Risk control methods, devices, equipment, and media for transaction systems based on relational knowledge graphs
By using a knowledge graph-based risk control method for transaction systems, and leveraging an anti-fraud decision engine and multi-hop queries, combined with a risk propagation attenuation coefficient and a three-level circuit breaker mechanism, the problem of isolated and static early warning systems in financial anti-fraud systems has been solved, enabling rapid and automated handling of group fraud.
Patent Information
- Application Number
- CN202611307726.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
Existing financial anti-fraud systems suffer from isolated early warning systems, static correlation analysis, and reliance on manual intervention, making it difficult to automatically execute tiered blocking operations for single-point high-risk early warnings.
The risk control method for transaction systems based on relational knowledge graphs obtains transaction flow data, uses an anti-fraud decision engine for rule matching, performs multi-hop queries on related entities, generates a risk propagation attenuation coefficient by combining edge attribute information, and automatically executes disposal strategies using a three-level circuit breaker mechanism.
It enables rapid identification and automated handling of group fraud, improves prevention and control efficiency, reduces time lag and manual intervention, and achieves real-time risk management.
Smart Images

Figure CN123118256A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial risk control, and in particular to risk control methods, devices, equipment and media for transaction systems based on relational knowledge graphs. Background Technology
[0002] Current anti-fraud systems in financial institutions primarily rely on rule engines and machine learning models to detect illegal transactions by individual users, which are characterized by networking, correlation, and speed. Knowledge graph technology, due to its ability to characterize entity relationships, has been increasingly applied to the identification of organized fraud.
[0003] Traditional methods have the following limitations: 1. Isolated early warning: After a high-risk user is discovered, accomplices may continue to commit crimes, resulting in a time lag in prevention and control.
[0004] 2. Staticization of association analysis: Existing knowledge graph technologies mostly involve batch scanning of the entire graph or static risk assessment based on the similarity of node attributes.
[0005] 3. Response relies on human intervention: Human intervention is required from early warning to response.
[0006] 4. The strategy lacks self-evolution: the parameter configuration of rules and models relies on expert experience or offline optimization.
[0007] In summary, how to automatically execute tiered blocking operations based on a single high-risk early warning is a problem that urgently needs to be solved. Summary of the Invention
[0008] In view of this, the purpose of this invention is to provide a risk control method, apparatus, device, and medium for a transaction system based on a relational knowledge graph, capable of automatically executing tiered blocking operations from a single high-risk early warning point. The specific solution is as follows: Firstly, this application provides a risk control method for a transaction system based on a relational knowledge graph, including: The system acquires transaction flow data from the target transaction system and uses an anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules to identify target entities in the association graph that meet preset fraud conditions. The association graph is a graph that shows the association relationship between various accounts in the target transaction system. The association graph includes entities representing accounts and association paths representing association relationships. Starting from the target entity, perform multi-hop queries on the association graph to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path; Based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, a risk propagation attenuation coefficient corresponding to each connecting edge on the associated path is generated. Based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path, the infection risk value corresponding to each associated entity is determined. Based on a preset three-level circuit breaker mechanism, it is determined whether the infection risk value is within a preset infection value range. Based on the determination result, the target associated entity that meets the conditions for gang fraud is identified, and the corresponding preset handling strategy is executed on the target associated entity and an alarm is triggered in the target transaction system.
[0009] Optionally, the step of performing multi-hop queries on the association graph, starting from the target entity, to obtain associated entities that have an association path with the target entity includes: Breadth-first queries are initiated simultaneously from the target entity and preset high-frequency nodes in the association graph. The query stops when the query results on both sides intersect, and the associated entity is obtained. And / or, before performing the multi-hop query, the connection edges in the association graph are filtered according to the creation time and / or interaction frequency of the connection edges to obtain a filtered association graph, and the multi-hop query is performed on the filtered association graph to obtain the associated entity.
[0010] Optionally, the edge attribute information includes the edge type, the frequency of interaction between the entities connected by the edge, and the time interval between the interaction between the entities connected by the edge. Accordingly, generating risk propagation attenuation coefficients corresponding to each connecting edge on the associated path based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path includes: For each connecting edge on the associated path, the corresponding risk propagation weight is determined according to the edge type; The frequency decay factor is determined based on the interaction frequency, and the time decay factor is determined based on the interaction time interval. Based on the risk propagation weight, the frequency decay factor, and the time decay factor, the risk propagation decay coefficient corresponding to the connection edge is generated.
[0011] Optionally, determining the corresponding frequency decay factor based on the interaction frequency and the corresponding time decay factor based on the interaction time interval includes: The frequency of interaction is normalized using a logarithmic function to obtain the frequency attenuation factor. The time decay factor is generated by calculating the exponential decay function of Newton's law of cooling for the interaction time interval.
[0012] Optionally, determining the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connection edge on the associated path includes: The preset risk score of the target entity and the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path are multiplied together based on the hop count of the associated path to obtain the infection risk value corresponding to the associated entity.
[0013] Optionally, the step of determining whether the infection risk value is within a preset infection value range based on a preset three-level circuit breaker mechanism, identifying the target associated entity that meets the conditions for gang fraud based on the determination result, and executing the corresponding preset handling strategy on the target associated entity and issuing an alarm in the target transaction system includes: If the infection risk value is within the first preset range, the corresponding target associated entity will be added to the monitoring list database to increase the frequency of feature calculation for the target associated entity and increase the dimension of collecting the behavioral trajectory of the target associated entity. If the infection risk value is within the second preset range, the corresponding target associated entity is added to the rule list library. When the target associated entity makes its next transaction, an enhanced identity authentication operation is triggered for the target associated entity, and the single transaction amount of the target associated entity is limited. Wherein, the infection risk value corresponding to the second preset range is greater than the infection risk value corresponding to the first preset range. If the infection risk value is within the third preset range, the account of the target associated entity is frozen, the transaction permissions of the target associated entity are suspended, and an alarm message is pushed to the risk control operation platform; wherein, the infection risk value corresponding to the third preset range is greater than the infection risk value corresponding to the second preset range.
[0014] Optionally, after executing the corresponding preset handling strategy on the target associated entity and issuing an alarm in the target transaction system, the method further includes: Obtain group activity data associated with the target entity and the associated entity within a preset time period, and determine the activity stagnation index based on the group activity data; Based on the activity stagnation index, determine whether the effect of the preset handling strategy meets the preset conditions; If not, then adjust the attenuation constant corresponding to the risk propagation weight and / or time attenuation factor used in the risk propagation attenuation coefficient.
[0015] Secondly, this application provides a risk control device for a transaction system based on a relational knowledge graph, comprising: The entity determination module is used to acquire transaction flow data from the target transaction system and use an anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules to determine the target entities in the association graph that meet the preset fraud conditions. The association graph is a graph that shows the association relationship between various accounts in the target transaction system. The association graph includes entities representing accounts and association paths representing association relationships. The information acquisition module is used to perform multi-hop queries on the association graph starting from the target entity to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path. The risk value determination module is used to generate a risk propagation attenuation coefficient corresponding to each connecting edge on the associated path based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, and to determine the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path. The transaction alarm module is used to determine whether the infection risk value is within a preset infection value range based on a preset three-level circuit breaker mechanism, identify the target associated entity that meets the conditions for gang fraud based on the determination result, execute the corresponding preset handling strategy on the target associated entity and issue an alarm in the target transaction system.
[0016] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned risk control method for a transaction system based on a relational knowledge graph.
[0017] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned risk control method for a transaction system based on a relational knowledge graph.
[0018] In summary, this application acquires transaction flow data from the target transaction system, uses an anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules, and identifies target entities in the association graph that meet preset fraud conditions. The association graph is a graph showing the relationship between various accounts in the target transaction system, and includes entities representing accounts and association paths representing the relationship. Starting from the target entity, multi-hop queries are performed on the association graph to obtain associated entities that have association paths with the target entity, as well as edge attribute information of each connecting edge on the association path. Based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the association path, a risk propagation attenuation coefficient corresponding to each connecting edge on the association path is generated. Based on the risk propagation attenuation coefficient corresponding to each connecting edge on the association path, the contagion risk value corresponding to each associated entity is determined. Based on a preset three-level circuit breaker mechanism, it is determined whether the contagion risk value is within a preset contagion value range. Based on the determination result, the target associated entity that meets the group fraud conditions is identified, and the corresponding preset handling strategy is executed on the target associated entity and an alarm is triggered in the target transaction system. As shown above, this application depicts the relationships between accounts based on a relational graph. After obtaining transaction data from the transaction system, it uses an anti-fraud decision engine to perform rule matching on the transaction data, thereby locating target entities in the graph that meet the fraud criteria. Subsequently, starting from the target entity, it expands along the graph in multiple hops, collecting attribute information of related entities and connecting edges. It then combines the inherent risk score of the target entity with the attributes of each edge to assign a risk propagation attenuation coefficient to each edge, thereby quantifying the contagion risk of each related entity. Finally, a three-level circuit breaker mechanism is introduced to classify the contagion risk within a range, identifying target related entities that need to be dealt with as part of a group fraud scheme. Simultaneously, it triggers system alarms and executes corresponding control strategies. This allows for rapid spread to the entire related network and automatic handling after the discovery of a single high-risk individual, significantly improving the efficiency of preventing group fraud. Furthermore, the three-level automatic circuit breaker strategy enables rapid isolation and crackdown before fraud groups can react. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0020] Figure 1 This application discloses a flowchart of a risk control method for a transaction system based on a relational knowledge graph. Figure 2This application discloses a specific three-level circuit breaker strategy flowchart; Figure 3 This application discloses a specific overall architecture diagram of risk control for a transaction system based on a relational knowledge graph. Figure 4 This is a schematic diagram of the structure of a risk control device for a transaction system based on a relational knowledge graph, as disclosed in this application. Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] Currently, anti-fraud systems in financial institutions primarily rely on rule engines and machine learning models to score the risk of individual transactions or users. Illegal transactions exhibit characteristics of being networked, interconnected, and rapid. Knowledge graph technology, due to its ability to characterize entity relationships, has been gradually applied to identifying organized fraud. Traditional methods suffer from the following limitations: isolated early warning systems, static correlation analysis, reliance on manual intervention, and a lack of self-evolving strategies. To address these technical issues, this application discloses a transaction system risk control method, device, equipment, and medium based on a relational knowledge graph, capable of automatically executing tiered blocking operations from a single high-risk early warning point.
[0023] See Figure 1 As shown in the figure, this invention discloses a risk control method for a transaction system based on a relational knowledge graph, including: Step S11: Obtain transaction flow data from the target transaction system, and use the anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules to determine the target entity in the association graph that meets the preset fraud conditions; the association graph is a graph that shows the association relationship between each account in the target transaction system, and the association graph includes entities representing accounts and association paths representing association relationships.
[0024] In this embodiment, the target transaction system can be a bank's core transaction system, a third-party payment platform, an internet consumer finance platform, a securities trading platform, or any other business system with an account system and fund transfer capabilities. Transaction flow data refers to the structured record generated by the target transaction system during its business operations, used to describe a fund transfer or an account behavior. It typically includes, but is not limited to, the following fields: transaction number, transaction time, payer account identifier, payee account identifier, transaction amount, transaction currency, transaction channel, transaction type, merchant number, device fingerprint used when the transaction was initiated, network protocol address, geographic location information, and transaction result status.
[0025] In one specific embodiment, transaction flow data can be captured in real time and written to a message queue, for example, to a designated topic in a Kafka message middleware, by deploying data acquisition probes along the transaction chain of the target transaction system; alternatively, transaction flow data can be obtained by subscribing to the change data capture log of the target transaction system database. Compared to timed batch extraction, the above real-time acquisition method can control the delay from data generation to availability to within seconds, thereby providing a data foundation for subsequent real-time risk management.
[0026] The association graph is a graph illustrating the relationships between various accounts in the target transaction system. The association graph includes entities representing accounts and association paths representing the relationships. It should be noted that the entities in this embodiment are not limited to the accounts themselves, but may further include media entities that are bound to or used by the accounts, such as device entities, mobile phone number entities, ID card entities, bank card entities, shipping address entities, network protocol address entities, and merchant entities. The association path is formed by one or more connecting edges, each connecting edge representing a direct association between two entities.
[0027] In one specific implementation, the edge types of the association graph may include, but are not limited to: transfer relationship edges, indicating that two account entities have had financial transactions; shared device relationship edges, indicating that two account entities have logged in or initiated transactions on the same device; same contact person relationship edges, indicating that two account entities have overlap in their address books or emergency contact information; shared delivery address relationship edges, indicating that two account entities have used the same delivery address; same network protocol address relationship edges, indicating that two account entities used the same network exit address within the same time window; and card binding relationship edges, indicating that two account entities have bound the same bank card or the same ID number. These various edge types respectively depict the relationships across the dimensions of funds, devices, social interaction, logistics, network, and identity, collectively forming a multi-dimensional and comprehensive portrayal of the group's relationships.
[0028] In this embodiment, the association graph is stored using a graph database. In a preferred embodiment, the graph database is Neo4j. Neo4j uses a native graph storage structure, where adjacent nodes are directly connected via physical pointers. Therefore, it eliminates the need for costly join operations during multi-hop traversal, maintaining stable traversal performance even when the number of relationships reaches billions. Furthermore, the Cypher query language provided by the graph database has excellent expressive power, enabling the declarative description of complex graph pattern matching logic. Of course, those skilled in the art will understand that any other graph database product that supports attribute graph models and multi-hop traversal can also be used.
[0029] It is understood that the construction process of the relational graph can extract account information, device information, transaction information, and binding relationship information from the business database, log system, and external data sources of the target transaction system; clean and standardize the extracted data, such as unifying the mask of mobile phone numbers, normalizing device fingerprints, and segmenting and standardizing address text; perform entity alignment based on the cleaned data, merging multiple records of the same real-world object in different data sources into the same entity in the graph; generate connection edges according to preset relationship extraction rules, and write edge attribute information for each connection edge; and batch write entities and connection edges into the graph database. After the graph is constructed, its real-time nature can be maintained through streaming incremental updates, that is, when the transaction system generates new transactions or new binding behaviors, the corresponding connection edges and their attributes are added or updated in the graph in real time.
[0030] The anti-fraud decision engine is a risk assessment component deployed in the transaction chain of the target transaction system, and it maintains a set of preset rules internally. The preset rules can be expert experience rules, such as "the same device logs into more than five different accounts within one hour", "the account has a large amount of money transferred in within 24 hours after opening the account and then transferred out in a scattered manner", "the account's cumulative daily transfer amount exceeds ten times the historical average", "the account makes multiple consecutive integer amount payments to merchants known to be involved in illegal transactions", etc. The preset rules can also be the output judgment conditions of a machine learning model, such as "the fraud probability output by the gradient boosting tree model is higher than 0.9".
[0031] The step of using an anti-fraud decision engine to perform rule matching on each transaction data according to preset rules to determine the target entity in the association graph that meets the preset fraud conditions can specifically include: concatenating the transaction data with pre-calculated account profile features and device profile features to obtain the feature vector required for rule matching; inputting the feature vector into the anti-fraud decision engine, which then executes the condition judgment of each preset rule sequentially or in parallel; when any preset rule is hit, or the weighted score of multiple hit rules exceeds a preset threshold, it is determined that the account corresponding to the transaction data meets the preset fraud conditions; and the entity corresponding to that account in the association graph is determined as the target entity.
[0032] Furthermore, while identifying the target entity, a preset risk score can be assigned to it. This preset risk score characterizes the severity of the risk posed by the target entity, and its value can be normalized to [0,1]. The preset risk score can be mapped based on the severity level of the triggered rule. For example, when the rule "the account has been listed as an account involved in a case by relevant departments" is triggered, the preset risk score is 1.0; when the rule "the account is suspected of cashing out" is triggered, the preset risk score is 0.7. In another specific implementation, the preset risk score can be directly taken from the fraud probability value output by the machine learning model.
[0033] Step S12: Starting from the target entity, perform a multi-hop query on the association graph to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path.
[0034] In this embodiment, after determining the target entity, a breadth-first multi-hop nearest neighbor query is performed on the association graph centered on the target entity to obtain all associated entities and connection edge information within a range of one to N hops. Here, N is a preset maximum number of hops, the value of which can be configured according to business scenarios and performance constraints. In one specific implementation, N is set to three. This is because in financial risk control practice, associations within three hops usually have strong business interpretability, while beyond three hops, the number of associated entities grows exponentially, and the actual risk correlation between them and the source node becomes very weak.
[0035] It should be noted that this step needs to complete the multi-hop neighbor query within milliseconds. Traditional graph database traversal queries can suffer from excessive latency due to the presence of supernodes or increased depth. A supernode refers to a node with an extremely high degree in the graph, such as a public network exit address shared by millions of users, or a merchant node on a well-known e-commerce platform with a large number of users. When a traversal path passes through a supernode, the number of its neighbors explodes instantly, causing a fan-out explosion at that node, leading to a query storm. This degrades the time taken for a single query from milliseconds to seconds or even minutes, completely failing to meet the requirements of online real-time processing. Therefore, this embodiment provides the following three optimization methods.
[0036] In one specific embodiment, the optimization method is an accelerated query based on bidirectional breadth-first search and materialized views. Breadth-first queries are initiated simultaneously from the target entity and preset high-frequency nodes in the association graph. The query stops when the query results from both sides intersect, thus obtaining the associated entity. Specifically, the preset high-frequency nodes refer to nodes in the association graph that are frequently accessed or have high degrees, such as popular merchant nodes or popular device nodes. When using bidirectional breadth-first search, the search expands simultaneously from both ends of the target entity and the preset high-frequency nodes towards the middle. Each end only needs to expand approximately d / 2 layers to meet in the middle, thus reducing the time complexity to approximately [missing value]. For example, if the average degree of the graph is 100 and the query hop count is four, then a one-way search needs to visit about 100 million nodes, while a two-way search only needs to visit about 20,000 nodes, a difference of about 5,000 times, thus reducing the query time by an order of magnitude.
[0037] The execution process of bidirectional breadth-first search can initialize two access sets, denoted as the forward set and the backward set. The initial element of the forward set is the target entity, and the initial element of the backward set is the preset high-frequency node. In each iteration, the set with fewer current elements is selected for expansion, that is, all neighbor nodes of all nodes in the set are added to the set. After each expansion, it is determined whether there is an intersection between the forward set and the backward set. If there is an intersection, it means that the two ends have met. At this time, the complete association path can be restored by backtracking, and the query stops. If there is no intersection and the total number of expanded layers has not reached the preset maximum number of hops, the next iteration continues.
[0038] Furthermore, this embodiment also employs materialized views to accelerate common query patterns. For common query patterns, the results are pre-calculated and stored in the materialized view. During a query, the materialized view is queried first, and only when the materialized view becomes invalid does it fall back to online traversal, thereby significantly reducing the amount of online computation.
[0039] Historical query logs are analyzed to identify the most frequent query patterns. For each identified query pattern, offline batch computation is used to pre-calculate all entity pairs in the graph that satisfy the pattern, along with their path information. The calculation results are stored in a cache as key-value pairs, where the key is the identifier of the starting entity, and the value is a list of all target entities and path information that the entity can reach through the pattern. The materialized view can be updated using a strategy combining incremental updates and timed full reconstruction. When edges are added or deleted in the graph, only the affected materialized view entries are partially updated. Simultaneously, a full reconstruction of the materialized view is performed every preset period, for example, every six hours, to eliminate errors that may accumulate during incremental updates. Furthermore, a validity period timestamp can be set for each materialized view entry. When a query-triggered entry has expired, it is determined to be invalid, and the online traversal is rolled back, triggering an asynchronous reconstruction of the entry.
[0040] In another specific embodiment, the optimization method is a supernode-oriented edge pruning strategy. Before performing the multi-hop query, the connecting edges in the association graph are filtered based on their creation time and / or interaction frequency to obtain a filtered association graph. The multi-hop query is then performed on the filtered association graph to obtain the associated entity. Specifically, before traversal, a large number of irrelevant edges are quickly filtered out based on attributes such as edge creation time and weight. For example, only connecting edges whose creation time is within the last ninety days and whose interaction frequency is greater than two are retained, thereby preventing the supernode from causing a query storm.
[0041] It's important to note that fraud rings typically exhibit a highly time-sensitive organizational structure, with relationships between members often forming within a short period. Furthermore, genuine group collaboration usually involves multiple repetitive interactions, while relationships based on a single interaction are likely due to chance, such as two strangers accidentally connecting to the same public Wi-Fi network. Therefore, by using the lightweight filtering conditions of time window and interaction frequency, we can filter out the vast majority of noisy edges with minimal computational cost, while retaining the majority of genuine group relationship edges.
[0042] To quickly filter out a large number of irrelevant edges, indexes can be pre-created in the graph database for the creation time and interaction frequency attributes of connecting edges. This allows the filtering conditions to be determined at the index level without loading the complete attribute data of the edges. Furthermore, for identified supernodes, a pre-set upper limit on the number of neighbors can be set. For example, it can be stipulated that a single node can expand outwards by a maximum of one thousand neighbors during traversal, and the edges can be truncated in descending order of weight, thus providing a deterministic upper bound guarantee for query time.
[0043] In another specific embodiment, the optimization method is query template compilation optimization based on the Cypher query language. This embodiment pre-compiles multi-hop query logic into a parameterized query template for the graph database. During actual querying, only the entity identifier parameters are replaced, without re-parsing and compiling the query statement, thus avoiding the additional overhead of repeated compilation. In the execution process of a graph database, the processing of a query statement typically involves multiple stages, including lexical analysis, syntax analysis, semantic verification, execution plan generation, execution plan optimization, and actual execution. The first five stages are collectively referred to as the compilation stage. For queries with fixed structures but variable parameters, the compilation stage often accounts for a significant proportion of the total execution time. By using parameterized template pre-compilation, the compilation results can be cached and reused in subsequent queries, allowing each query to only execute the final actual execution stage. For example, after using parameterized template pre-compilation, the average execution time of a single three-hop query is reduced from approximately forty milliseconds to approximately twelve milliseconds.
[0044] Step S13: Based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, generate the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path, and determine the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path.
[0045] In this embodiment, for each associated entity found, the risk value inherited from the target entity along the propagation path is calculated. It should be noted that this step requires quantifying the dynamic attenuation process of risk propagation along the edge, rather than simply performing a static score transfer. The calculation of the risk propagation attenuation coefficient comprehensively considers the following factors.
[0046] The first factor is the source node risk score. The source node risk score, which is the preset risk score of the target entity, is used to characterize the original risk severity of the high-risk entity itself.
[0047] The second factor is the edge type weight. For each connection edge on the associated path, a corresponding risk propagation weight is determined according to the edge type. Based on the risk propagation strength of the association, corresponding weight values are preset for different edge types, and a dynamic mapping table is maintained. The weight of the transfer relationship edge is 0.8, the weight of the shared device relationship edge is 0.6, and the weight of the same contact person relationship edge is 0.5.
[0048] The above weighting is based on the objective differences in the risk transmission intensity reflected by different relationships: the transfer relationship directly reflects the fund transfers between two accounts and is the most direct evidence of gang cooperation, so it is given the highest weight; the shared device relationship indicates that two accounts have been operated by the same physical device. Although normal situations such as family members sharing devices cannot be ruled out, in gang crime scenarios, batch registration and batch operation are often concentrated on a few devices, so this relationship also has strong indicative significance; the same contact person relationship is relatively weak, because the overlap of two users' address books may only mean that they have mutual friends, and has no necessary connection with fraud cooperation.
[0049] It should be noted that the dynamic mapping table is stored in the configuration center in the form of key-value pairs, where the key is the edge type identifier and the value is the corresponding weight value. Risk control operations personnel can adjust the weight values online through the management interface of the configuration center. The adjustment results are sent to each computing node in real time through the configuration push mechanism and take effect without restarting the service.
[0050] The third factor is the frequency attenuation factor. The corresponding frequency attenuation factor is determined based on the interaction frequency. The higher the interaction frequency, the weaker the attenuation of risk propagation, meaning the risk is more easily transmitted between the two entities. This embodiment uses a logarithmic function for normalization, and its calculation formula is as follows: ; in, freq This indicates the frequency of interactions between the entities connected by the edge.
[0051] The fourth factor is the time decay factor. The corresponding time decay factor is determined based on the interaction time interval. The time decay factor is calculated based on the time interval of the most recent interaction between entities; the longer the time interval, the stronger the decay of risk transmission. This embodiment uses the exponential decay function of Newton's law of cooling for calculation: ; Where λ is the decay constant and Δt is the time interval in hours.
[0052] This embodiment employs a full-path risk cumulative multiplication model, meaning that for a K-hop neighbor, the risk decays hop by hop along the path, which aligns with the objective laws of risk propagation. Based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, a risk propagation attenuation coefficient is generated for each connecting edge on the associated path. This includes: determining the corresponding risk propagation weight for each connecting edge on the associated path based on the edge type; determining the corresponding frequency attenuation factor based on the interaction frequency; determining the corresponding time attenuation factor based on the interaction time interval; and generating the risk propagation attenuation coefficient corresponding to the connecting edge based on the risk propagation weight, the frequency attenuation factor, and the time attenuation factor. Specifically, the calculation formula for the single-step attenuation coefficient, i.e., the risk propagation attenuation coefficient corresponding to a single connecting edge, is as follows: ; Here, i refers to a step in a multi-hop query. For example, if you query from a node along an edge to find the first neighbor, and then query the neighbor's neighbor along an edge from the first neighbor, i would refer to 1 or 2.
[0053] Furthermore, the preset risk score of the target entity and the risk propagation attenuation coefficients corresponding to each connecting edge on the associated path are multiplied together based on the hop count order of the associated path to obtain the infection risk value corresponding to the associated entity. Specifically, the infection risk value of the associated entity is equal to the product of the source node's risk score and the propagation coefficients of all edges on the path, and its calculation formula is as follows: ; in, R0 represents the infection risk value corresponding to the associated entity located at the k-th hop; R0 represents the preset risk score of the target entity; k represents the number of hops in the associated path; This indicates a series multiplication operation.
[0054] It should be noted that when there are multiple different association paths between the target entity and a certain associated entity, the final infection risk value of the associated entity can be determined as follows: calculate the infection risk value corresponding to each association path separately, and take the maximum value as the final infection risk value of the associated entity. Alternatively, a probability merging method can be used to determine the final infection risk value.
[0055] Furthermore, to control the computational scale, a pruning mechanism can be introduced during the computation process. Specifically, during breadth-first expansion, once the cumulative infection risk value of an intermediate node falls below a preset minimum handling threshold (e.g., below 0.3), expansion to deeper levels along that node ceases. This is because in the multiplicative model, the infection risk value monotonically decreases with the number of hops. If the risk value of an intermediate node is already below the minimum handling threshold, the risk values of all its subsequent downstream nodes will inevitably be even lower, and no handling action will be triggered. Therefore, further expansion to these nodes is meaningless. By introducing this pruning mechanism, the number of nodes that need to be accessed can be significantly reduced without losing any effective handling results.
[0056] Step S14: Based on the preset three-level circuit breaker mechanism, determine whether the infection risk value is within the preset infection value range, determine the target associated entity that meets the conditions for gang fraud according to the determination result, and execute the corresponding preset disposal strategy on the target associated entity and issue an alarm in the target transaction system.
[0057] In this embodiment, if the infection risk value is within a first preset range, the corresponding target associated entity is added to the monitoring list, increasing the frequency of feature calculation for the target associated entity and adding dimensions for collecting the behavioral trajectory of the target associated entity; if the infection risk value is within a second preset range, the corresponding target associated entity is added to the rule list, and when the target associated entity makes its next transaction, an enhanced identity authentication operation is triggered for the target associated entity, and the single transaction amount of the target associated entity is limited; wherein, the infection risk value corresponding to the second preset range is greater than the infection risk value corresponding to the first preset range; if the infection risk value is within a third preset range, the account of the target associated entity is frozen, the transaction permissions of the target associated entity are suspended, and an alarm message is pushed to the risk control operation platform; wherein, the infection risk value corresponding to the third preset range is greater than the infection risk value corresponding to the second preset range. Specifically, as shown... Figure 2 As shown, based on the calculated infection risk value Automatically match and execute a three-level handling strategy based on preset thresholds: Observation level (0.3≤ <0.5): Silently add associated entities to the monitoring list, increase the feature calculation frequency to every minute, and add collection dimensions such as behavioral trajectories.
[0058] Challenge level (0.5≤ <0.7): Add the associated entity to the rule list. When the associated entity initiates a transaction again, an enhanced identity authentication strategy, such as facial recognition or SMS verification code, will be automatically triggered, and the amount of its single transaction will be limited.
[0059] Fuse rating (0.7≤ ): Immediately freeze the account of the associated entity, suspend all transaction permissions, and send an alert to the risk control operations platform.
[0060] Furthermore, the system acquires group activity data associated with the target entity and the related entities within a preset time period, and determines an activity stagnation index based on the group activity data. Based on the activity stagnation index, it determines whether the effect of the preset handling strategy meets preset conditions. If not, it adjusts the risk propagation weight and / or the decay constant corresponding to the time decay factor used in the risk propagation decay coefficient. Specifically, after the circuit breaker strategy is executed, the overall activity status of the related groups within time period T is continuously monitored. If the activities of the related groups significantly cease or show no significant change, it determines whether the circuit breaker is effective and automatically increases or decreases the weight of the relevant edge types. and time decay constant Specifically, the comprehensive indicator of gang activity stagnation is defined. For example, the result can be obtained by weighted summing the decline rates of three indicators: transaction volume, login events, and net cash flow. ; in, The rate of decline in group transaction volume. Group login frequency attenuation rate The group's net cash flow attenuation rate, of which 1.
[0061] Based on stagnation If the threshold is exceeded, adjust the parameters using a fixed-step strategy. To adjust the step size, the default is 0.02. Set upper and lower bounds (0.1 and 1.0) to prevent parameter drift from becoming uncontrollable. efficient: ; Invalid / Questionable: ; Time decay constant Adjustments: The larger the value, the faster the time decay (i.e., the more sensitive it is to time), and the adjustment direction is opposite to the weight.
[0062] efficient: ; Invalid / Questionable: ; in, The default value is 0.01. You can set the upper and lower bounds, such as 0.5 for the upper bound and 0.01 for the lower bound.
[0063] As described above, this embodiment of the application depicts the relationships between accounts based on a relational graph. After obtaining transaction system flow data, it uses an anti-fraud decision engine to perform rule matching on the flow data, thereby locating target entities in the graph that meet the fraud criteria. Subsequently, starting from the target entity, it expands along the graph in multiple hops, collecting attribute information of related entities and connecting edges. It then combines the inherent risk score of the target entity with the attributes of each edge to assign a risk propagation attenuation coefficient to each edge, thereby quantifying the infection risk of each related entity. Finally, a three-level circuit breaker mechanism is introduced to classify the infection risk within a range, identifying target related entities that need to be dealt with as part of a group fraud scheme. Simultaneously, it triggers system alarms and executes corresponding control strategies. This allows for rapid spread to the entire related network and automatic handling after the discovery of a single high-risk individual, significantly improving the efficiency of preventing group fraud. Furthermore, the three-level automatic circuit breaker strategy enables rapid isolation and crackdown before fraud groups can react.
[0064] As can be seen from the previous embodiment, this application discloses a risk control method for a transaction system based on a relational knowledge graph, which can automatically execute tiered blocking operations based on a single high-risk warning. Next, taking a third-party payment platform as an example, it will address issues such as... Figure 3 The risk control method for the transaction system based on the relational knowledge graph is explained in detail.
[0065] At 10:00:00 on a certain day, user account A initiated a transfer transaction of 48,000 yuan on the platform. The data acquisition probe captured the transaction details in real time and wrote them to the message queue.
[0066] First, the anti-fraud decision engine consumes the transaction record from the message queue, concatenates it with pre-calculated account profile features, and then performs rule matching. The matching results show that the transaction matches two preset rules: "account opened less than 72 hours ago and single transfer amount greater than 30,000 yuan" and "the recipient account is in the list of merchants involved in fraud." Based on this, the anti-fraud decision engine determines that account A meets the preset fraud conditions, identifies its corresponding entity in the association graph as the target entity, and assigns it a preset risk score of 1.0. This process takes approximately 30 milliseconds.
[0067] Secondly, the early warning and fission process was triggered. Starting with target entity A, the system performed a bidirectional breadth-first multi-hop query on the association graph after edge pruning. The pruning condition was set to retain only connections created within the last 90 days and with an interaction frequency greater than 2. The query yielded 12 one-hop related entities, 87 two-hop related entities, and 634 three-hop related entities, while also returning the edge attribute information for all connections along each association path. Since approximately 60% of the query requests hit the materialized view, the total time for this multi-hop query was approximately 18 milliseconds.
[0068] Then, the infection risk value is calculated for each of the 733 related entities. Taking related entity B as an example, entity B is directly connected to target entity A through a transfer relationship edge. The interaction frequency of this edge is 86 times, and the time interval between the most recent interaction and the current time is 2 hours. Based on this, the edge type weight is calculated to be 0.8, the frequency decay factor is approximately 0.64, the time decay factor is approximately 0.98, and the risk propagation decay coefficient is approximately 0.50. Therefore, the infection risk value of entity B is approximately 0.50. Taking related entity C as another example, entity C is connected to entity B through a shared device relationship edge. The interaction frequency of this edge is 41 times, and the time interval is 5 hours. The calculated risk propagation decay coefficient of this edge is approximately 0.33. Therefore, the infection risk value of entity C is approximately 0.50 multiplied by 0.33, which equals 0.17.
[0069] Next, the infection risk values were determined based on a three-tiered circuit breaker mechanism. The determination showed that 3 related entities had infection risk values falling within the third preset range of 0.7 or higher, and were therefore classified as circuit breaker level; 9 related entities had infection risk values falling within the second preset range of 0.5 or higher and less than 0.7, and were classified as challenge level; 41 related entities had infection risk values falling within the first preset range of 0.3 or higher and less than 0.5, and were classified as observation level; the remaining 680 related entities had infection risk values below 0.3, and no action was triggered.
[0070] Then, the executor executes corresponding actions in parallel based on the above judgment results: for the three circuit breaker-level entities, the account service is immediately invoked to freeze their accounts and issue payment gateway interception rules, while simultaneously pushing alarms to the risk control operation platform; for the nine challenge-level entities, they are added to the rule list database, configured to trigger facial recognition verification in their next transaction, and the single transaction limit is reduced to 5,000 yuan; for the 41 observation-level entities, they are silently added to the monitoring list database and their feature calculation frequency is increased. The total execution time for this step is approximately 240 milliseconds. In this way, from the capture of suspicious transactions of account A to the completion of the hierarchical handling of the entire related group, the total end-to-end time is approximately 300 milliseconds, achieving rapid blocking before the fraud group completes the transfer of funds.
[0071] After the action is completed, a 24-hour delay effect retrospective message is sent to the message queue. 24 hours later, the message is consumed, and the system queries the graph database to statistically analyze the transaction volume, login frequency, and net cash flow data of the aforementioned 733 related entities in the 24 hours before and after the action. The statistics show that the group's transaction volume decay rate is 0.91, the group's login frequency decay rate is 0.54, and the group's net cash flow decay rate is 0.88. Calculated using the weighted formula, the group's activity stagnation degree is approximately 0.82 (0.5 x 0.91, 0.2 x 0.54, 0.3 x 0.88). Since the stagnation degree of 0.82 is greater than the preset threshold of 0.6, the system determines that the action is valid. Accordingly, the system automatically increases the weight of the transfer relationship edge type involved in this action from 0.8 to 0.82, and the weight of the shared device relationship edge type from 0.6 to 0.62. Meanwhile, the time decay constant, having reached its lower bound, remains unchanged at 0.01. The above parameter adjustment results are written to the configuration center and sent to each computing node in real time, taking effect immediately in subsequent risk contagion calculations.
[0072] See Figure 4 As shown, this embodiment of the invention discloses a risk control device for a transaction system based on a relational knowledge graph, comprising: The entity determination module 11 is used to acquire transaction flow data in the target transaction system and use the anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules to determine the target entity in the association graph that meets the preset fraud conditions; the association graph is a graph that shows the association relationship of each account in the target transaction system, and the association graph includes entities representing accounts and association paths representing association relationships. The information acquisition module 12 is used to perform multi-hop queries on the association graph starting from the target entity to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path. The risk value determination module 13 is used to generate a risk propagation attenuation coefficient corresponding to each connecting edge on the associated path based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, and to determine the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path. The transaction alarm module 14 is used to determine whether the infection risk value is within a preset infection value range based on a preset three-level circuit breaker mechanism, determine the target associated entity that meets the conditions for gang fraud based on the determination result, execute the corresponding preset handling strategy on the target associated entity and issue an alarm in the target transaction system.
[0073] As shown above, this application depicts the relationships between accounts based on a relational graph. After obtaining transaction data from the transaction system, it uses an anti-fraud decision engine to perform rule matching on the transaction data, thereby locating target entities in the graph that meet the fraud criteria. Subsequently, starting from the target entity, it expands along the graph in multiple hops, collecting attribute information of related entities and connecting edges. It then combines the inherent risk score of the target entity with the attributes of each edge to assign a risk propagation attenuation coefficient to each edge, thereby quantifying the contagion risk of each related entity. Finally, a three-level circuit breaker mechanism is introduced to classify the contagion risk within a range, identifying target related entities that need to be dealt with as part of a group fraud scheme. Simultaneously, it triggers system alarms and executes corresponding control strategies. This allows for rapid spread to the entire related network and automatic handling after the discovery of a single high-risk individual, significantly improving the efficiency of preventing group fraud. Furthermore, the three-level automatic circuit breaker strategy enables rapid isolation and crackdown before fraud groups can react.
[0074] In some specific embodiments, the information acquisition module 12 may specifically include: An entity query unit is configured to simultaneously initiate a breadth-first query from the target entity and a preset high-frequency node in the association graph, stop the query when the query results on both sides intersect, and obtain the associated entity; and / or, before performing the multi-hop query, filter the connection edges in the association graph according to the creation time and / or interaction frequency of the connection edges to obtain a filtered association graph, and perform the multi-hop query on the filtered association graph to obtain the associated entity.
[0075] In some specific implementations, the edge attribute information includes the edge type, the frequency of interaction between the entities connected by the edge, and the time interval between the interaction between the entities connected by the edge; correspondingly, the risk value determination module 13 may specifically include: The weight determination unit is used to determine the corresponding risk propagation weight for each connecting edge on the associated path according to the edge type; A factor determination unit is used to determine a corresponding frequency decay factor based on the interaction frequency and a corresponding time decay factor based on the interaction time interval. The coefficient generation unit is used to generate the risk propagation attenuation coefficient corresponding to the connection edge based on the risk propagation weight, the frequency attenuation factor, and the time attenuation factor.
[0076] In some specific implementations, the factor determination unit may specifically include: The factor acquisition subunit is used to perform logarithmic function normalization on the interaction frequency to obtain the frequency attenuation factor. The factor generation subunit is used to calculate the exponential decay function of Newton's law of cooling for the interaction time interval, and generate the corresponding time decay factor.
[0077] In some specific implementations, the risk value determination module 13 may specifically include: The risk value acquisition unit is used to perform a multiplication operation on the preset risk score of the target entity and the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path, based on the hop count order of the associated path, to obtain the infection risk value corresponding to the associated entity.
[0078] In some specific implementations, the transaction alarm module 14 may specifically include: The first risk value determination unit is used to add the corresponding target associated entity to the monitoring list if the infection risk value is in the first preset range, thereby increasing the frequency of feature calculation of the target associated entity and increasing the dimension of collecting the behavioral trajectory of the target associated entity. The second risk value determination unit is used to add the corresponding target associated entity to the rule list library if the infection risk value is in the second preset range, and to trigger an enhanced identity authentication operation for the target associated entity when the target associated entity makes the next transaction, and to limit the single transaction amount of the target associated entity; wherein, the infection risk value corresponding to the second preset range is greater than the infection risk value corresponding to the first preset range. The third risk value determination unit is used to freeze the account of the target associated entity, suspend the transaction permissions of the target associated entity, and push alarm information to the risk control operation platform if the infection risk value is in the third preset range; wherein, the infection risk value corresponding to the third preset range is greater than the infection risk value corresponding to the second preset range.
[0079] In some specific implementations, the transaction system risk control device based on relational knowledge graphs may further include: The indicator determination unit is used to acquire group activity data associated with the target entity and the associated entity within a preset time period, and determine the activity stagnation index based on the group activity data; The indicator judgment unit is used to determine whether the effect of the preset handling strategy meets the preset conditions based on the activity stagnation index. A weight or constant adjustment unit is used to adjust the attenuation constant corresponding to the risk propagation weight and / or time attenuation factor used in the risk propagation attenuation coefficient if no.
[0080] Furthermore, embodiments of this application also disclose an electronic device, Figure 5This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the risk control method for a transaction system based on a relational knowledge graph disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be a computer.
[0081] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.
[0082] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0083] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the risk control method for a transaction system based on a relational knowledge graph, which is executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.
[0084] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned risk control method for a transaction system based on a relational knowledge graph. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.
[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.
[0086] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0087] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0088] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0089] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A risk control method for a transaction system based on a relational knowledge graph, characterized in that, include: Obtain transaction flow data from the target transaction system, and use an anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules in order to identify target entities in the association graph that meet preset fraud conditions. The association graph is a graph that shows the association relationship between various accounts in the target transaction system. The association graph includes entities representing accounts and association paths representing association relationships. Starting from the target entity, perform multi-hop queries on the association graph to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path; Based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, a risk propagation attenuation coefficient corresponding to each connecting edge on the associated path is generated. Based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path, the infection risk value corresponding to each associated entity is determined. Based on a preset three-level circuit breaker mechanism, it is determined whether the infection risk value is within a preset infection value range. Based on the determination result, the target associated entity that meets the conditions for gang fraud is identified, and the corresponding preset handling strategy is executed on the target associated entity and an alarm is triggered in the target transaction system.
2. The risk control method for a transaction system based on a relational knowledge graph according to claim 1, characterized in that, The step of performing multi-hop queries on the association graph, starting from the target entity, to obtain associated entities that have a path to the target entity includes: Breadth-first queries are initiated simultaneously from the target entity and preset high-frequency nodes in the association graph. The query stops when the query results on both sides intersect, and the associated entity is obtained. And / or, before performing the multi-hop query, the connection edges in the association graph are filtered according to the creation time and / or interaction frequency of the connection edges to obtain a filtered association graph, and the multi-hop query is performed on the filtered association graph to obtain the associated entity.
3. The risk control method for a transaction system based on a relational knowledge graph according to claim 1, characterized in that, The edge attribute information includes the edge type, the frequency of interaction between the entities connected by the edge, and the time interval between the interaction between the entities connected by the edge. Accordingly, generating risk propagation attenuation coefficients corresponding to each connecting edge on the associated path based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path includes: For each connecting edge on the associated path, the corresponding risk propagation weight is determined according to the edge type; The frequency decay factor is determined based on the interaction frequency, and the time decay factor is determined based on the interaction time interval. Based on the risk propagation weight, the frequency decay factor, and the time decay factor, the risk propagation decay coefficient corresponding to the connection edge is generated.
4. The risk control method for a transaction system based on a relational knowledge graph according to claim 3, characterized in that, The step of determining the corresponding frequency decay factor based on the interaction frequency and the corresponding time decay factor based on the interaction time interval includes: The frequency of interaction is normalized using a logarithmic function to obtain the frequency attenuation factor. The time decay factor is generated by calculating the exponential decay function of Newton's law of cooling for the interaction time interval.
5. The risk control method for a transaction system based on a relational knowledge graph according to claim 1, characterized in that, The step of determining the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connection edge on the associated path includes: The preset risk score of the target entity and the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path are multiplied together based on the hop count of the associated path to obtain the infection risk value corresponding to the associated entity.
6. The risk control method for a transaction system based on a relational knowledge graph according to claim 1, characterized in that, The process of determining whether the infection risk value falls within a preset infection value range based on a preset three-level circuit breaker mechanism, identifying target related entities that meet the conditions for group fraud based on the determination result, and executing corresponding preset handling strategies on the target related entities and issuing an alarm in the target transaction system includes: If the infection risk value is within the first preset range, the corresponding target associated entity will be added to the monitoring list database to increase the frequency of feature calculation for the target associated entity and increase the dimension of collecting the behavioral trajectory of the target associated entity. If the infection risk value is within the second preset range, the corresponding target associated entity is added to the rule list library. When the target associated entity makes its next transaction, an enhanced identity authentication operation is triggered for the target associated entity, and the single transaction amount of the target associated entity is limited. Wherein, the infection risk value corresponding to the second preset range is greater than the infection risk value corresponding to the first preset range. If the infection risk value is within the third preset range, the account of the target associated entity is frozen, the transaction permissions of the target associated entity are suspended, and an alarm message is pushed to the risk control operation platform; wherein, the infection risk value corresponding to the third preset range is greater than the infection risk value corresponding to the second preset range.
7. The risk control method for a transaction system based on a relational knowledge graph according to claim 1, characterized in that, After executing the corresponding preset handling strategy on the target associated entity and issuing an alarm in the target transaction system, the method further includes: Obtain group activity data associated with the target entity and the associated entity within a preset time period, and determine the activity stagnation index based on the group activity data; Based on the activity stagnation index, determine whether the effect of the preset handling strategy meets the preset conditions; If not, then adjust the attenuation constant corresponding to the risk propagation weight and / or time attenuation factor used in the risk propagation attenuation coefficient.
8. A risk control device for a transaction system based on a relational knowledge graph, characterized in that, include: The entity determination module is used to acquire transaction flow data of each transaction in the target transaction system, and use the anti-fraud decision engine to perform rule matching on each transaction flow data according to preset rules to determine the target entities in the association graph that meet the preset fraud conditions. The association graph is a graph that shows the association relationship between various accounts in the target transaction system. The association graph includes entities representing accounts and association paths representing association relationships. The information acquisition module is used to perform multi-hop queries on the association graph starting from the target entity to obtain the associated entities that have an association path with the target entity, as well as the edge attribute information of each connecting edge on the association path. The risk value determination module is used to generate a risk propagation attenuation coefficient corresponding to each connecting edge on the associated path based on the preset risk score of the target entity and the edge attribute information of each connecting edge on the associated path, and to determine the infection risk value corresponding to each associated entity based on the risk propagation attenuation coefficient corresponding to each connecting edge on the associated path. The transaction alarm module is used to determine whether the infection risk value is within a preset infection value range based on a preset three-level circuit breaker mechanism, identify the target associated entity that meets the conditions for gang fraud based on the determination result, execute the corresponding preset handling strategy on the target associated entity and issue an alarm in the target transaction system.
9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the risk control method for a transaction system based on a relational knowledge graph as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the risk control method for a transaction system based on a relational knowledge graph as described in any one of claims 1 to 7.
