Financial risk reasoning method and system based on positive and negative association, electronic device and computer readable storage medium

Abstract

The application discloses a financial risk reasoning method and system based on positive and negative association, an electronic device and a computer readable storage medium. The method comprises the following steps: acquiring a financial extension graph containing positive association information and negative association information, and a financial reasoning rule set containing a graph pattern, an antecedent predicate set and a consequent predicate; matching the rule graph pattern in the financial extension graph to obtain a matching instance; checking whether the matching instance meets the antecedent predicate, instantiating the matching instance that passes the check according to the consequent predicate to generate a candidate reasoning result including a positive or negative conclusion; performing conflict detection and resolution on the candidate reasoning result to obtain a target risk reasoning result; and outputting the target risk reasoning result and reasoning explanation information. The positive and negative evidence fusion reasoning is realized, the false positive rate is reduced through conflict detection and multi-dimensional priority resolution, and auditable explanation information is outputted, so that the fine risk control and strong supervision scene demand are met.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a financial risk reasoning method, system, electronic device, and computer-readable storage medium based on positive and negative correlations. Background Technology

[0002] As financial institutions continue their digital transformation, online lending, interbank transactions, and multi-channel payments are expanding rapidly, creating a large-scale, multi-type heterogeneous network of connections among a vast number of business entities, including customers, accounts, bank cards, and transaction devices. In core financial risk control scenarios such as anti-fraud and anti-money laundering, the industry commonly relies on financial knowledge graphs, graph association rules, graph neural networks, and machine learning models to identify risks. This is used to uncover hidden risks that are difficult to identify with traditional single-dimensional data, such as implicit connections between entities, fraudulent networks, multi-layered suspicious fund flows, and transmission paths of high-risk entities. It also facilitates business tracing and regulatory review of the risk assessment process.

[0003] Currently, the mainstream technologies for financial risk analysis and risk reasoning fall into two main categories. The first is graph rule and business rule engine systems built upon expert experience, which generate risk alerts based on preset association and threshold conditions. The second is risk prediction schemes based on machine learning and graph neural networks, which output risk scores, link prediction results, and entity association determination results through model training. Both types of technologies have basic application value in daily risk control operations, but they have inherent limitations in complex, multi-evidence, and strongly regulatory financial business scenarios: they cannot be compatible with various verified refutational business facts, and they struggle to suppress redundant alerts using excluded evidence. Summary of the Invention

[0004] The purpose of this application is to propose a financial risk reasoning method based on positive and negative correlations, which aims to solve the problems that existing solutions cannot achieve the fusion reasoning of positive and negative business evidence, accurately control risk false alarms, and are not suitable for refined risk control and strong regulatory scenarios.

[0005] This application provides a financial risk reasoning method based on positive and negative correlations, the method including: Obtain a financial extension graph and a set of financial inference rules. The financial extension graph includes at least positive and negative association information between financial entities. Each financial inference rule in the set of financial inference rules includes a graph pattern, a set of antecedent predicates, and an antecedent predicate. Both the set of antecedent predicates and the antecedent predicates contain predicates that represent positive associations and predicates that represent negative associations. Perform graph pattern matching on the graph patterns of each financial inference rule in the financial extended graph to obtain matching instances corresponding to each financial inference rule; For each matching instance, verify whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiate the matching instances that pass the verification according to the consequent predicate of the financial inference rule to generate candidate inference results, which include positive or negative conclusions. The candidate reasoning results are subjected to conflict detection and resolution to obtain the target risk reasoning result; Output the target risk reasoning result and the corresponding reasoning explanation information.

[0006] Accordingly, this application provides a financial risk reasoning system based on positive and negative associations, including: The data access module is used to access financial business data, business verification data, and model prediction data; The financial extended graph construction module is used to construct a financial extended graph based on financial business data, business verification data, and model prediction data. The financial extended graph includes nodes representing financial entities, first-type edges representing confirmed relationships, second-type edges representing confirmed non-relationships, and attribute information for each node and edge. The rule management module is used to store the set of financial reasoning rules. Each financial reasoning rule in the set of financial reasoning rules includes a graph pattern, a set of antecedent predicates and an consequent predicate. Both the set of antecedent predicates and the consequent predicates contain predicates that represent positive associations and predicates that represent negative associations. The candidate rule discovery module is used to generate candidate financial inference rules according to the candidate rule generation strategy, calculate the quality index of each candidate financial inference rule, and filter the complete set of financial inference rules according to the quality index and store them in the rule management module. The graph pattern matching module is used to perform graph pattern matching on the financial extended graph for each financial inference rule, and obtain the matching instance corresponding to each financial inference rule. The predicate verification and reasoning module is used to verify whether each matching instance satisfies the antecedent predicate set of the corresponding financial reasoning rule, and to instantiate the matching instances that pass the verification according to the consequent predicate of the financial reasoning rule to generate candidate reasoning results, which include positive or negative conclusions. The conflict detection module is used to detect conflicts between candidate inference results and identify conflicts between candidate inference results and positive correlation information, negative correlation information, manually verified facts, model prediction results and other candidate inference results in the financial extension graph. The conflict resolution module is used to calculate the priority score of each conflict candidate reasoning result according to the preset conflict resolution strategy parameters when a conflict is detected, determine the resolution result based on the priority score, and determine the target risk reasoning result based on the resolution result. The explanation generation module is used to output the target risk reasoning result and the reasoning explanation information corresponding to the target risk reasoning result; The feedback update module is used to write back the results of manual review and business feedback data to the financial extension graph, update the negative association information in the financial extension graph, and / or optimize the confidence of financial inference rules, and / or adjust the conflict resolution strategy of the conflict resolution module.

[0007] Accordingly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the above-described financial risk reasoning method based on positive and negative associations.

[0008] Accordingly, this application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the above-described financial risk reasoning method based on positive and negative associations.

[0009] This application obtains a financial extended graph containing positive and negative association information, as well as a set of financial inference rules containing graph patterns, antecedent predicate sets, and consequent predicates. Both antecedent and consequent predicates contain predicates representing positive and negative associations. The rule graph patterns are matched within the financial extended graph to obtain matching instances. The matching instances are verified to satisfy the antecedent predicates. Matching instances that pass the verification are instantiated according to the consequent predicates, generating candidate inference results with either affirmative or negative conclusions. Conflict detection and resolution are performed on the candidate inference results to obtain the target risk inference result. The target risk inference result and inference explanation information are then output. This achieves fusion inference of positive and negative evidence, reduces the false positive rate through conflict detection and multi-dimensional priority resolution, and outputs auditable explanation information, meeting the needs of refined risk control and strong regulatory scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] in: Figure 1 A schematic diagram of a financial risk reasoning system based on positive and negative associations provided in this application embodiment.

[0012] Figure 2 This is a flowchart illustrating a financial risk reasoning method based on positive and negative associations, provided as an embodiment of this application.

[0013] Figure 3 This is a flowchart illustrating another financial risk reasoning method based on positive and negative associations, provided as an embodiment of this application.

[0014] Figure 4 This is a structural block diagram of a financial risk reasoning device based on positive and negative associations, provided for an embodiment of this application.

[0015] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] This application provides a method, apparatus, computer-readable storage medium, and electronic device for financial risk reasoning based on positive and negative associations. Specifically, the financial risk reasoning method based on positive and negative associations in this application can be executed by an electronic device, which can be a terminal or a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, and big data and artificial intelligence platforms.

[0018] To address the aforementioned issues, this application provides a financial risk reasoning method, system, electronic device, and computer-readable storage medium based on positive and negative correlations. This enables the fusion reasoning of positive and negative evidence, reduces the false alarm rate through conflict detection and multi-dimensional priority resolution, and outputs auditable and explanatory information, thus meeting the needs of refined risk control and strong regulatory scenarios.

[0019] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0020] For example, please see Figure 1 , Figure 1 A schematic diagram of a financial risk reasoning system based on positive and negative associations provided in this application embodiment.

[0021] like Figure 1As shown, the financial risk reasoning system based on positive and negative associations provided in this application may include: a data access module 101, a financial extended graph construction module 102, a rule management module 103, a candidate rule discovery module 104, a graph pattern matching module 105, a predicate verification and reasoning module 106, a conflict detection module 107, a conflict resolution module 108, an explanation generation module 109, and a feedback update module 110. Specifically, the functions and connections of each module are as follows: The data access module 101 is used to access financial business data, business verification data and model prediction data, and pass them to the financial extended graph construction module 102.

[0022] The financial extended graph construction module 102 is used to construct a financial extended graph based on the accessed data. The financial extended graph includes nodes representing financial entities, first-type edges representing confirmed relationships, second-type edges representing confirmed non-relationships, and attribute information of each node and edge. The graph data is then provided to the candidate rule discovery module 104, the graph pattern matching module 105, and the conflict detection module 107.

[0023] The rule management module 103 is used to store a set of financial reasoning rules. Each rule includes a graph pattern, a set of antecedent predicates, and an antecedent predicate. Both the antecedent predicate set and the antecedent predicate contain predicates representing positive associations and predicates representing negative associations. The rule management module 103 provides rule templates to the candidate rule discovery module 104 and executable rules to the graph pattern matching module 105.

[0024] The candidate rule discovery module 104 is used to generate candidate financial inference rules according to the candidate rule generation strategy, calculate the quality index of each candidate rule, and filter the complete set of financial inference rules according to the quality index and store them in the rule management module 103.

[0025] The graph pattern matching module 105 is used to obtain executable rules from the rule management module 103, perform graph pattern matching on the graph patterns of each rule in the financial extended graph, obtain matching instances, and pass them to the predicate verification and reasoning module 106.

[0026] The predicate verification and reasoning module 106 is used to verify whether each matching instance satisfies the set of antecedent predicates of the corresponding rule. The matching instances that pass the verification are instantiated according to the consequent predicates to generate candidate reasoning results. The candidate reasoning results include positive or negative conclusions. During the verification process, the attribute service, graph query service and model service are called and verified in order of cost from low to high.

[0027] The conflict detection module 107 is used to receive candidate inference results and compare them with positive association information, negative association information, manually verified facts, model prediction results and other candidate inference results in the financial extension graph to identify at least one of the following: conflict between positive and negative conclusions, conflict between mutually exclusive attribute comparisons, conflict between machine learning predicates, conflict between rules and manual verification, and conflict between multi-rule inference.

[0028] The conflict resolution module 108 is used to calculate the priority score of each conflict candidate inference result according to the conflict resolution strategy parameters when a conflict is detected, determine the resolution result according to the priority score, and determine the target risk inference result based on the resolution result; the resolution result includes at least one of retaining positive inference, retaining negative inference, adjusting risk score, marking as pending manual review or including in the observation list.

[0029] The explanation generation module 109 is used to generate explanation results from rule hit records, graphical evidence, model scores and conflict resolution records, and output the target risk reasoning results and their corresponding reasoning explanation information.

[0030] The feedback update module 110 is used to write back the manual review results and business feedback data to the financial extension graph, update the negative association information, optimize the rule confidence and / or adjust the conflict resolution strategy.

[0031] Furthermore, this application provides a financial risk reasoning method based on positive and negative correlations. For example, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating a financial risk reasoning method based on positive and negative correlations, provided as an embodiment of this application. The specific process of this financial risk reasoning method based on positive and negative correlations can be as follows: 201. Obtain the financial extension graph and the set of financial reasoning rules.

[0032] In this application, the financial extension graph includes at least positive and negative association information between financial entities. Positive association information is represented by positive edges to indicate confirmed financial business relationships; negative association information is represented by negative edges to indicate confirmed non-existent or excluded financial business relationships, which can be used to distinguish between the two states of "unknown relationship" and "excluded relationship".

[0033] In some embodiments, prior to the step "obtaining the financial extension graph and the set of financial inference rules", the method may further include the following steps: Acquire financial business data, business verification data, and model prediction data; generate financial entity nodes based on the entity information contained in the financial business data, and construct positive association edges between corresponding entity nodes based on the confirmed business relationships between entities in the financial business data; construct negative association edges between corresponding entity nodes based on the verified exclusion relationships between entities recorded in the business verification data, and add verification-type attribute information to the negative association edges; add model prediction-type attribute information to entity nodes, positive association edges, and negative association edges respectively according to the model prediction data; configure semantic label information and business attribute information for all entity nodes, positive association edges, and negative association edges to complete the construction of the financial extension graph.

[0034] Financial business data refers to original business records such as customer information, accounts, and transactions, used to generate nodes and positive edges. Business verification data refers to data containing definitive conclusions such as manual verification and exclusion, and KYC (Know Your Customer) verification, used to generate negative edges and verification attributes. Model prediction data refers to the risk scores output by machine learning models, used to generate model prediction attributes.

[0035] In this application, the construction of the financial extension graph is based on financial business data, business verification data, and model prediction data, as follows: Financial entity nodes V are created based on the entity information in the financial business data. Node types can include customers, accounts, transactions, merchants, devices, enterprises, etc. Forward edges E+ between nodes are created based on the confirmed business relationships between entities in the financial business data. Each forward edge can be represented as (u, l, v), indicating that a confirmed financial business relationship with the relationship label l exists between node u and node v. Examples include a customer holding an account, an account logging into a device, an account transferring funds to another account, a customer being a legal entity of a company, a customer matching a list of entities, and an account belonging to a fraud group.

[0036] Based on the verified exclusion relationships between entities recorded in the business verification data, negative edges E- are created between nodes, and verification attribute information (such as verification source, verification time, and evidence credibility) is added to the negative edges. Each negative edge can be represented as (u, ¬l, v), indicating that the relationship label l between nodes u and v has been confirmed to be non-existent, excluded, or the opposite of the target relationship. For example, confirming that two accounts do not belong to the same fraud gang, confirming that a customer is not the beneficial owner of a certain enterprise, confirming that a certain entity is not the same entity as the entities on the list, confirming that a certain transaction is not a fraudulent transaction, etc. Negative edges are different from "no edge" relationships. If (u, l, v) does not belong to E+ and (u, ¬l, v) does not belong to E-, it means that the relationship has not been observed or verified and cannot be directly regarded as confirmed to be non-existent.

[0037] Add model prediction class attribute information (FA) to nodes and edges based on the model prediction data, such as model prediction score, model name, and model version, for use by machine learning predicates in subsequent rules.

[0038] Configure label information L and attribute information FA for all nodes and edges. L represents the label function, used to identify the business semantics corresponding to node type, edge type, and variables in the rules; FA represents the attribute set, including node attributes, edge attributes, and evidence attributes, such as transaction amount, transaction time, account opening duration, customer risk level, device fingerprint, verification source, verification time, evidence credibility, model score, etc. After completing the above configuration, the financial extended graph G=(V, E+, E-, L, FA) is obtained.

[0039] For any pair of nodes (u, v) and relation label l, within the same time window and the same business task, a positive edge (u, l, v) and a negative edge (u, ¬l, v) must not exist simultaneously. If both appear in the business data, the conflict detection and resolution process will begin.

[0040] The financial reasoning rules in the financial reasoning rule set include graph patterns, antecedent predicate sets, and consequent predicates. Both the antecedent predicate set and the consequent predicates contain predicates that represent positive associations and predicates that represent negative associations.

[0041] In this application, the financial reasoning rules can take the form of graph association rules containing negative predicates, denoted as: phi = Q[x](X→p0); where Q[x] represents the financial graph pattern, Q consists of pattern nodes, positive pattern edges, negative pattern edges, and label functions, and x is a list of rule variables, with each variable corresponding one-to-one with a pattern node in the graph pattern. X represents the set of antecedent predicates, consisting of the conjunction of one or more predicates. The rule is triggered only when the matching instance satisfies all predicates in X. p0 represents the consequent predicate, i.e., the target predicate inferred after the rule is triggered. It can be a positive risk relationship, a negative risk relationship, an attribute conclusion, or a model predicate conclusion. X→p0 represents the dependency relationship, i.e., when the set of antecedent predicates is true, the consequent predicate is inferred to be true.

[0042] Both the antecedent predicate set and the consequent predicate set contain predicates representing positive associations and predicates representing negative associations. Among them, the predicates representing positive associations are used to express the relationships that have been confirmed to exist in the financial extension graph, such as hold(customer, account) (customer holds an account), login(account, device) (account login device), transfer(account1, account2) (transfer between accounts), same_entity(customer, blacklist_subject) (customer and blacklist subject are the same entity), fraud_collusion(account1, account2) (accounts are associated with a fraud gang), etc.

[0043] Predicates representing negative associations are used to express relationships that have been confirmed to be non-existent or have been verified and excluded in the financial extension graph, such as NOT same_owner(customer1, customer2) (confirming that the two customers are not the same owner), NOT fraud_collusion(account1, account2) (confirming that the two accounts are not associated with a fraud gang), NOT same_entity(customer, blacklist_subject) (confirming that the customer and the blacklist subject are not the same entity), NOT beneficial_owner(customer, enterprise) (confirming that the customer is not the beneficial owner of a certain enterprise), etc.

[0044] Financial reasoning rules can simultaneously express risk discovery and risk elimination. The following two specific examples illustrate this: Risk Detection: Q[a, b, d](transfer(a, b) AND login(a, d) AND login(b, d) AND NOTfraud_collusion_excluded(a, b) AND M_fraud(a, b)>= eta→fraud_collusion(a, b)); This rule means that if there is a transaction relationship between account a and account b, and both have logged into the same device d, and there are no verified and excluded gang relationship negation edges in the financial extension graph (i.e., NOT fraud_collusion_excluded(a, b) is not true, corresponding to the definition of the negation edge predicate in Section 3.4 of the specification, indicating that the relationship has not been excluded), and the fraud model's prediction scores for account a and account b exceed the preset threshold, then it is inferred that there is a fraud gang relationship between account a and account b.

[0045] Risk exclusion: Q[c, e](name_similar(c, e) AND blacklist_subject(e) AND manual_screening(c, e) = pass→NOT same_entity(c, e)); This rule means: if customer c and blacklist subject e have similar names, and e is a blacklist subject, but the manual screening result is passed, then it is inferred that customer c and blacklist subject e are not the same entity.

[0046] In some embodiments, obtaining a set of financial reasoning rules may include the following operations: Obtain the complete set of financial reasoning rules and the target business task; based on the target business task, select a subset of rules from the complete set of financial reasoning rules that match the target business task, and use it as the financial reasoning rule set; The matching criteria include at least one of the following: node type, relationship label, task label, business line, applicable region, time window, and regulatory strategy.

[0047] The target business task refers to the specific risk control problem that needs to be solved in this reasoning process, which may include different scenarios such as anti-fraud investigation, anti-money laundering reporting, credit approval assistance, and list screening. Different business tasks focus on different types of financial entities, relationship types, and risk characteristics, so the entire set of rules needs to be screened in a targeted manner.

[0048] In this application, the complete set of financial inference rules and the current target business task can be obtained first. Based on the target business task, a subset of rules matching the task is selected from the complete set of financial inference rules, which serves as the set of financial inference rules actually used in the current inference process. The selection criteria include at least one of the following: node type (i.e., the type of entity node to which the rule applies), relationship label (i.e., the type of relationship to which the rule applies), task label (i.e., the type of business task to which the rule applies), business line (i.e., the business line to which the rule applies), applicable region (i.e., the geographical scope to which the rule applies), time window (i.e., the time range to which the rule applies), and regulatory strategy (i.e., the category of regulatory requirements corresponding to the rule).

[0049] The above filtering criteria can be pre-configured as metadata tags for rules. During filtering, the tags of the target business task are quickly matched with the rule metadata, retaining only rules relevant to the current task and filtering out irrelevant rules. Through this filtering, only rules relevant to the current business task are passed to the graph pattern matching module, avoiding invalid matching on the entire rule set, significantly reducing the computational overhead of graph pattern matching, and improving inference efficiency.

[0050] In some embodiments, prior to the step "obtaining the set of financial inference rules", the method may further include the following operations: Obtain candidate rule generation strategies; generate candidate financial reasoning rules based on candidate rule generation strategies; calculate the quality index of each candidate financial reasoning rule; select rules that meet preset conditions from the candidate financial reasoning rules based on the quality index to obtain the complete set of financial reasoning rules.

[0051] In this application, the complete set of financial inference rules can be automatically constructed through a candidate rule discovery and quality screening mechanism, as follows: Candidate rule generation strategies can be developed using at least one of the following methods: Business template-based generation: Preset shared equipment templates, short-term cash flow templates, list similarity templates, beneficiary owner exclusion templates, and pre-loan implicit association templates, etc., and generate candidate rules based on node types, edge types, thresholds, and predicates in the instantiated templates for business tasks; Historical risk case mining: Extract high-frequency graph patterns from confirmed fraud cases, suspicious transaction cases, rejected loan applications, list hit cases, and closed false alarm cases; extract positive evidence combinations from positive cases and negative evidence combinations from false alarm or exclusion cases to generate candidate rules; Supervised learning model path generation: Under fixed graph patterns, convert the attributes, relationships, negative edges, and model scores of matching instances into structured features, train decision trees or gradient boosting trees, convert model path conditions into antecedent predicates, and leaf node labels into consequent predicates to generate candidate rules; Generative or heuristic search-based generation: Generate candidate rules through heuristic search, genetic algorithms, reinforcement learning, or diffusion models, and include the generated results in the candidate rule set after filtering by rule validity, business interpretability, and quality indicators.

[0052] For each generated candidate rule, a quality metric is calculated. The quality metric includes at least one of the following: Support: Used to measure the coverage of a rule in a financial extended graph. For a rule phi = Q[x] (X → p0), support can be defined as the number of matching instances that satisfy the antecedent predicate set X, or the number of matching instances that satisfy both the antecedent and consequent: support_X(phi) = |{h | h is match of Q in G and h |= X}|, which means: count the number of matching instances h in the financial extended graph G that satisfy both the graph pattern Q and the antecedent predicate set X, i.e., the number of matching instances that trigger the antecedent of the rule; support_phi(phi) = |{h | h is match of Q in G and h |= X and h |= p0}|, which means: count the number of matching instances h in the financial extended graph G that simultaneously satisfy the graph pattern Q, the antecedent predicate set X, and the consequent predicate p0, i.e., the number of matching instances where both the antecedent and consequent are true. The smaller the support_phi(phi), the more difficult it is for the consequent to be true when the antecedent of the rule is true, and the stricter the rule.

[0053] Confidence: This measures the probability that the consequent is true if the antecedent is true. It is defined as: confidence(phi) = support_phi(phi) / support_X(phi). This formula means that the confidence level equals the number of matching instances that satisfy both the antecedent and consequent divided by the number of matching instances that only satisfy the antecedent. It reflects the conditional probability that "if the antecedent is true, the consequent is also true." The value ranges from 0 to 1; the closer to 1, the more reliable the rule. When support_X(phi) is 0, the rule has no practical meaning and should not be included in the set of valid rules.

[0054] Coverage: Measures the proportion of target risk objects covered by a rule. For example, in anti-fraud tasks, the proportion of confirmed fraudulent account pairs covered by a rule to the total number of confirmed fraudulent account pairs can be calculated.

[0055] Conflict rate: This measures the proportion of rule inference results that conflict with known facts, human verification results, or other highly reliable rules. It is defined as: conflict_rate(phi) = conflict_count(phi) / inference_count(phi). This formula means that the conflict rate equals the number of times the rule's inference result conflicts with facts or other rules divided by the total number of inferences, reflecting the degree of contradiction between the rule and existing knowledge. A higher conflict rate indicates lower rule reliability, and the rule should be filtered, downgraded, or subject to manual review.

[0056] Consistency: Check whether the rule itself contains mutually exclusive predicates, such as the same antecedent containing both the positive edge predicate l(x, y) and the corresponding negative edge predicate NOT l(x, y), or containing both mutually exclusive attribute comparison conditions (e.g., score>= 80 and score<80).

[0057] Business interpretability: Assess whether the rules can be mapped to risk paths, verification evidence, or model outputs with clear business implications, and ensure that the rules can be understood and audited in actual business.

[0058] Based on the aforementioned quality indicators, rules meeting preset conditions are selected from the candidate rules to obtain the complete set of financial inference rules. These preset conditions include: support greater than a preset minimum support threshold, confidence greater than a preset minimum confidence threshold, conflict rate less than a preset maximum conflict rate threshold, passing consistency checks, and achieving a preset standard for business interpretability score. After selection, rules meeting these conditions are included in the complete set of financial inference rules.

[0059] 202. Perform graph pattern matching on the graph patterns of each financial inference rule in the financial extended graph to obtain the matching instance corresponding to each financial inference rule.

[0060] In this application, graph pattern matching is a prerequisite step for performing rule-based reasoning. For each financial reasoning rule phi=Q[x](X →p0), it is necessary to find a matching graph pattern Q in the financial extended graph G. A matching instance can be represented as a mapping from rule variables to entity nodes in the graph: h: x→V; where x is the list of rule variables and V is the set of entity nodes in the financial extended graph.

[0061] Matching instances must meet the following conditions: Node label compatibility: For each pattern node in the graph pattern, the mapped entity node label should be consistent with the pattern node label, or satisfy a preset label compatibility relationship. For example, the "Account" node in the graph pattern can be mapped to both the "Debit Card Account" node and the "Credit Card Account" node in the financial extension graph; Positive edge matching: For each positive pattern edge (xi, l, xj) in the graph pattern, there should be a corresponding positive edge (h(xi), l, h(xj)) in the financial extension graph. For example, if the graph pattern contains the positive edge login(account, device), then the financial extension graph must have a corresponding positive edge indicating login; Negative edge matching: For each negative pattern edge (xi, ¬l, xj) in the graph pattern, there should be a corresponding negative edge (h(xi), ¬l, h(xj)) in the financial extension graph. For example, if the graph pattern contains the negative edge NOTfraud_collusion(account1, account2), then the financial extension graph must have a corresponding negative edge indicating that the association with the fraud gang has been excluded. Negative edge matching differs from "no edge" matching. Matching is successful only when a negative edge is explicitly present in the graph. Unobserved or unverified relations do not meet the negative edge matching condition. Antecedent predicate verification: For attribute comparison predicates (such as transaction amount exceeding a threshold), variable comparison predicates (such as two accounts sharing a mobile phone number) and machine learning predicates (such as model score exceeding a threshold) in the rule antecedent, verification is performed based on the mapped entity attributes, edge attributes, and model output. Only when all predicates are satisfied is the matching instance output as a valid matching instance for subsequent inference.

[0062] In this application, the set of antecedent predicates and the set of consequent predicates may include the following predicate types: Attribute comparison predicates are used to compare the attribute values ​​of nodes or edges with preset thresholds or constants to determine their magnitude or equivalence. Operators include =, !=, >, >=, <, and <=. Operators such as greater than and less than, equal to and not equal to can form mutually exclusive or dual relationships for subsequent conflict detection. For example, `transaction.amount>50000`, `account.open_days<30`, `customer.risk_score>=80`, and `merchant.chargeback_rate>0.05`.

[0063] The variable comparison predicate is used to compare whether the attribute values between different nodes or edges meet the preset relation condition, and can express risk control features such as shared mobile phone number, shared address, similar transaction amount, and similar transaction time. For example, account1.phone=account2.phone, transaction1.amount~=transaction2.amount, transaction1.time-transaction2.time<1h.

[0064] The positive edge predicate is used to express a confirmed existing relation in the graph. For example, hold(customer, account) indicates that the customer holds an account, login(account, device) indicates that the account logs in to a device, transfer(account1, account2) indicates transfer between accounts, and same_entity(customer, blacklist_subject) indicates that the customer and the list subject are the same entity.

[0065] The negative edge predicate is used to express a relation that is confirmed not to exist or has been verified and excluded in the graph. For example, NOT same_owner(customer1, customer2), NOT fraud_collusion(account1, account2), NOT beneficial_owner(customer, enterprise), NOT same_entity(customer, blacklist_subject). A negative edge predicate is different from "edge not found". Only when a negative edge (u, ¬l, v) with an evidence source exists in the financial extended graph, can the matching instance satisfy the NOT l(u, v) predicate.

[0066] The machine learning predicate is used to compare the model prediction score with a preset threshold to express that the prediction result is valid. For example, M_fraud(account1, account2)>=eta_fraud indicates that the prediction score of the fraud model reaches the fraud threshold. The model can be graph neural network, gradient boosting decision tree, logistic regression, entity resolution model, etc.

[0067] The negative machine learning predicate is used to compare the model prediction score with a preset threshold to express that the prediction result is not valid. For example, M_fraud(account1, account2)<eta_fraud indicates that the prediction score does not reach the threshold, which can be used to exclude low-confidence risk relations or serve as a comparison basis during conflict detection.

[0068] 203. For each matching instance, verify whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiate the matching instances that pass the verification according to the consequent predicate of the financial inference rule to generate candidate inference results.

[0069] In some embodiments, the step "for each matching instance, verify whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiate the matching instances that pass the verification according to the consequent predicate of the financial inference rule to generate candidate inference results" may include the following operations: Obtain the set of antecedent predicates for the financial inference rule corresponding to the matching instance; verify whether the matching instance satisfies each predicate in the antecedent predicate set in order of computational cost from low to high; if any predicate is not satisfied, the matching instance is determined to have failed the verification and the inference for the matching instance is terminated; if all predicates are satisfied, the matching instance is determined to have passed the verification, and the consequent predicates of the financial inference rule are instantiated according to the node mapping in the matching instance to generate candidate inference results.

[0070] In this application, for each matching instance, the set of antecedent predicates X corresponding to the financial inference rule is first obtained. A matching instance is a mapping h: x→V from a rule variable to an entity node in the financial extended graph, satisfying all conditions for graph pattern matching (node ​​label compatibility, positive edge matching, and negative edge matching). Based on this, it is further verified whether the matching instance satisfies all predicates in the rule's antecedent predicate set.

[0071] Verification can be performed in ascending order of computational cost. Computational cost refers to the time and resource consumption required to verify various predicate types. Different predicate types exhibit significant differences in computational cost due to variations in data access, computational complexity, and external service calls. Specifically, the algorithm first verifies edge existence predicates and negative edge predicates. These two types of predicates only involve graph query operations and have the lowest computational cost. Then, it verifies attribute comparison predicates (e.g., whether the transaction amount exceeds a threshold) and variable comparison predicates (e.g., whether two accounts share a device). These two types of predicates involve reading and comparing attribute values ​​and have moderate computational cost. Finally, it verifies machine learning model predicates. These predicates require calling external model services for prediction calculations and have the highest computational cost. By verifying predicates in order of increasing cost, matching instances that do not meet the conditions can be identified early and terminated in a timely manner, avoiding high-cost model calls and thus reducing overall computational overhead.

[0072] If any predicate in the set of antecedent predicates is not satisfied, the matching instance is determined to have failed the verification, the reasoning for the matching instance is terminated, and the matching instance will no longer participate in subsequent steps.

[0073] If all predicates in the antecedent predicate set satisfy (i.e., h|=X), then the matching instance is deemed to have passed the verification. At this point, the consequent predicate p0 of the financial inference rule is instantiated according to the node mapping in the matching instance, that is, the variables in the consequent predicate are replaced with the entity nodes corresponding to h, resulting in the candidate inference result alpha=h(p0). The candidate inference result can be: a valid risk relationship (e.g., "Account A and Account B are associated with a fraud gang"), a invalid risk relationship (e.g., "Account A and Account B are not associated with a fraud gang"), valid entity identity (e.g., "Customer C and List Entity S are the same entity"), invalid entity identity (e.g., "Customer C and List Entity S are not the same entity"), a suspicious transaction, a transaction whose risk has been eliminated, or a customer having or not having an association with a high-risk entity.

[0074] For each candidate inference result, an inference record is generated, including rule number, rule version, matching entity, triggering predicate, rule confidence, rule support, model score, evidence source, inference time, and applicable task, for use in subsequent conflict detection, interpretation generation, and closed-loop update.

[0075] The candidate inference results include affirmative and negative conclusions. An affirmative conclusion indicates that the rule infers that a certain relationship or risk state is true; a negative conclusion indicates that the rule infers that a certain relationship or risk state is false. Affirmative and negative conclusions together constitute the set of candidate inference results, corresponding to positive and negative associations in the financial extension graph, respectively. Affirmative conclusions are typically generated by rules containing positive edge predicates or positive machine learning predicates, while negative conclusions are typically generated by rules containing negative edge predicates or negative machine learning predicates.

[0076] 204. Perform conflict detection and resolution on the candidate reasoning results to obtain the target risk reasoning results.

[0077] In some embodiments, the step "performing conflict detection and resolution on candidate inference results to obtain target risk inference results" may include the following operations: Retrieve multi-dimensional comparison benchmarks for conflict verification. These benchmarks include at least positive and negative correlation information within the financial extension graph, manually verified facts, model prediction results, and inference results generated by other rules. Based on each comparison benchmark, perform conflict detection on candidate inference results, identifying at least one of the following conflict types: conflict between positive and negative inference conclusions, mutual exclusion conflict between attribute comparison predicates, conflict between machine learning predicates and negative machine learning predicates, conflict between candidate inference results and manually verified facts, and mutual exclusion inference conflict between different candidate inference results. If a conflict is detected, calculate the priority score for each conflict candidate inference result according to preset conflict resolution strategy parameters. Resolve the conflict candidate inference results based on the priority scores to obtain resolution results. Resolution results include at least one of retaining positive inference, retaining negative inference, adjusting risk scores, marking as pending manual review, or adding to the observation list. Based on the resolution results, determine the target risk inference result.

[0078] In this context, positive and negative association information in the financial extension graph refers to confirmed and non-confirmed financial business relationships; manually verified facts refer to definitive conclusions generated through manual verification; model prediction results refer to risk scores output by machine learning models; and inference results generated by other rules refer to candidate conclusions generated by different financial inference rules. The aforementioned benchmarks cover factual evidence from three different sources: rule-based inference, manual verification, and model prediction, providing multi-dimensional references for conflict detection.

[0079] A conflict between positive and negative inference conclusions refers to the simultaneous existence of both positive and negative inferences for the same node pair and the same relation label. For example, rule A infers fraud_collusion(accountA, accountB) (account A and account B are associated with a fraud gang), while rule B infers NOT fraud_collusion(accountA, accountB) (account A and account B are not associated with a fraud gang), and the two are mutually exclusive.

[0080] Mutually exclusive conflicts between attribute comparison predicates refer to situations where the same entity attribute or relation attribute is inferred to have mutually exclusive values ​​or mutually exclusive intervals. For example, rule C infers that customer.risk_score >= 80 (the customer's risk level is not lower than 80 points), while rule D infers that customer.risk_score < 80 (the customer's risk level is lower than 80 points), which is contradictory.

[0081] Conflict between machine learning predicates and negative machine learning predicates: refers to that the same model predicate and its negative form hold true simultaneously. For example, rule E contains M_fraud(accountA, accountB)>=eta (the fraud model prediction score reaches the threshold), while rule F contains M_fraud(accountA, accountB)<eta (the fraud model prediction score does not reach the threshold), and the two are mutually exclusive.

[0082] Conflict between candidate inference results and manually verified facts: refers to that the rule inference result is inconsistent with the manual verification conclusion. For example, a rule infers fraud_collusion(accountA, accountB), but the manually verified fact is NOT fraud_collusion(accountA, accountB) (it has been confirmed that there is no fraud gang association), and the two are contradictory.

[0083] Mutually exclusive inference conflict between different candidate inference results: refers to that different rules give mutually exclusive conclusions for the same object. For example, rule R1 infers suspicious_flow(accountA, accountB, accountC) (there is a suspicious capital link), while rule R2 infers NOT suspicious_flow(accountA, accountB, accountC) (there is no suspicious capital link), and the two are mutually exclusive.

[0084] Wherein, the conflict resolution strategy parameters include at least one of the following: data source credibility, manual verification priority, rule confidence, model prediction score, number of evidences, evidence time freshness, regulatory rule mandatory priority and business scenario weight.

[0085] For the candidate conclusion alpha, the calculation method of the priority score priority(alpha) is as follows: priority(alpha)=w_source×source_score(alpha)+w_rule×rule_confidence(alpha)+w_model×model_score(alpha)+w_evidence×evidence_count(alpha)+w_time×freshness(alpha)+w_reg×regulation_priority(alpha) The parameters are defined as follows: source_score(alpha) represents the credibility of the data source. Different scores can be set for sources such as manual verification, regulatory lists, real-name verification, model inference, and weak rule inference. For example, the credibility of manual verification is higher than that of model inference, and model inference is higher than that of weak rule inference. rule_confidence(alpha) represents the confidence of the rule that generated the candidate inference result, with a value ranging from 0 to 1. model_score(alpha) represents the degree of support of the relevant machine learning model for the candidate inference result. evidence_count(alpha) represents the number of independent pieces of evidence that support the candidate inference result. freshness(alpha) represents the time freshness of the candidate inference result or evidence. Newer evidence is given a higher score. regulation_priority(alpha) represents the mandatory weight of high-priority rules such as regulatory rules, strong lists, and judicial freezes. w_source, w_rule, w_model, w_evidence, w_time, and w_reg are configurable weight coefficients for business purposes, and the importance of each factor can be adjusted according to different business scenarios.

[0086] Conflict resolution can follow these priority principles: highly credible manual verification results take precedence over automatic rule inference results; high-priority facts such as regulatory mandatory lists, judicial freezes, and real-name verification take precedence over general model predictions; under the same source, newer evidence takes precedence over older evidence; conclusions supported by multiple independent pieces of evidence take precedence over conclusions supported by a single weak piece of evidence; conclusions with both low rule confidence and model scores should be downgraded or placed under review; if the scores of positive and negative conclusions are close and involve high-risk business, a manual review should be output instead of an automatically given definitive conclusion.

[0087] The conflicting candidate reasoning results are resolved based on priority scores to obtain the resolution results. The resolution results include at least one of the following: retaining positive inferences (adding positive conclusions to the risk conclusion set), retaining negative inferences (adding negative conclusions to the risk exclusion set), adjusting risk scores (lowering or increasing risk scores according to the degree of conflict), marking as pending manual review (adding to the review queue), or including in the observation list (entering the gray list or observation list).

[0088] The target risk inference result is determined based on the resolution results. If the resolution result retains the positive inference, the target risk inference result is the conclusion that the corresponding risk relationship is valid; if it retains the negative inference, the target risk inference result is the conclusion that the corresponding risk relationship is invalid or has been excluded; if it adjusts the risk score, the target risk inference result is the adjusted risk score; if it is marked as pending manual review, the target risk inference result is in a pending manual review status; if it is included in the observation list, the target risk inference result is the conclusion that it is included in the observation list. For cases where the graph facts are not updated temporarily, only candidate inferences and conflict records are retained, and the final conclusion is not output.

[0089] 205. Output the target risk reasoning result and the corresponding reasoning explanation information.

[0090] The inference and explanation information includes at least one of the following: Final Conclusion Type: The conclusion category of the target risk reasoning result, such as "related to a fraud gang exists," "not related to a fraud gang exists," or "pending manual review." Hit Rule Information: The rule information that generated this conclusion, including rule number, name, version, confidence level, support level, and applicable task. Graph Pattern Instance: The specific financial entity to which the rule variable is bound, such as customer, account, transaction, or device. Positive and / or Negative Evidence: Positive evidence supporting the conclusion (shared equipment, fund transfers, etc.) or negative evidence (manual verification exclusion, KYC exclusion, etc.). Model Predicate Information: The model name, version, prediction score, threshold, and predicate judgment result involved. Conflict Information: Whether a conflict occurred, the type of conflict, the conflicting parties, and their sources. Resolution Basis: The priority strategy adopted, the credibility of the evidence, the freshness of the time, the rule confidence level, and the reason for the final selection. Recommended Actions: Automatic interception, manual review, continuous observation, alarm cancellation, priority reduction, supplementary investigation, etc.

[0091] In some embodiments, the reasoning explanation information can be output in the form of structured JSON, tables, graph path views, case reports, or rule hit details pages. By outputting the above explanation information, business personnel can be provided with an auditable chain of evidence, rule links, conflict resolution explanations, and final risk conclusions, meeting the requirements of decision interpretability in financial risk control scenarios.

[0092] In some embodiments, the method may further include the following steps: Write the objects marked as awaiting manual review into the manual review queue; obtain the manual review results corresponding to the objects awaiting manual review, as well as the business feedback data; update the negative association information in the financial extension graph, optimize the confidence level of the financial inference rules, and / or adjust the conflict resolution strategy of the resolution process based on the manual review results and business feedback data.

[0093] This application also includes a feedback and update mechanism, as detailed below: Objects marked as requiring manual review during conflict resolution are written into the manual review queue, and the corresponding manual review results and business feedback data are obtained. Business feedback data includes at least one of the following: alarm closure records (alarm event handling results), model iteration feedback data (model prediction accuracy verification feedback), and newly added business verification facts (new facts confirmed in subsequent verifications).

[0094] Based on the results of manual review and business feedback data, perform at least one of the following operations: Update the negative association information in the financial extension graph: write the exclusion relationship confirmed by manual review or the negative relationship confirmed in the newly added verification facts into the negative edge; Optimize the confidence of financial inference rules: revise the rule confidence statistics based on the review results; Adjust the conflict resolution strategy for conflict resolution processing: adjust the conflict resolution strategy parameters (such as at least one of the following: source credibility weight, rule confidence weight, model score weight, evidence quantity weight, time freshness weight, and regulatory enforcement priority weight) based on feedback.

[0095] Through the aforementioned feedback and update mechanism, the results of manual review and business feedback data are continuously written back to the financial extension graph and rule evaluation process, forming a complete closed loop of "reasoning → output → review → write-back → optimization", which enables the denial association information, rule confidence and conflict resolution strategies to be continuously iterated and optimized.

[0096] In this application, the thresholds, weights, and constraints involved in each of the above stages can be configured as needed. Optional parameters include, but are not limited to, the following: minimum rule support (min_support), used to filter rules with insufficient coverage samples; minimum rule confidence (min_confidence), used to filter rules with insufficient reliability; maximum conflict rate (max_conflict_rate), used to filter rules with too many conflicts with high-confidence facts; model predicate threshold (eta), which can be configured according to business tasks, model versions, and risk levels; negation edge confidence threshold (neg_conf_threshold), negation edges below this threshold are not used as strong negation predicates in inference; conflict resolution weights (w_source, w_rule, w_model, w_evidence, w_time, w_reg), used to adjust the importance of each factor in priority score calculation; maximum number of nodes and maximum number of edges in the graph pattern, used to control the search complexity in the candidate rule discovery process; rule effective time window, used to handle the issue of financial relationships and verification conclusions changing over time; and manual review trigger threshold, used to determine whether conflict conclusions are automatically resolved or enter the review queue.

[0097] The above parameters can be flexibly adjusted within the framework of this application according to specific business scenarios to adapt to the needs of different financial institutions, different business tasks and different risk preferences, without departing from the protection scope of this application.

[0098] This application discloses a financial risk reasoning method based on positive and negative associations. The method includes: acquiring a financial extended graph containing positive and negative association information, and a set of financial reasoning rules containing graph patterns, antecedent predicate sets, and consequent predicates, where both antecedent and consequent predicates contain predicates representing positive and negative associations; matching the rule graph patterns in the financial extended graph to obtain matching instances; verifying whether the matching instances satisfy the antecedent predicates, and instantiating the verified matching instances according to the consequent predicates to generate candidate reasoning results including affirmative or negative conclusions; performing conflict detection and resolution on the candidate reasoning results to obtain the target risk reasoning result; and outputting the target risk reasoning result and reasoning explanation information. This method achieves fusion reasoning with positive and negative evidence, reduces the false alarm rate through conflict detection and multi-dimensional priority resolution, and outputs auditable explanation information, meeting the needs of refined risk control and strong regulatory scenarios.

[0099] Based on the above description, the following examples will further illustrate the financial risk reasoning method based on positive and negative associations in this application. For example, please refer to... Figure 3 , Figure 3 A flowchart illustrating another financial risk reasoning method based on positive and negative correlations provided in this application embodiment. The complete flow of this financial risk reasoning method based on positive and negative correlations can be as follows: Obtain financial business data, verification data, and model prediction results, and construct a financial extended graph G=(V, E+, E-, L, FA) based on the above data. The financial extended graph includes nodes representing financial entities, positive edges representing confirmed relationships, negative edges representing confirmed non-existent or verified exclusion relationships, and attribute information for each node and edge.

[0100] At the rule level, financial GAR+ rules (i.e., financial graph association rules containing negative predicates) are defined. Each rule includes a graph pattern, a set of antecedent predicates, and an consequent predicate. Simultaneously, a candidate rule discovery and screening mechanism is used to generate candidate rules based on business templates, historical case mining, supervised learning model paths, or generative search. Quality indicators such as support, confidence, and conflict rate are calculated to obtain a set of effective rules for graph pattern matching and iterative rule optimization.

[0101] During the inference phase, graph pattern matching is performed on the rule's graph pattern within the financial extended graph to generate matching instances. For each matching instance, the antecedent predicates (positive edge predicate, negative edge predicate, attribute comparison predicate, variable comparison predicate, machine learning predicate, and negative machine learning predicate) are verified in ascending order of computational cost. If any predicate is not satisfied, the inference for that instance is terminated; if all are satisfied, the consequent predicate is instantiated, generating candidate risk inferences.

[0102] Subsequently, conflict detection is performed on the candidate risk inferences, comparing the positive and negative inference conclusions to identify at least one of the following: conflicts between positive and negative conclusions, attribute comparison conflicts, machine learning predicate conflicts, rule and manual verification conflicts, and multi-rule inference conflicts. If a conflict is detected, a priority score is calculated based on data source credibility, rule confidence, model prediction score, amount of evidence, time freshness, regulatory rule enforcement priority, and business scenario weight. The conflict is then resolved based on the priority score to obtain the target risk inference result.

[0103] Finally, risk conclusions and explanatory information are output, including rule hit records, graphical evidence chains, and conflict resolution basis. Objects marked for manual review are written into the manual review queue, and the manual review results and business feedback data are obtained and written back to the financial extension graph and rule evaluation process. This is used to update negation edges, optimize rule confidence, and adjust conflict resolution strategies, achieving a closed-loop iterative reasoning process.

[0104] The financial risk reasoning method based on positive and negative associations provided in this application can be applied to the following areas: anti-fraud gang identification, transaction fraud and credit card fraud identification, anti-money laundering suspicious fund flow identification, pre-loan risk control and credit approval, merchant risk identification, list screening and entity analysis, alarm review and case management, financial knowledge graph governance, and precision marketing and customer management.

[0105] The following three specific application scenarios further illustrate the implementation of this application.

[0106] (I) Example of Anti-fraud Gang Identification: Taking the reasoning of account fraud gang relationships as an example, the financial extended graph includes nodes such as customers, accounts, devices, mobile phone numbers, IPs, addresses, bank cards, merchants, transactions, and alarm events.

[0107] Positive edges include: accounts held by the customer, mobile phone numbers bound to accounts, account login devices, IP addresses used for accounts, transfers from one account to another, transactions conducted by accounts at merchants, and accounts that have been confirmed by historical cases as belonging to the same fraud gang. Negative edges include: manual verification confirming that account A and account B do not belong to the same fraud gang, customer appeals confirming that transaction T does not constitute fraud, device fingerprint verification confirming that device D does not belong to a high-risk device cluster, alarm closing confirming that account A does not match the fraud pattern M, and merchant verification confirming that the transaction dispute does not constitute fraud acquiring.

[0108] Taking the reasoning of account-fraud gang relationships as an example, the rules are as follows: phi1 = Q[a,b,d,t](transfer(a,b) AND login(a,d) AND login(b,d) ANDt.amount>50000 AND t.time_gap<3d AND neg_confidence(fraud_collusion,a,b)<theta_neg AND M_fraud(a,b)>= eta ->fraud_collusion(a,b)); This rule indicates that: if there is a transaction relationship between account a and account b, both accounts have logged in through the same device d, the transaction amount exceeds 50,000 yuan, the transaction time interval is less than 3 days, there is no high-confidence negative edge (a,¬fraud_collusion,b) in the financial expansion graph (or the confidence of this negative edge is lower than the current service threshold), and the prediction score of the fraud model for account a and account b exceeds the preset threshold, then it is inferred that there is a fraud gang association between account a and account b. By introducing the neg_confidence(fraud_collusion,a,b)<theta_neg condition, automatic inference of gang relationships can be avoided when there is already strong exclusion evidence.

[0109] Another exclusion rule is as follows: phi2 = Q[a,b](manual_review(a,b).result = excluded AND manual_review(a,b).confidence>= 0.95 ->NOT fraud_collusion(a,b)); When phi1 infers that account A and account B have a fraudulent relationship, and phi2 or human verification gives NOT fraud_collusion(A, B), the system triggers a conflict detection. If the credibility of the human verification is higher than the priority of the rule inference, the negative inference is retained, and the output is "The account pair has positive evidence of shared devices and transactions, but a highly credible human verification has already ruled it out. The final conclusion is that the fraudulent relationship is not established or needs to be reviewed."

[0110] (II) Example of identifying suspicious money laundering funding links: Taking the identification of suspicious fund flows as an example, the financial extension graph includes nodes such as customers, accounts, enterprises, legal persons, beneficial owners, transactions, regions, industries, blacklisted entities, and suspicious transaction reports.

[0111] Example rules are as follows: phi3 = Q[a, b, c, t1, t2, s](transfer(a, b) AND transfer(b, c) AND t1.amount~= t2.amount AND t2.time - t1.time<24h AND related_to(c, s) AND high_risk_subject(s) AND NOT same_entity(owner(a), s) AND M_aml(a, b, c)>= eta_aml ->suspicious_flow(a, b, c)); The rule states that if accounts a, b, and c form a short-term, closely related funding chain, the end account is related to a high-risk entity, and the model determines the chain to be suspicious, then the funding chain is inferred to be suspicious. At the same time, the rule explicitly checks the identity exclusion relationship of the list to avoid mistakenly using excluded entities with the same name as strong risk evidence.

[0112] The exclusion rule can be expressed as: phi4 = Q[c, e](kyc_check(c, e).result = not_beneficial_owner ANDevidence_source(c, e) in {KYC, KYB, manual_due_diligence} ->NOT beneficial_owner(c, e)); When the system detects a suspicious funding chain but also has a highly credible beneficial owner exclusion relationship, it can lower the risk score, mark it as pending review, or only retain the other unexcluded parts of the chain as clues for investigation.

[0113] (III) Example of pre-loan risk control: Taking the identification of implicit connections between applicants and high-risk entities as an example, the financial extension graph includes nodes such as applicant, loan application, enterprise, legal person, guarantor, contact person, device, address, overdue account, and blacklisted entity.

[0114] Example rules are as follows: phi5 = Q[x, e, d, p](apply_for(x, loan) AND login(x, d) AND login(p, d)AND overdue_subject(p) AND related_to(p, e) AND high_risk_enterprise(e) ANDNOT related_to(x, e) AND M_credit_default(x)>= eta_credit ->preloan_risk(x)) The rule states that if applicant x shares equipment with historically overdue entity p, p is associated with high-risk enterprise e, and the applicant has a high default model score, then it can be inferred that the applicant has pre-loan risk; however, if there is NOTrelated_to(x, e) that has been verified and excluded, then this negative evidence should enter the interpretation and conflict resolution process to avoid simply directly linking the applicant with high-risk enterprises.

[0115] To facilitate better implementation of the financial risk reasoning method based on positive and negative associations provided in this application, this application also provides a financial risk reasoning device based on positive and negative associations, which is based on the aforementioned financial risk reasoning method based on positive and negative associations. The meanings of the terms used are the same as in the aforementioned financial risk reasoning method based on positive and negative associations, and specific implementation details can be found in the descriptions in the method embodiments.

[0116] Please see Figure 4 , Figure 4A structural block diagram of a financial risk reasoning device based on positive and negative associations provided in this application embodiment is shown. The device includes: an acquisition unit 301, used to acquire a financial extension graph and a set of financial reasoning rules, wherein the financial extension graph includes at least positive association information and negative association information between financial entities, and each financial reasoning rule in the set of financial reasoning rules includes a graph pattern, an antecedent predicate set, and a consequent predicate, wherein both the antecedent predicate set and the consequent predicate contain predicates representing positive associations and predicates representing negative associations; and a matching unit 302, used to execute the graph pattern of each financial reasoning rule in the financial extension graph. The graph pattern matching process obtains matching instances corresponding to each financial inference rule. The verification unit 303 verifies whether each matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiates the verified matching instances according to the consequent predicate of the financial inference rule to generate candidate inference results, which include affirmative or negative conclusions. The processing unit 304 performs conflict detection and resolution processing on the candidate inference results to obtain the target risk inference result. The output unit 305 outputs the target risk inference result and the inference explanation information corresponding to the target risk inference result.

[0117] In some embodiments, the acquisition unit 301 may include: a first acquisition subunit, used to acquire the complete set of financial inference rules and the target business task; and a filtering subunit, used to filter a subset of rules that match the target business task from the complete set of financial inference rules, as a set of financial inference rules, based on the target business task; wherein the matching conditions include at least one of node type, relationship label, task label, business line, applicable region, time window, and regulatory strategy. In some embodiments, the apparatus may further include: a second acquisition unit, configured to acquire candidate rule generation strategies, the candidate rule generation strategies including at least one of a business template-based generation strategy, a historical risk case mining strategy, a supervised learning model path generation strategy, and a generative search generation strategy; a generation unit, configured to generate candidate financial inference rules according to the candidate rule generation strategies; a calculation unit, configured to calculate the quality indicators of each candidate financial inference rule, the quality indicators including at least one of support, confidence, coverage, conflict rate, consistency, and business interpretability; and a filtering unit, configured to filter rules that meet preset conditions from the candidate financial inference rules according to the quality indicators, to obtain a complete set of financial inference rules.

[0118] In some embodiments, the antecedent predicate set and consequent predicate include at least one of positive edge predicates, negative edge predicates, attribute comparison predicates, variable comparison predicates, machine learning predicates, and negative machine learning predicates; the verification unit 303 may include: a second acquisition subunit, used to acquire the antecedent predicate set of the financial inference rule corresponding to the matching instance; a verification subunit, used to sequentially verify whether the matching instance satisfies each predicate in the antecedent predicate set in order of computational cost from low to high; a first determination subunit, used to determine that the matching instance has failed the verification and terminate the inference of the matching instance if any predicate is not satisfied; and a second determination subunit, used to determine that the matching instance has passed the verification if all predicates are satisfied, and instantiate the consequent predicate of the financial inference rule according to the node mapping in the matching instance. The system generates candidate inference results, including: attribute comparison predicates, which compare the attribute values ​​of nodes or edges in the financial extended graph with preset thresholds or constants; variable comparison predicates, which compare whether the attribute values ​​of different nodes or edges in the financial extended graph satisfy preset attribute relationship conditions; positive edge predicates, which express confirmed relationships in the financial extended graph; negative edge predicates, which express confirmed non-existent or excluded relationships in the financial extended graph; machine learning predicates, which compare the prediction scores output by the machine learning model for financial entities or relationships with preset thresholds to express that the prediction result is valid; and negative machine learning predicates, which compare the prediction scores output by the machine learning model for financial entities or relationships with preset thresholds to express that the prediction result is invalid.

[0119] In some embodiments, the processing unit 304 may include: a retrieval subunit, used to retrieve multi-dimensional comparison benchmarks for conflict verification, the comparison benchmarks including at least positive and negative correlation information within the financial extended graph, manually verified facts, model prediction results, and inference results generated by other rules; a conflict detection subunit, used to perform conflict detection on candidate inference results based on each comparison benchmark, identifying at least one of the following conflict types: conflict between positive and negative inference conclusions, mutual exclusion conflict between attribute comparison predicates, conflict between machine learning predicates and negative machine learning predicates, conflict between candidate inference results and manually verified facts, and mutual exclusion inference conflict between different candidate inference results; a conflict resolution subunit, used to, if a conflict is detected, calculate the priority score of each conflict candidate inference result according to preset conflict resolution strategy parameters, and resolve the conflict candidate inference result according to the priority score to obtain a resolution result, the resolution result including at least one of retaining positive inference, retaining negative inference, adjusting risk score, marking as pending manual review, or adding to the observation list; and a determination subunit, used to determine the target risk inference result based on the resolution result.

[0120] In some embodiments, the apparatus may further include: a writing unit for writing objects marked as pending manual review into a manual review queue; a third acquisition unit for acquiring the manual review result corresponding to the object pending manual review, and business feedback data, wherein the business feedback data includes at least one of alarm closure records, model iteration feedback data, and newly added business verification facts; and an updating unit for updating the negative association information in the financial extension graph, optimizing the confidence level of the financial inference rules, and / or adjusting the conflict resolution strategy of the resolution process based on the manual review result and the business feedback data.

[0121] In some embodiments, the apparatus may further include: a fourth acquisition unit, configured to acquire financial business data, business verification data, and model prediction data; a first construction unit, configured to generate financial entity nodes based on the subject information contained in the financial business data, and construct positive association edges between corresponding entity nodes based on the confirmed business relationships between subjects in the financial business data; a second construction unit, configured to construct negative association edges between corresponding entity nodes based on the verified exclusion relationships between subjects recorded in the business verification data, and add verification-type attribute information to the negative association edges; an adding unit, configured to add model prediction-type attribute information to entity nodes, positive association edges, and negative association edges respectively according to the model prediction data; and a configuration unit, configured to configure semantic label information and business attribute information for all entity nodes, positive association edges, and negative association edges to complete the construction of the financial extension graph.

[0122] In some embodiments, the reasoning explanation information includes at least one of the following: the final conclusion type of the target risk reasoning result; rule information that hits the target risk reasoning result; graph pattern instance corresponding to the target risk reasoning result; positive evidence and / or negative evidence supporting the target risk reasoning result; model predicate information involved in the target risk reasoning result; conflict information corresponding to the target risk reasoning result; resolution basis corresponding to the target risk reasoning result; and suggested actions for handling the target risk reasoning result.

[0123] Accordingly, embodiments of this application also provide an electronic device. For example... Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 400 includes a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, and a computer program stored in the memory 402 and executable on the processor. The processor 401 and the memory 402 are electrically connected. Those skilled in the art will understand that... Figure 5 The electronic device structures shown herein do not constitute a limitation on electronic devices and may include, but are not limited to, those shown. Figure 5 It can show more or fewer parts, or combine certain parts, or arrange different parts.

[0124] The processor 401 is the control center of the electronic device 400. It connects various parts of the electronic device 400 through various interfaces and lines. By running or loading software programs and / or modules stored in the memory 402, and calling data stored in the memory 402, it performs various functions of the electronic device 400 and processes data, thereby monitoring the electronic device 400 as a whole.

[0125] In this embodiment, the processor 401 in the electronic device 400 loads the instructions corresponding to the processes of one or more applications into the memory 402 according to the following steps, and the processor 401 runs the applications stored in the memory 402 to realize various functions: Obtain the financial extension graph and the set of financial inference rules; perform graph pattern matching on the graph patterns of each financial inference rule in the financial extension graph to obtain matching instances corresponding to each financial inference rule; for each matching instance, verify whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiate the matching instances that pass the verification according to the consequent predicate of the financial inference rule to generate candidate inference results; perform conflict detection and resolution on the candidate inference results to obtain the target risk inference result; output the target risk inference result and the inference explanation information corresponding to the target risk inference result.

[0126] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0127] Optional, such as Figure 5 As shown, the electronic device 400 may further include a display 403 and an input unit 404. The processor 401 is electrically connected to both the display 403 and the input unit 404. Those skilled in the art will understand that... Figure 5 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0128] The display 403 can be used to display a graphical user interface (GUI) and receive operation commands generated by the user interacting with the GUI. The display 403 may include a display panel and a touch panel. The display panel can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the electronic device. These graphical user interfaces can be composed of graphics, guidance information, icons, video, and any combination thereof. Optionally, the display panel can be configured using a liquid crystal display (LCD), an organic light-emitting diode (OLED), or other similar devices. The touch panel can be used to collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel), generate corresponding operation commands, and execute the corresponding program. Optionally, the touch panel may include a touch detection device and a touch controller.

[0129] The touch detection device detects the user's touch location and the signal generated by the touch operation, transmitting the signal to the touch controller. The touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to the processor 401. It can also receive and execute commands from the processor 401. The touch panel can cover the display panel. When the touch panel detects a touch operation on or near it, it transmits the information to the processor 401 to determine the type of touch event. Subsequently, the processor 401 provides corresponding visual output on the display panel based on the type of touch event. In this embodiment, the touch panel and display panel can be integrated into the display 403 to achieve input and output functions. However, in some embodiments, the touch panel and display panel can be implemented as two independent components to achieve input and output functions. That is, the display 403 can also be used as part of the input unit 404 to achieve input functions.

[0130] The input unit 404 can be used to receive input numbers, characters, or user characteristic information (such as fingerprints, iris, facial information, etc.), and generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control.

[0131] In some embodiments, the electronic device may further include an audio circuit, which can provide an audio interface between the user and the electronic device via a speaker and a microphone. The audio circuit can convert received audio data into electrical signals and transmit them to the speaker, where the speaker converts them into sound signals for output. Conversely, the microphone converts collected sound signals into electrical signals, which are then received by the audio circuit, converted back into audio data, and output to processor 401 for processing. The audio data is then transmitted via radio frequency circuitry to, for example, another electronic device, or output to memory 402 for further processing. The audio circuit may also include an earphone jack to provide communication between peripheral headphones and the electronic device.

[0132] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0133] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of computer programs that can be loaded by a digital signal processor to execute steps in any of the financial risk reasoning methods based on positive and negative associations provided in embodiments of this application.

[0134] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0135] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0136] Since the computer program stored in the computer-readable storage medium can execute the steps in any of the financial risk reasoning methods based on positive and negative associations provided in the embodiments of this application, the beneficial effects that any of the financial risk reasoning methods based on positive and negative associations provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0137] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A financial risk reasoning method based on positive and negative correlations, characterized in that, The method includes: Obtain a financial extension graph and a set of financial inference rules, wherein the financial extension graph includes at least positive and negative association information between financial entities, and each financial inference rule in the set of financial inference rules includes a graph pattern, an antecedent predicate set and a consequent predicate, wherein both the antecedent predicate set and the consequent predicate contain predicates representing positive associations and predicates representing negative associations. The graph patterns of each financial inference rule are matched in the financial extended graph to obtain matching instances corresponding to each financial inference rule. For each matching instance, it is verified whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and the matching instance that passes the verification is instantiated according to the consequent predicate of the financial inference rule to generate a candidate inference result, which includes a positive conclusion or a negative conclusion. The candidate reasoning results are subjected to conflict detection and resolution to obtain the target risk reasoning result; Output the target risk reasoning result and the corresponding reasoning explanation information; Specifically, for each matching instance, verifying whether the matching instance satisfies the antecedent predicate of the corresponding financial inference rule, and instantiating the verified matching instances according to the consequent predicate of the financial inference rule to generate candidate inference results, includes: Obtain the set of antecedent predicates of the financial inference rule corresponding to the matching instance; In order of computational cost from low to high, verify whether the matching instance satisfies each predicate in the antecedent predicate set; If any predicate is not satisfied, the matching instance is determined to have failed the verification, and the reasoning for the matching instance is terminated. If all predicates are satisfied, the matching instance is determined to pass the verification. The consequent predicates of the financial inference rule are instantiated according to the node mapping in the matching instance to generate the candidate inference result.

2. The method according to claim 1, characterized in that, Obtain the set of financial reasoning rules, including: Obtain the complete set of financial reasoning rules and the target business task; Based on the target business task, a subset of rules that match the target business task is selected from the complete set of financial reasoning rules, and this subset is used as the financial reasoning rule set. The matching conditions include at least one of node type, relationship label, task label, business line, applicable region, time window, and regulatory strategy. Before obtaining the set of financial inference rules, the method further includes: Obtain candidate rule generation strategies, wherein the candidate rule generation strategies include at least one of the following: a business template-based generation strategy, a historical risk case mining strategy, a supervised learning model path generation strategy, and a generative search generation strategy. Based on the candidate rule generation strategy, candidate financial reasoning rules are generated; Calculate the quality metrics for each candidate financial inference rule, wherein the quality metrics include at least one of support, confidence, coverage, conflict rate, consistency, and business interpretability; Based on the quality indicators, rules that meet the preset conditions are selected from the candidate financial reasoning rules to obtain the complete set of financial reasoning rules.

3. The method according to claim 1, characterized in that, The set of antecedent predicates and the set of consequent predicates include at least one of positive edge predicates, negative edge predicates, attribute comparison predicates, variable comparison predicates, machine learning predicates, and negative machine learning predicates; The attribute comparison predicate is used to compare the attribute values ​​of nodes or edges in the financial extended graph with a preset threshold or constant, indicating a magnitude or equivalence relationship. The variable comparison predicate is used to compare whether the attribute values ​​between different nodes or edges in the financial extended graph satisfy a preset attribute relationship condition. The positive edge predicate is used to express a confirmed relationship in the financial extended graph. The negative edge predicate is used to express a confirmed non-existent or excluded relationship in the financial extended graph. The machine learning predicate is used to compare the prediction score output by the machine learning model for a financial entity or relationship with a preset threshold to indicate that the prediction result is valid. The negative machine learning predicate is used to compare the prediction score output by the machine learning model for a financial entity or relationship with a preset threshold to indicate that the prediction result is invalid.

4. The method according to claim 3, characterized in that, The process of performing conflict detection and resolution on the candidate reasoning results to obtain the target risk reasoning result includes: Retrieve a multi-dimensional comparison benchmark for conflict verification, wherein the comparison benchmark includes at least the positive correlation information and the negative correlation information in the financial extension graph, manually verified facts, model prediction results, and inference results generated by other rules; Based on each comparison benchmark, conflict detection is performed on the candidate inference results to identify at least one of the following conflict types: conflict between positive inference conclusion and negative inference conclusion, mutual exclusion conflict between the attribute comparison predicates, conflict between the machine learning predicate and the negative machine learning predicate, conflict between the candidate inference results and the manually verified facts, and mutual exclusion inference conflict between different candidate inference results. If a conflict is detected, the priority score of each conflict candidate reasoning result is calculated according to the preset conflict resolution strategy parameters. The conflict candidate reasoning results are then resolved according to the priority scores to obtain the resolution results. The resolution results include at least one of the following: retaining positive inference, retaining negative inference, adjusting risk score, marking as pending manual review, or including in the observation list. Based on the resolution results, the target risk reasoning result is determined.

5. The method according to claim 4, characterized in that, The method further includes: Write the objects marked as awaiting manual review into the manual review queue; Obtain the manual review result corresponding to the object to be manually reviewed, as well as the business feedback data, wherein the business feedback data includes at least one of alarm case closure records, model iteration feedback data, and newly added business verification facts; Based on the results of the manual review and the business feedback data, update the negative association information in the financial extension graph, optimize the confidence level of the financial inference rules, and / or adjust the conflict resolution strategy of the resolution process.

6. The method according to claim 1, characterized in that, Before obtaining the financial extension graph and the set of financial inference rules, the method further includes: Acquire financial business data, business verification data, and model prediction data; Financial entity nodes are generated based on the subject information contained in the financial business data, and positive association edges are constructed between the corresponding entity nodes based on the confirmed business relationships between the subjects in the financial business data. Based on the verified exclusion relationships between the entities recorded in the business verification data, negative association edges are constructed between the corresponding entity nodes, and verification class attribute information is added to the negative association edges; Based on the model prediction data, add model prediction class attribute information to the entity node, the positive association edge, and the negative association edge respectively; Configure semantic label information and business attribute information for all entity nodes, positive association edges, and negative association edges to complete the construction of the financial extension graph.

7. The method according to any one of claims 1-6, characterized in that, The reasoning and explanation information includes at least one of the following: The final conclusion type of the target risk reasoning results; Rule information that matches the target risk reasoning result; The graph pattern instance corresponding to the target risk reasoning result; Positive and / or negative evidence supporting the reasoning results of the target risk; The model predicate information involved in the target risk reasoning results; The conflict information corresponding to the target risk reasoning result; The basis for resolving the target risk reasoning results; Recommended actions to be taken based on the inference results of the target risk.

8. A financial risk reasoning system based on positive and negative correlations, characterized in that, The system includes: The data access module is used to access financial business data, business verification data, and model prediction data; The financial extended graph construction module is used to construct a financial extended graph based on the financial business data, the business verification data, and the model prediction data; the financial extended graph includes nodes representing financial entities, first-type edges representing confirmed relationships, second-type edges representing confirmed non-relationships, and attribute information of each node and each edge; The rule management module is used to store a set of financial reasoning rules. Each financial reasoning rule in the set of financial reasoning rules includes a graph pattern, a set of antecedent predicates, and an antecedent predicate. Both the set of antecedent predicates and the antecedent predicates contain predicates that represent positive associations and predicates that represent negative associations. The candidate rule discovery module is used to generate candidate financial inference rules according to the candidate rule generation strategy, calculate the quality index of each candidate financial inference rule, and filter the complete set of financial inference rules according to the quality index and store them in the rule management module. The graph pattern matching module is used to perform graph pattern matching on the financial extended graph for each financial inference rule, so as to obtain the matching instance corresponding to each financial inference rule. The predicate verification and reasoning module is used to verify whether each matching instance satisfies the antecedent predicate set of the corresponding financial reasoning rule, and to instantiate the matching instances that pass the verification according to the consequent predicate of the financial reasoning rule to generate candidate reasoning results, which include positive conclusions or negative conclusions. The conflict detection module is used to perform conflict detection on the candidate inference results and identify conflicts between the candidate inference results and positive correlation information, negative correlation information, manually verified facts, model prediction results and other candidate inference results in the financial extended graph. The conflict resolution module is used to calculate the priority score of each conflict candidate reasoning result according to the preset conflict resolution strategy parameters when a conflict is detected, determine the resolution result according to the priority score, and determine the target risk reasoning result based on the resolution result. An explanation generation module is used to output the target risk reasoning result and the reasoning explanation information corresponding to the target risk reasoning result; The feedback update module is used to write back the manual review results and business feedback data to the financial extension graph, update the negative association information in the financial extension graph, and / or optimize the confidence of the financial inference rules, and / or adjust the conflict resolution strategy of the conflict resolution module.

9. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the financial risk reasoning method based on positive and negative associations as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the financial risk reasoning method based on positive and negative associations as described in any one of claims 1 to 7.

Images