant statistical correlation; specifically, in implementation, for the features remaining after the first priority screening, they are selected using "MI". norm The core criterion for secondary screening was "≥0.3 and p-value <0.05", where MI... norm A p-value ≥0.3 indicates a moderate association strength, representing a clear dependency between the feature and the overdue label. A p-value <0.05 indicates a significant statistical association, suggesting that the association is not random. During screening, the original calculation data and verification records of the feature's two indicators are retrieved first, and the MI is checked one by one. norm Check whether it is in the range of [0.3, 0.5) and at the same time confirm whether the p value falls within the range of (0.01, 0.05) to avoid overlapping with the first priority feature standard.

[0070] For features that meet the criteria, the integrity of the feature calculation process needs to be verified. For example, whether the entropy data in the mutual information value calculation and the contingency table construction logic for the chi-square test conform to the specifications are required. If there are borderline indicator values ​​or questionable calculation basis, the corresponding sample fragments from the "feature-label" dataset need to be retrieved and recalculated. Features that meet the criteria are uniformly marked as "moderately correlated - significantly reliable," and their MI is recorded simultaneously. norm The specific numerical values ​​and p-value ranges facilitate the differentiation of feature association strength differences during subsequent feature combination mining, ensuring that features included in the second priority not only meet the indicator requirements but also have clear traceability.

[0071] In S3, features that meet only one indicator but not both indicators are eliminated are removed. Specifically, in implementation, for the features remaining after the first two priority screenings, the core criterion for elimination is "both indicators meeting the standard simultaneously." The focus is on investigating situations where only one indicator meets the corresponding standard but the other does not. This includes two categories: one where the mutual information value meets the standard but the chi-square test fails, i.e., feature MI... norm ≥0.3 (association meets the standard) but p-value ≥0.05 (no significant statistical association), for example, an entity with the feature MI of "low number of lending relationships". norm =0.32 but p-value =0.06, although there is some dependence on the overdue label, the association lacks statistical reliability; another type is where the chi-square test meets the standard but the mutual information value fails, that is, p-value <0.05 (significant association exists) but MInorm <0.3 (insufficient association strength), such as the "degree centrality - medium" feature p-value = 0.04 but MI norm =0.28, and its statistical association may originate from random factors rather than a substantial risk association.

[0072] During the screening process, an indicator comparison log should be established for each feature to be judged, clearly marking the type of indicator and its specific value, and specifying the threshold value (such as MI). norm Features with a p-value of 0.29 and a p-value of 0.051 need to be recalculated and verified using the original "feature-label" sample fragments. Only after confirming there are no calculation errors should the removal operation be performed. Simultaneously, the name of the removed feature, indicator details, and the reason for removal should be recorded to create a feature screening traceability file. This ensures that all features within the core feature set meet the requirement of "both association strength and statistical significance are qualified," providing accurate and reliable input data for subsequent feature discretization and rule mining.

[0073] In S4, the core features are discretized using quartiles and credit business thresholds. Specifically, in implementation, the core feature set and the corresponding "feature-label" dataset after dual-index screening are retrieved first. Statistical analysis tools are used to sort out the numerical distribution of each core feature and calculate the quartiles of each feature, namely Q1 (25th quartile), Q2 (50th quartile), and Q3 (75th quartile), to clarify the natural distribution range of the feature values.

[0074] At the same time, the joint credit risk control team, combined with historical risk cases and industry experience, determines the corresponding business risk thresholds for different types of core characteristics. For example, for business association characteristics such as "number of guarantee relationships", the critical value of the number of guarantee relationships in past overdue cases is mainly referenced. For network structure characteristics such as "between centrality", the threshold standard is set according to the criticality of the entity in the flow of funds in the credit network.

[0075] The quartile intervals are integrated and compared with business thresholds, and the interval boundaries are adjusted accordingly to ensure that the adjusted intervals not only conform to the statistical regularity of the feature values, but also accurately match the risk perception of credit business. For example, a core feature is divided into [0,2), [2,5), and [5,+∞) according to the quartiles. Based on the experience that "three or more related transactions are likely to trigger risk transmission" in business, the middle interval is adjusted to [2,3) and [3,+∞) to enhance risk differentiation.

[0076] After discretization, the numerical data of each feature is converted into a classification format of "feature name-interval level". The interval level is defined as "low", "medium" and "high" based on business risk. At the same time, the quartile calculation results, business threshold basis and final interval division standard of each feature are recorded to form a feature discretization traceability document, which provides standardized feature input for subsequent association rule mining.

[0077] In S4, overdue labels are used as constraints to extract feature combinations that meet business requirements as candidate itemsets. Specifically, this involves first retrieving the discretized core feature data and the corresponding "feature-label" dataset, focusing the data filtering on samples with an overdue label of 1, thus achieving targeted constraints on the overdue labels. The joint credit business team clarifies the business requirements for feature combinations, the core of which includes label support ≥ 5%, the inclusion of high-risk feature dimensions as determined by business experience, and avoidance of redundant combinations without practical risk control significance. First, 1-itemsets are generated, i.e., the "feature name-interval" form of a single discretized feature, such as "number of guarantee relationships - high" or "betweenness centrality - high." Label support is calculated for each 1-itemset: label support = number of overdue samples containing that 1-itemset / total number of overdue samples. 1-itemsets with label support < 5% are removed, and those meeting the requirements are retained as the basic itemsets.

[0078] Subsequently, k-itemsets (k≥2) are generated based on the basic itemsets. During the generation process, a high-risk constraint mechanism is introduced, retaining only combinations that contain at least one "high-risk 1-itemset". The high-risk 1-itemsets must simultaneously meet the strong association criterion in the dual-indicator screening, i.e., MI. norm The range is defined as ≥0.5 or p-value <0.01, and the business risk level is "high".

[0079] After generating each k-itemset, it is synchronously verified whether it conforms to business rules such as "no repeated feature dimensions" and "combination logic is consistent with the credit risk transmission law". For example, combinations with logical contradictions such as "high number of guarantee relationships" and "medium number of guarantee relationships" are excluded. Finally, a set of candidate itemsets that meet the business requirements is formed, and the composition characteristics, label support and business compliance verification results of each itemset are recorded.

[0080] In S4, combinations containing high-risk single features are retained to prune invalid calculations. Specifically, in implementation, a "high-risk single feature list" is first established based on the dual-indicator screening results and the business risk level labeling after feature discretization. Features in the list must simultaneously meet two conditions: first, they must reach a strong correlation standard in the dual-indicator screening, i.e., the normalized mutual information value MI. norm ≥0.5 or p-value <0.01, and secondly, the discretized interval level is defined as "high" risk by credit business rules, such as "number of guarantee relationships - high" or "between centrality - high".

[0081] In the k-itemset (k≥2) generation stage, a pre-validation mechanism is introduced. After candidate combinations are generated, they are first matched with the list, retaining only combinations that contain at least one feature from the list, and directly filtering out combinations that do not contain any high-risk single features. At the same time, auxiliary validation rules are set in conjunction with credit business logic. If there is a significant logical conflict between high-risk single features in a combination and other features, such as "low number of lending relationships" and "high number of guarantee relationships" but the entity has no actual fund transaction record, it must also be included in the pruning scope. Through this dual constraint of "feature list matching + business logic validation", the amount of invalid calculation is reduced while ensuring that all retained candidate combinations have potential overdue risk association attributes.

[0082] In S4, effective association rules are selected based on confidence level, lift, and coverage to meet risk prediction requirements. Specifically, this involves first retrieving the candidate feature combination itemset formed after overdue-oriented pruning, simultaneously associating it with the "feature-label" dataset and overdue sample details, constructing a rule selection index calculation matrix, and clarifying the sample statistical dimensions corresponding to each candidate combination. First, three core indicators are calculated: confidence level is obtained by "number of overdue samples containing this rule combination ÷ total number of samples containing this rule combination," with a core standard of ≥60% to ensure the accuracy of the rule's risk assessment; lift is calculated by first obtaining the prior probability of the overdue label by "total number of overdue samples ÷ total number of samples," then dividing the confidence level by this prior probability, with a result ≥1.5 to prove that the rule has a stronger risk differentiation ability than random judgment, avoiding the selection of redundant rules with no actual risk control value; coverage is calculated by "number of overdue samples containing this rule combination ÷ total number of overdue samples," and must meet the requirement of ≥3% to ensure that the rule can cover a certain proportion of risk scenarios.

[0083] During the calculation process, the accuracy of the indicator values ​​needs to be verified. For example, the confidence level should be retained to two decimal places. For samples with critical values ​​(such as combinations of indicator values ​​that are close to but do not meet the standard), the statistical caliber of the samples needs to be re-verified to eliminate sample counting errors or data matching biases. After the indicator calculation is completed, a three-dimensional joint screening is performed, retaining only combinations that meet all three indicators as valid association rules. At the same time, a secondary verification is carried out in conjunction with the credit business logic. For example, if a rule is "low number of lending relationships, high number of guarantee relationships", it needs to be confirmed whether the combination conforms to the business understanding of "no actual capital inflow but heavy guarantee liability that easily leads to overdue payments", and rules that are logically contradictory or contrary to historical risk patterns are eliminated. Finally, the compliant rules are sorted from high to low confidence level and marked with levels such as "high confidence - high coverage" and "high confidence - medium coverage". The specific indicator values, sample statistical basis and business verification results of each rule are recorded simultaneously to form a structured list of valid association rules, providing accurate input for the subsequent construction of the rule base.

[0084] In S5, the rules are validated using a test set, and the qualifying rules are retained to form a rule base. For the entity to be predicted, the risk is determined by matching the rule base; if no match is found, the GNN model is called to output the result. Specifically, the standardized "feature-label" dataset is randomly divided into a training set and a test set in a 7:3 ratio. The training set is used for association rule mining, and the test set is specifically used for rule validity validation. During the validation phase, the candidate association rule list and test set data are retrieved, and a rule matching and indicator calculation matrix is ​​constructed. Each "feature-discrete value" combination in the test set is compared with the rule, and three core validation indicators are calculated for each rule. The matching accuracy is calculated as "number of correctly matched test samples ÷ total number of matched test samples," and must be ≥75%. The coverage is "number of matched test samples ÷ total number of test samples," and must be ≥80%. The F1 score is calculated by combining precision and recall using the formula 2 × (precision × recall) ÷ (precision + recall), and must be ≥0.6.

[0085] Rules that fail to meet single-indicator standards are directly removed. After all rules are verified, the compliant rules are integrated to form the final rule base. The rule base must meet the comprehensive requirements of "single rule accuracy ≥ 70% + overall coverage ≥ 85%" to ensure coverage of most risk scenarios and accurate judgment. For entities to be predicted, their network structure features and business-related features are first extracted using Neo4j graph query and the NetworkX toolkit. After being converted into numerical values, interval division is completed according to the core feature discretization standard, generating a feature set to be matched in the form of "feature name-interval". This feature set is then fully matched with all valid rules in the rule base. If one or more rules are matched, it is directly judged as high risk; if no rules are matched, a pre-trained GNN model is called, using the entity's embedding vector in the knowledge graph as the model input. The model outputs the overdue probability. When the probability is ≥ 0.5, it is judged as high risk; otherwise, it is low risk. The final output should include the risk level, the specific matching rule content or the probability value output by the GNN model, and list the core features and feature values ​​on which the judgment is based, to ensure the interpretability and traceability of credit decisions.

[0086] This application also protects a knowledge graph-based credit delinquency prediction and analysis system, including:

[0087] The feature extraction module is used to extract the network structure features and business association features of entities based on the credit knowledge graph and transform them into a numerical raw feature set.

[0088] The data preprocessing module is used to match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples;

[0089] The feature filtering module is used to quantify the distinguishing ability of features from the dimensions of correlation strength and statistical significance, filter features according to a specified priority, and remove features that meet only one indicator but do not meet both indicators at the same time.

[0090] The association rule mining module is used to discretize core features based on credit business thresholds using quartiles, extract a candidate set of feature combinations with overdue labels as constraints, and filter effective association rules according to confidence, lift, and coverage.

[0091] The rule base building module is used to validate rules with a test set and retain the qualifying rules to form a rule base.

[0092] The risk assessment module is used to determine the risk of the entity to be predicted by matching it with the rule base. If there is no match, the GNN model is called to output the result.

[0093] This application also protects a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the knowledge graph-based credit delinquency prediction and analysis method provided in the embodiments of this invention.

[0094] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0095] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit programs for use by or in conjunction with an instruction execution system, system, or device.

[0096] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0097] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A knowledge graph-based method for predicting and analyzing credit delinquency, characterized in that, Including the following steps: Based on the credit knowledge graph, the network structure features and business association features of entities are extracted and transformed into a numerical raw feature set; Match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples; The ability to distinguish features is quantified from the dimensions of correlation strength and statistical significance. The features are selected by priority. The first priority retains features with high correlation strength or strong statistical significance, the second priority retains features with medium correlation strength and significant statistical correlation, and features that meet only one indicator but not both indicators are eliminated. The core features are discretized based on the quartiles and credit business thresholds. Then, overdue labels are used as constraints to extract feature combinations that meet business requirements as candidate sets. Combinations with high-risk single features are retained to prune invalid calculations. Finally, effective association rules are selected based on confidence, lift, and coverage that meet risk prediction requirements. The rules are validated using a test set, and the compliant rules are retained to form a rule base. For the entity to be predicted, the risk is determined by matching it against the rule base. If there is no match, the GNN model is called to output the result.

2. The knowledge graph-based credit delinquency prediction and analysis method according to claim 1, characterized in that, The process of extracting network structure features and business association features of entities based on the credit knowledge graph and transforming them into a numerical raw feature set involves, specifically, querying statements based on the credit knowledge graph and extracting core features of entities, including network structure features reflecting the density of entity graph associations and business association features reflecting the credit business associations of entities. All extracted features are then transformed into identifiable numerical data and associated with the corresponding entity IDs to form a raw feature set containing unique entity identifiers and various numerical features.

3. The knowledge graph-based credit delinquency prediction and analysis method according to claim 1, characterized in that, The process of matching the original feature set with overdue records from the past three years, adding overdue binary classification labels, and removing invalid samples involves using the entity ID in the original feature set as the association key to match customer overdue records from the credit business system from the past three years, adding overdue binary classification labels to the entities (label 1 if an overdue record exists, 0 if not), cleaning the data after adding labels, and removing invalid samples with no corresponding overdue labels or missing key features to form a feature label dataset.

4. The knowledge graph-based credit delinquency prediction and analysis method according to claim 1, characterized in that, The method of quantifying the feature differentiation ability from the dimensions of association strength and statistical significance is as follows: a dual indicator system of mutual information value and chi-square test is adopted. The mutual information value is used to quantify the dependency relationship between the feature and the overdue label. A mutual information value of not less than 0.3 indicates that the association degree meets the standard. The chi-square test is used to construct a contingency table of feature values ​​and overdue labels. By calculating the chi-square statistic and looking up the corresponding p value, the association significance is judged. A p value less than 0.05 indicates that there is a significant association between the feature and the label, and a p value less than 0.01 indicates a strong significant association.

5. The knowledge graph-based credit delinquency prediction and analysis method according to claim 4, characterized in that, The first priority retains features with high correlation strength or strong statistical significance. Specifically, the first priority retains features with both high correlation strength and strong statistical significance. During the screening process, feature indicators are verified. Features that meet the criteria of high correlation strength or strong statistical significance are included in the core feature set. Features that meet both criteria are marked with priority levels.

6. The knowledge graph-based credit delinquency prediction and analysis method according to claim 5, characterized in that, The second priority retains features with moderate correlation strength and significant statistical correlation. Specifically, for the features remaining after the first priority screening, a second screening is conducted based on the dual-indicator standard as the core condition. During the screening, the original data and verification records of the dual indicators of the feature are retrieved first, and the indicator range is checked to avoid overlap with the first priority standard. For the compliant features, the integrity of the calculation process needs to be verified. If the indicator is critical or questionable, it is recalculated. The compliant features are marked with the corresponding level, and the indicator values ​​and ranges are recorded simultaneously.

7. The credit delinquency prediction and analysis method based on knowledge graphs according to claim 1, characterized in that, The process of removing features that meet only one indicator but not both indicators simultaneously involves using simultaneous compliance of both indicators as the core criterion. After priority screening, features that meet only one indicator are investigated, including those that meet the correlation strength requirement but have no significant statistical correlation, and those that have a significant statistical correlation but have low correlation strength. During screening, an indicator ledger is created for each feature, marking the compliance type and value. Indicators that are at the critical threshold need to be recalculated and verified. Once it is confirmed that there are no errors, the feature is removed. At the same time, the information of the removed features is recorded to form a traceability file, ensuring that all features in the core feature set meet the dual requirements of correlation strength and statistical significance.

8. The credit delinquency prediction and analysis method based on knowledge graphs according to claim 1, characterized in that, The discretization of core features using quartiles and credit business thresholds involves the following steps: First, the core feature set and corresponding feature label dataset, filtered by dual indicators, are retrieved. The distribution of feature values ​​is analyzed, and quartiles are calculated to clarify the natural distribution range. Business risk thresholds are determined for different types of core features. The boundaries are adjusted by integrating the quartile ranges and business thresholds, taking into account both statistical regularity and credit risk perception. After discretization, the numerical features are converted into a "feature name-range level" classification format. The range level is defined as low, medium, and high according to business risk. The relevant calculations and basis are recorded to form a traceability document.

9. The credit delinquency prediction and analysis method based on knowledge graphs according to claim 1, characterized in that, The process involves using overdue labels as constraints to extract feature combinations that meet business requirements as candidate item sets. Specifically, this involves retrieving discretized core features and corresponding feature label data, focusing on samples with an overdue label of 1, generating a 1-itemset for a single discretized feature, calculating label support and removing substandard item sets as the base item set, generating k-itemsets based on this, and retaining only combinations containing high-risk 1-itemsets. High-risk 1-itemsets must meet the dual-indicator strong correlation standard and a high business risk level. Business rules are verified to exclude logically contradictory combinations, ultimately forming a candidate item set that meets the requirements and recording relevant information.

10. A system utilizing the knowledge graph-based credit delinquency prediction and analysis method according to any one of claims 1-9, characterized in that, include: The feature extraction module is used to extract the network structure features and business association features of entities based on the credit knowledge graph and transform them into a numerical raw feature set. The data preprocessing module is used to match the original feature set with overdue records from the past 3 years, add overdue binary classification labels, and remove invalid samples; The feature filtering module is used to quantify the distinguishing ability of features from the dimensions of correlation strength and statistical significance, filter features according to a specified priority, and remove features that meet only one indicator but do not meet both indicators at the same time. The association rule mining module is used to discretize core features based on credit business thresholds using quartiles, extract a candidate set of feature combinations with overdue labels as constraints, and filter effective association rules according to confidence, lift, and coverage. The rule base building module is used to validate rules with a test set and retain the qualifying rules to form a rule base. The risk assessment module is used to determine the risk of the entity to be predicted by matching it with the rule base. If there is no match, the GNN model is called to output the result.

Images