Risk synergic control method, device, equipment and medium

Abstract

This invention relates to the field of artificial intelligence technology and discloses a risk collaborative control method, apparatus, device, and medium, comprising: acquiring risk feature vectors output by risk control analysis sub-models deployed locally from at least two data source nodes; aligning the risk feature vectors from the at least two data source nodes to construct a cross-node risk correlation graph; assigning feature weights to each risk feature vector based on its risk contribution in the cross-node risk correlation graph, and generating a fused feature vector by combining the risk feature vectors and feature weights; inputting the fused feature vector into a risk collaborative analysis model, and determining a target risk disposal strategy from a preset set of disposal strategies using a weighted voting mechanism; and executing corresponding business intervention operations according to the target risk disposal strategy. This invention can be applied to fintech scenarios, solving the technical problem of fragmented multi-source risk control data in existing risk control systems, which prevents institutions from coordinating risk control efforts.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a risk collaborative control method, apparatus, equipment and medium. Background Technology

[0002] Currently, existing financial risk control systems suffer from numerous technical deficiencies, resulting in ineffective risk control and prevention. On one hand, the industry generally suffers from data silos. Traditional risk control relies solely on internal business data from financial institutions for risk assessment, failing to integrate with external data sources such as business registration and taxation. This leads to a one-sided risk profile, an inability to construct a complete risk portrait of the entity, and insufficient accuracy in risk identification. On the other hand, existing risk control systems lack cross-institutional and cross-industry collaborative sharing capabilities. Each risk control entity operates independently, and risk information related to different institutions cannot be shared in a timely manner, making it difficult to predict cross-entity transmission risks. Therefore, when a single institution experiences business risk, the risk can easily spread rapidly, hindering coordinated early warning and response, and potentially triggering regional chain financial risks.

[0003] Therefore, there is an urgent need for a financial risk control technology solution that can solve the problem of fragmented multi-source risk control data and the inability of institutions to coordinate risk control, and make up for the shortcomings of existing system risk control. Summary of the Invention

[0004] This invention provides a risk collaborative control method, device, computer equipment, and medium based on artificial intelligence, to solve the technical problem that existing risk control systems suffer from fragmented multi-source risk control data, which prevents institutions from coordinating risk control efforts.

[0005] Firstly, a risk collaborative control method is provided, applied to a central node, wherein the central node is communicatively connected to at least two data source nodes, the method comprising: Obtain risk feature vectors output by the risk control analysis sub-model deployed locally on the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes; The risk feature vectors of the at least two data source nodes are aligned to construct a cross-node risk association graph, wherein the data source nodes are the graph nodes and the risk feature vectors are the node attributes of the graph nodes. Based on the risk contribution of each risk feature vector in the cross-node risk association graph, feature weights are assigned to each risk feature vector, and a fused feature vector is generated by combining the risk feature vectors and the feature weights. The fused feature vector is input into a pre-trained risk collaborative analysis model, and a weighted voting mechanism is used to determine the target risk management strategy from a preset set of management strategies. Execute corresponding business intervention operations based on the target risk management strategy.

[0006] Secondly, a risk collaborative control device is provided, applied to a central node, which communicates with at least two data source nodes, including: The feature acquisition module is used to acquire risk feature vectors output by the risk control analysis sub-model deployed locally by the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes; The graph construction module is used to perform feature alignment on the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph, wherein the cross-node risk association graph uses the data source nodes as graph nodes and the risk feature vectors as the node attributes of the graph nodes; The weight allocation module is used to allocate feature weights to each risk feature vector based on the risk contribution of each risk feature vector in the cross-node risk association graph, and to generate a fused feature vector by combining the risk feature vector and the feature weights. The strategy determination module is used to input the fused feature vector into a pre-trained risk collaborative analysis model and determine the target risk disposal strategy from a preset set of disposal strategies using a weighted voting mechanism. The strategy execution module is used to execute corresponding business intervention operations according to the target risk management strategy.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned risk collaborative control method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned risk collaborative control method.

[0009] The aforementioned risk collaborative control method, device, computer equipment, and storage medium can be applied to a central node that communicates with at least two data source nodes. The central node acquires risk feature vectors output by locally deployed risk control analysis sub-models from at least two data source nodes, without acquiring or storing the original business risk control data from each node. The local sub-models of the data source nodes perform pre-analysis of the data, and the original data is stored locally, effectively preventing the leakage of original business privacy data during cross-node transmission. By aligning the risk feature vectors of at least two data source nodes, a cross-node risk correlation graph is constructed. This graph can be used to clarify the relationships between various data source nodes, breaking down the fragmentation of multi-party risk control data and facilitating the identification of relationships between nodes. This approach addresses hidden risks and, by assigning feature weights to each risk feature vector based on its risk contribution in the cross-node risk correlation graph, generates a fused feature vector by combining the risk feature vectors and their weights. This allows for the allocation of corresponding feature weights based on the risk contribution of each vector, integrating multi-dimensional risk control information, and generating a unified feature vector to improve the accuracy of risk feature analysis. By inputting the fused feature vector into the risk collaborative analysis model and using a weighted voting mechanism to select the optimal handling strategy, it avoids errors in judgment based on a single risk control dimension and improves the credibility of risk assessment results. Furthermore, by executing corresponding business intervention operations based on the target risk handling strategy, it achieves multi-node linked risk control, promptly preventing the spread of risks across institutions and preventing regional chain financial risks. Attached Figure Description

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

[0011] Figure 1 This is a schematic diagram of an application environment for a risk collaborative control method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a risk collaborative control method in one embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 4 yes Figure 3 A flowchart illustrating a specific implementation of step S24; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 6This is a schematic diagram of a risk collaborative control device in one embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 8 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] 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, not all, of the embodiments of the present invention. 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.

[0013] The risk collaborative control method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can be configured as a central node, and the client can be configured as a data source node. The data source node has a risk control analysis sub-model deployed locally, and the server communicates with at least two clients (i.e., data source nodes). The server obtains risk feature vectors output by the locally deployed risk control analysis sub-models of at least two data source nodes. It does not acquire or store the original business risk control data of each node; the local sub-model of the data source node performs the pre-analysis of the data, and the original data is retained locally, effectively avoiding the leakage of original business privacy data during cross-node transmission. By aligning the risk feature vectors of at least two data source nodes, a cross-node risk correlation graph is constructed. This graph can be used to clarify the relationships between various data source nodes, breaking down the fragmentation of multi-party risk control data and helping to identify hidden risk correlations between nodes. Furthermore, by analyzing the risk feature vectors in the cross-node risk correlation graph... The system assigns feature weights to each risk feature vector based on its risk contribution, and combines these with the feature weights to generate a fused feature vector. This fused feature vector assigns corresponding feature weights based on the risk contribution of each vector, integrates multi-dimensional risk control information, and generates a unified feature vector, thus improving the accuracy of risk feature analysis. By inputting the fused feature vector into a risk collaborative analysis model, a weighted voting mechanism is used to select the optimal handling strategy, avoiding errors in judgment based on a single risk control dimension and improving the credibility of risk assessment results. Based on the target risk handling strategy, corresponding business intervention operations are executed, achieving multi-node linked risk control, timely preventing the spread of risk across institutions, and preventing regional chain financial risks. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 As shown, Figure 2 This is a flowchart illustrating a risk collaborative control method provided in an embodiment of the present invention. The method is applied to a central node, wherein the central node communicates with at least two data source nodes, and includes the following steps S10-S50: S10: Obtain the risk feature vectors output by the risk control analysis sub-model deployed locally by the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes.

[0015] The risk collaborative control method provided by this invention can be applied to intelligent risk collaborative control engines in various application scenarios. The intelligent risk collaborative control engine is implemented through a server-side component, which can output risk feature vectors in real time from at least two data source nodes via a locally deployed risk control analysis sub-model. The data source nodes can be distributed communication nodes deployed by different institutions such as banks, tax authorities, industrial and commercial bureaus, and IoT agencies, responsible for collecting business risk control data from their respective institutions. The risk control analysis sub-model is a dedicated model deployed locally on the data source nodes for processing business risk control data. Business risk control data can include raw business data collected by various institutions, such as bank transaction data, industrial and commercial registration data, tax declaration data, IoT device behavior data, and insurance claim data.

[0016] The central node retrieves risk feature vectors uploaded by each data source node after processing by its locally deployed risk control analysis sub-model from a public message queue. The public message queue serves as an intermediate communication carrier connecting the various data source nodes, facilitating the exchange of vector data. Data source nodes can unidirectionally upload risk feature vectors to the public message queue; each risk feature vector is a standardized risk control representation vector output by the data source node after processing business risk control data. In this embodiment, the central node does not retrieve the original business risk control data from any data source node to avoid cross-node transmission leakage of original business privacy data.

[0017] In some embodiments, the data source nodes and the central node interact and transmit data using a streaming processing mode. Each data source node collects financial business risk control data in real time, processes it through its local risk control analysis sub-model to generate risk feature vectors, and then continuously pushes them to the central node in a streaming, sharded format. The central node establishes a dedicated streaming receiving channel to receive sharded data in real time and perform splicing and verification. It only transmits the processed feature vectors and not the original business risk control data. Through the streaming bidirectional transmission architecture, low-latency real-time interaction between business data and risk feature vectors can be achieved, adapting to high-frequency real-time risk control scenarios in the financial sector. Through the streaming incremental transmission mechanism, newly added business risk control information can be updated in real time, avoiding data lag and connection loss issues caused by batch transmission. This ensures that the central node obtains risk feature vectors with complete time series, providing real-time and complete data support for subsequent full-process risk assessment.

[0018] S20: Perform feature alignment on the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph, wherein the cross-node risk association graph uses the data source nodes as graph nodes and the risk feature vectors as node attributes of the graph nodes.

[0019] After obtaining risk feature vectors from at least two data source nodes, feature alignment is performed on the multiple risk feature vectors. Using the data source nodes as graph nodes and the risk feature vectors as node attributes of the corresponding graph nodes, a cross-node risk association graph is constructed. Feature alignment refers to the process of eliminating heterogeneous differences between vectors output by different data source nodes; the cross-node risk association graph is a topological graph representing the association relationships and risk control attributes between data source nodes.

[0020] Furthermore, such as Figure 3 As shown, step S20 above, which involves feature alignment of the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph, includes the following steps: S21: Obtain a pre-trained feature alignment mapping subnet, which is trained using the risk collaborative analysis model as the teacher model and historical risk feature samples through knowledge distillation. S22: Classify and aggregate the risk feature vectors of the at least two data source nodes according to the node identifier; S23: Input the aggregated risk feature vectors into the feature alignment mapping subnet in sequence, and perform dimension mapping and distribution normalization on the aggregated risk feature vectors according to the standard feature rules learned by the teacher model to obtain a structure-aligned risk feature vector group. S24: Perform graph association operation on the risk feature vector group to obtain vector attribute association parameters, wherein the vector attribute association parameters include node attribute parameters and association edge parameters; S25: Perform graph encapsulation processing based on the node attribute parameters and the associated edge parameters to obtain a cross-node risk association graph.

[0021] For steps S21-S25, firstly, a pre-stored feature alignment mapping subnet is obtained. This feature alignment mapping subnet uses the risk collaborative analysis model as the teacher model. The central node pre-acquires historical risk feature samples collected in the database. Iterative training of the subnet is completed based on the knowledge distillation mode, replicating the standard feature rules learned by the teacher model. After training, it is stored locally for subsequent vector alignment.

[0022] After obtaining the feature alignment mapping subnet, the risk feature vectors from multiple parties are classified and aggregated. Specifically, risk feature vectors from at least two data source nodes are obtained, the node identifier of the data source node bound to each vector is read, and the vectors are grouped and classified according to the node identifier. Risk feature vectors corresponding to the same data source node are grouped together to distinguish vector data produced by different institutions and avoid cross-node vector mixing.

[0023] The grouped risk feature vectors are input one by one into the feature alignment mapping subnet. According to the standard feature rules learned by the teacher model, the grouped risk feature vectors are subjected to dimension adaptation mapping and data distribution normalization to eliminate the dimensional differences and numerical distribution differences of the output vectors of the data source nodes of different institutions. Finally, the structure-aligned risk feature vector group is output.

[0024] After obtaining the structure-aligned risk feature vector group, the built-in graph operation rules are used to perform graph association operations on the structure-aligned risk feature vector group in batches, extract the node attribute parameters of each data source node, calculate the association edge parameters between each node, and form vector attribute association parameters based on the node attribute parameters and association edge parameters.

[0025] After obtaining the vector attribute association parameters, the graph is encapsulated based on the obtained node attribute parameters and association edge parameters. The data source node is used as the graph node, and the corresponding node attribute parameters are bound. The graph nodes with association relationships are connected by the association edge parameters to construct a cross-node risk association graph.

[0026] As can be seen from the above scheme, the feature alignment mapping subnet uses the risk collaborative analysis model as the teacher model and completes knowledge distillation training using historical risk feature samples. This enables it to match backend risk control judgment standards, unify vector feature rules, and improve vector alignment adaptability. By aggregating vectors from various data sources through node identifiers, it distinguishes between different types of business data such as banking, industry and commerce, taxation, and IoT, avoiding the problem of mixed data operations. Simultaneously, it performs dimensional mapping and distribution normalization on the aggregated vectors, eliminating differences in vector dimensions and values ​​across different institutions and addressing the pain point of difficult collaborative modeling of heterogeneous data across institutions. Finally, by combining node attribute parameters and associated edge parameters to encapsulate the graph, it intuitively reflects the business and risk relationships of each node, providing stable data support for subsequent vector weighting and risk scoring judgment, and improving the overall accuracy of risk control judgment.

[0027] Further, in step S23 above, the collected risk feature vectors are sequentially input into the feature alignment mapping subnet, and the collected risk feature vectors are subjected to dimension mapping and distribution normalization according to the standard feature rules learned by the teacher model to obtain a structure-aligned risk feature vector group, including the following steps: S231: Sort and organize the collected risk feature vectors to obtain an ordered set of risk feature vectors; S232: Sequentially extract a single set of risk feature vectors from the ordered set of risk feature vectors and perform admission verification processing according to the verification rules of the feature alignment mapping subnet to obtain the risk feature vector to be aligned; S233: Obtain the standard feature rules learned by the teacher model, and perform dimensional mapping on the risk feature vector to be aligned according to the standard feature rules to obtain a risk feature vector with unified dimensions; S234: Perform distribution normalization on the risk feature vector with the unified dimension to eliminate differences in numerical distribution and obtain a normalized risk feature vector. S235: Summarize and integrate all normalized risk feature vectors to obtain a structure-aligned risk feature vector group.

[0028] It should be noted that the feature alignment mapping subnet in this embodiment is a dedicated processing subnet trained using the risk collaborative analysis model as the teacher model. It internally carries admission verification rules and standard feature rules. The admission verification rules are the criteria used to determine whether a vector can be input into the feature alignment mapping subnet for computational processing. The standard feature rules are unified feature specifications generated by training the teacher model, used to standardize the dimension and numerical range of the output vectors from different data sources, smoothing out feature differences between data sources.

[0029] For steps S231-S235, the collected risk feature vectors are first sorted and organized to obtain an ordered set of risk feature vectors. The sorting process can arrange the risk feature vectors according to data source sequence, business time sequence, etc., to generate an ordered set of risk feature vectors, thus adapting to the sequential processing logic of the feature alignment mapping subnet.

[0030] Single risk feature vectors are sequentially extracted from the ordered set of risk feature vectors. Combined with the built-in access verification rules of the feature alignment mapping subnet, the vector format and field integrity are checked, invalid risk feature vectors are removed, and the risk feature vectors to be aligned are obtained.

[0031] Obtain the standard feature rules learned by the teacher model, and perform dimension mapping on the risk feature vector to be aligned obtained from the verification according to the standard feature rules. That is, correct the dimension parameters of the risk feature vector to be aligned, unify the vector dimension specifications, eliminate the dimension differences of heterogeneous vectors from various data sources, and obtain a risk feature vector with unified dimensions.

[0032] The risk feature vector after unifying the dimensions is subjected to data distribution normalization operation to unify the standard of vector numerical distribution, eliminate the vector numerical deviation caused by different data sources, and obtain the normalized risk feature vector.

[0033] Repeat steps S231 to S234 to complete the verification and alignment of all vectors in the risk feature vector set, integrate all normalized risk feature vectors, and complete the bidirectional alignment of vector dimension and data distribution to obtain a structurally aligned risk feature vector group.

[0034] As can be seen from the above scheme, the entire process can regulate the vector input sequence and ensure smooth serial processing of subnets by sorting and organizing the collected risk feature vectors; by performing admission verification on each vector, incomplete and invalid vectors can be screened out, improving the accuracy of subsequent feature processing; by calling the standard feature rules learned by the teacher model to complete the vector dimension mapping, the vector dimension specifications across data sources can be unified, and the problem of original dimension heterogeneity can be resolved; by performing distribution normalization on the dimension-unified vectors, the numerical distribution deviation of multiple data sources can be smoothed out; by integrating the fully normalized vectors to form aligned vector groups, vector global standardization can be achieved, ensuring the accuracy and uniformity of risk assessment results across institutions.

[0035] Furthermore, such as Figure 4 As shown, step S24 above, which involves performing graph association operations on the risk feature vector group to obtain vector attribute association parameters, includes the following steps: S241: Obtain the preset subject matching rules, perform subject attribution matching on all vectors in the risk feature vector group, divide the vector set corresponding to the same business subject, and obtain the subject grouping result; S242: Traverse any two risk feature vectors in the main grouping results to form a pairing vector, and use a preset similarity measurement function to perform similarity calculation on the pairing vector to obtain the initial association score; S243: The initial association score is filtered using a preset association threshold. First pairing vectors that are lower than the association threshold are removed, and second pairing vectors that are higher than the association threshold and their corresponding initial association scores are retained. S244: Construct the associated edges of the second pairing vector and map the initial associated scores of the second pairing vector to the associated edge parameters of the second pairing vector; S245: Map the self-attribute features of all vectors in the risk feature vector group to node attribute parameters, and integrate the node attribute parameters with the associated edge parameters to obtain vector attribute association parameters.

[0036] For steps S241-S245, firstly, a preset subject matching rule is obtained. This subject matching rule refers to the matching standard used to determine whether different data source vectors belong to the same business subject. Business subjects can include risk control objects such as bank account opening companies, tax filing entities, business registration companies, and IoT control objects. Using this subject matching rule, subject attribution matching is performed on all risk feature vectors within the structured aligned risk feature vector group. The business subject corresponding to each vector is identified, and all risk feature vectors belonging to the same business subject are integrated and aggregated to complete the vector subject division, resulting in subject grouping results.

[0037] After obtaining the main grouping results, the paired vector calculation is performed. Specifically, any two risk feature vectors in the main grouping results are traversed, and paired vectors are formed. Using a preset similarity metric function, the association similarity of each paired vector is calculated one by one, and the initial association score corresponding to each paired vector is quantified. This initial association score is a quantitative score used to characterize the strength of the risk association between the two paired vectors.

[0038] After obtaining the initial association scores of each pair of vectors, a preset association threshold is obtained. The initial association scores are then filtered using this threshold. The values ​​of each pair of vectors are compared with the association threshold. The first pair of vectors with scores lower than the association threshold are removed, while the second pair of vectors with scores higher than the association threshold are retained. At the same time, the initial association scores corresponding to the second pair of vectors are also retained.

[0039] Next, node association edges are constructed for each group of second pairing vectors. The initial association scores corresponding to the second pairing vectors are mapped and assigned to the edge attribute data of the corresponding association edges, thereby obtaining the association edge parameters corresponding to each group of second pairing vectors.

[0040] Finally, the central node extracts the vector attribute features inherent in each risk feature vector within the risk feature vector group, maps the vector attribute features one by one to the node attribute parameters of the corresponding data source node, and integrates all node attribute parameters with the generated associated edge parameters to obtain the vector attribute association parameters.

[0041] As can be seen from the above scheme, the entire process uses preset subject matching rules to match and group vector subjects, which can accurately collect cross-node risk vectors of the same business subject, avoid cross-contamination of data from different subjects, and improve the targeting of association operations. By using a similarity metric function to solve the association score of paired vectors, the risk association degree of nodes can be digitally determined, avoiding subjective errors caused by manual judgment. By using an association threshold to filter paired vectors, invalid weak association vectors can be eliminated, reducing redundant operations and reducing the computational cost of graph modeling. By mapping vector attributes and the association score of paired vectors, node attribute parameters and association edge parameters are generated and integrated, which can completely characterize the attributes of data source nodes and the risk association relationships between nodes, ensuring the accuracy of graph construction.

[0042] S30: Based on the risk contribution of each risk feature vector in the cross-node risk association graph, assign feature weights to each risk feature vector, and combine the risk feature vectors and the feature weights to generate a fused feature vector.

[0043] After constructing the cross-node risk correlation graph, the vector parameters stored in the cross-node risk correlation graph are retrieved, and the risk contribution of each risk feature vector is determined one by one to distinguish between high-value and low-value risk vectors. Based on the contribution, appropriate feature weights are assigned to each risk feature vector to reduce the interference of low-value vectors on the risk control results. Then, the weighted risk feature vectors are standardized and formatted, multi-dimensional risk control features are integrated, and invalid and redundant features are eliminated to obtain a fused feature vector.

[0044] Furthermore, such as Figure 5 As shown, step S30 above, which involves allocating feature weights to each risk feature vector based on its risk contribution in the cross-node risk association graph, and generating a fused feature vector by combining the risk feature vectors and the feature weights, includes the following steps: S31: Extract business time-series attribute parameters from the risk feature vector of the cross-node risk association graph. The business time-series attribute parameters are used to characterize the business occurrence time of the business risk control data. S32: Obtain a preset time decay function, perform time decay calculation processing based on the time decay function and the business time sequence attribute parameters to obtain a risk decay factor, and assign the risk decay factor to the risk contribution degree of the risk feature vector in the time dimension. S33: Perform global weight reduction processing on the risk feature vector according to the risk attenuation factor to obtain the weighted risk feature vector. The longer the business occurred, the larger the corresponding risk attenuation factor and the greater the reduction in vector weight. S34: The reduced-weight risk feature vector is normalized using a normalization function to obtain a normalized risk feature vector; S35: Perform feature dimensionality reduction processing on the normalized risk feature vector to obtain the fused feature vector.

[0045] For steps S31-S35, business time-series attribute parameters are first extracted from the risk feature vectors contained in the cross-node risk association graph. These parameters represent the time of occurrence of various types of original business risk control data. Each risk feature vector within the graph is matched with a corresponding business time-series attribute parameter, ensuring a one-to-one correspondence between business time traceability.

[0046] After extracting the business time-series attribute parameters, a preset time decay function is obtained. This time decay function is a computational function that quantifies the impact of the business occurrence time on risk assessment. Using this time decay function, combined with the business time-series attribute parameters bound to a single risk feature vector, time-series decay calculations are performed to obtain the risk decay factor corresponding to each risk feature vector. The obtained risk decay factor is then assigned as the risk contribution degree of the corresponding risk feature vector in the time dimension. This risk contribution degree is used to measure the reference value of the risk feature vector in risk control assessment.

[0047] After obtaining the risk decay factors corresponding to each risk feature vector, a global weight reduction process is performed on the risk feature vectors within the cross-node risk correlation graph based on all risk decay factors, resulting in weighted risk feature vectors. Specifically, the risk decay factor value is related to the timeliness of the business; the older the business occurred, the larger the corresponding risk decay factor value, and the greater the reduction in the weight of the corresponding risk feature vector; conversely, the closer the business occurred to the current risk control assessment period, the smaller the corresponding risk decay factor value, and the smaller the reduction in the weight of the corresponding risk feature vector. Through time-series differentiated weight reduction, the assessment weight of outdated and old business risk control data is weakened.

[0048] After completing the global vector weight reduction process, a preset normalization function is obtained. Using this normalization function, the numerical normalization process is performed on all the risk feature vectors that have completed the weight reduction process. The numerical magnitude of the weighted vectors from different institutions is uniformly calibrated to obtain the normalized risk feature vectors, which facilitates subsequent unified calculation and processing.

[0049] Finally, the normalized risk feature vector is subjected to feature dimensionality reduction processing to remove redundant and invalid risk control features within the vector, retain the core judgment features of related business risks, compress the vector dimension, integrate effective risk control features from multiple institutions and dimensions, and finally generate a fusion feature vector that is adapted to the input of the risk collaborative analysis model.

[0050] As can be seen from the above scheme, the entire process, by extracting the business time-series attribute parameters bound to the risk feature vector, can accurately trace the occurrence time of business risk control data, providing data support for subsequent timeliness weight determination. By combining the time decay function with the time-series parameters to calculate the risk decay factor, the time dimension risk contribution of business vectors in different time periods can be quantified, aligning with the logic of financial risk control timeliness assessment. Global vector weight reduction using the risk decay factor weakens the influence of outdated business data on risk control and strengthens the proportion of real-time business risk assessment. Through normalization and feature dimensionality reduction in a step-by-step processing of vectors, the numerical specifications of vectors across the entire domain can be unified, redundant vector features can be eliminated, data volume can be compressed, and high-quality fused feature vectors can be generated, improving the accuracy and computational efficiency of backend model risk scoring and disposal strategy determination.

[0051] Furthermore, step S33 above, namely, performing global weight reduction processing on the risk feature vector according to the risk attenuation factor to obtain the weighted risk feature vector, includes the following steps: S331: Obtain the features and weight parameters of each dimension within the risk feature vector; S332: Synchronously superimpose the risk attenuation factor onto the weight parameters of each dimension feature, and perform a global scaling operation on the weight parameters to obtain the adjusted weight parameters; S333: The risk feature vector after weighting is obtained by weighting and summing the features of each dimension and the adjusted weight parameters.

[0052] For steps S331-S333, firstly, obtain the features of each dimension and the corresponding weight parameters contained in the risk feature vector. The dimension features are the individual risk control features after being split, including basic institutional risk control features such as transaction features, equipment features, and geographical location features. The weight parameters are the initial proportion values ​​of the dimension features, which are used to characterize the basic risk control influence of the single-dimensional features.

[0053] After obtaining the features and weight parameters of each dimension of the risk feature vector, the determined risk attenuation factor is synchronously superimposed on the weight parameter corresponding to each dimension feature. Combined with the risk attenuation factor, a global scaling operation is performed on the weight parameter to achieve dynamic weight correction, resulting in the adjusted weight parameter. Specifically, the older the business event, the larger the risk attenuation factor, and the greater the reduction in weight parameter after global scaling; conversely, the more recent the business event, the smaller the risk attenuation factor, and the smaller the reduction in weight parameter.

[0054] After obtaining the adjusted weight parameters for each dimension of features, a weighted summation operation is performed based on the features of each dimension and the adjusted weight parameters to obtain the reduced-weight risk feature vector. Specifically, each dimension feature within a single risk feature vector is multiplied pairwise by the adjusted weight parameters to obtain a single-item weighted feature value. Then, all single-item weighted feature values ​​are summed and normalized into a standardized vector, ultimately yielding the reduced-weight risk feature vector.

[0055] For example, the risk feature vector after weight reduction for each dimension can be calculated using: weights = softmax a·(λ1·transaction feature + λ2·device feature + λ3·geographical location), where a is the risk attenuation factor and λ is the weight parameter corresponding to each dimension feature. For instance, if a bank's historical transaction vector from six months ago has a relatively large risk attenuation factor a1, the corresponding transaction feature λ1 coefficient is scaled down by a1. Conversely, if the real-time transaction vector's risk attenuation factor a2 is relatively small, the corresponding λ1 coefficient is scaled down only slightly, achieving differentiated weight scaling.

[0056] As can be seen from the above scheme, the entire process, by extracting the features of each dimension within the vector and the initial weight parameters, can decompose various risk control features and achieve refined weight management. By scaling the dimension weights with risk decay factors, the feature weights can be adjusted according to the differences in business timeliness, conforming to the time-series risk control judgment rules. By weighted summing of each dimension feature with the adjusted weights, vector weight correction can be completed, and time-series weight reduction can be completed in an orderly manner.

[0057] S40: Input the fused feature vector into the pre-trained risk collaborative analysis model, and use a weighted voting mechanism to determine the target risk disposal strategy from the preset disposal strategy set.

[0058] It should be noted that the weighted voting mechanism refers to the voting rules that combine multi-dimensional risk control scores to collaboratively determine risk control and disposal solutions.

[0059] After obtaining the fused feature vector, a pre-trained risk collaborative analysis model is used, and the fused feature vector is input into the risk collaborative analysis model. The model conducts risk assessment from multiple dimensions, such as institutional association, business compliance, and transaction security, and outputs risk scores for each group. Through a weighted voting mechanism, combined with the preset weights corresponding to the risk scores of each dimension, the comprehensive risk level is calculated. Then, based on the comprehensive risk level, the target risk disposal strategy is determined from the preset disposal strategy set.

[0060] Furthermore, in step S40 above, the risk collaborative analysis model internally deploys risk decision nodes with at least two dimensions. The aforementioned input of the fused feature vector into the pre-trained risk collaborative analysis model, and the use of a weighted voting mechanism to determine the target risk management strategy from a pre-set set of management strategies, includes the following steps: S41: Input the fused feature vector into the risk decision nodes of the at least two dimensions respectively, and perform risk assessment through the risk decision nodes of the at least two dimensions to obtain at least two risk scores; S42: Using a preset weighted voting rule, calculate the weighted average of at least two risk scores to determine the comprehensive risk score; S43: Obtain a preset risk threshold range and determine whether the comprehensive risk score is within the risk threshold range; S44: If the comprehensive risk score is within the risk threshold range, then the risk warning intervention method is determined to be the target risk management strategy; S45: If the comprehensive risk score is not within the risk threshold range, then the risk-free intervention method is determined as the target risk management strategy.

[0061] It should be noted that the risk collaborative analysis model has at least two dimensions of risk decision nodes in its internal structure. These risk decision nodes are independent decision units within the model and can include various distributed decision nodes such as institutional compliance, transaction security, anti-fraud, credit assessment, and post-loan management. Each type of node independently evaluates and scores, without interfering with each other's evaluation logic.

[0062] For steps S41-S45, after obtaining the fused feature vector, the fused feature vector is input into risk decision nodes of at least two dimensions. Risk assessment is performed through each dimension's risk decision node to obtain at least two risk scores. Specifically, the fused feature vector is synchronously distributed to various risk decision nodes. The institutional compliance decision node combines vector data from industrial and commercial and tax data source nodes to determine cross-institutional cooperation compliance risks and outputs a compliance risk score. The transaction security decision node combines vector data from bank and IoT data source nodes to determine fund transfer and device access security risks and outputs a transaction security score. The remaining distributed decision nodes synchronously output corresponding specific risk scores, with each decision node outputting a set of exclusive risk scores.

[0063] After obtaining multiple risk scores, a pre-defined weighted voting rule is used to calculate the weighted average of each risk score to determine the overall risk score. This weighted voting rule is adapted to multiple decision-making nodes, allowing for adjustments to the weighting of each node's score based on the business type. All risk scores are then weighted and calculated to arrive at a comprehensive risk score.

[0064] After obtaining the comprehensive risk score, a pre-stored risk threshold range is retrieved. This risk threshold range is a critical score range used to classify whether a business has risk control implications, and it is determined by historical risk control business data. Scores within the range are considered to indicate the presence of business-related risks, while scores outside the range are considered to indicate the absence of such risks. The central node compares the comprehensive risk score with the risk threshold range to determine whether the score falls within it.

[0065] If the overall risk score is determined to be within the risk threshold range, it indicates that there is a potential risk of cross-node business linkage. In this case, the risk warning intervention method is retrieved and determined as the target risk handling strategy. If the overall risk score is determined not to be within the risk threshold range, it indicates that the cross-node business is compliant and risk-free, and there is no need to issue control instructions. In this case, the risk-free intervention method is determined as the target risk handling strategy.

[0066] As can be seen from the above scheme, the entire process involves risk assessment at multiple risk decision-making nodes, enabling the examination of business risks from multiple perspectives, including compliance, transactions, credit, and anti-fraud, thus overcoming the problem of biased judgments based on a single assessment dimension. By calculating the risk scores of each group through weighted voting rules, the weight of each node can be allocated according to the business type, resulting in an objective and reasonable comprehensive risk score that ensures the score results align with actual business operations. By comparing the comprehensive risk score with preset risk threshold ranges, a unified judgment standard can be established, quickly distinguishing between risky and compliant businesses. By selecting appropriate handling strategies based on the score ranges, precise intervention in risky businesses and exemption from intervention in compliant businesses can be achieved, enabling tiered risk control management, reducing unnecessary intervention operations, and improving the efficiency of risk handling in financial scenarios.

[0067] S50: Execute the corresponding business intervention operation according to the target risk management strategy.

[0068] After determining the target risk management strategy, corresponding business intervention operations are executed according to the strategy. For example, based on the target risk management strategy, the corresponding data source node of the risky business is located, a control instruction containing the target risk management strategy is encapsulated, and sent to the data source node through the communication link; the central node receives the execution receipt from the data source node and supervises the data source node to complete the corresponding intervention actions such as business release, risk review, and business control in accordance with the instruction requirements.

[0069] For example, in cross-entity risk control scenarios in the financial sector, the central node can be a regional financial risk control platform or a government-finance collaborative management platform, which serves as a unified management platform. The central node communicates and connects with three types of data source nodes: banks, industrial and commercial bureaus, and tax bureaus, to implement collaborative management of corporate credit risk. The specific application process is as follows: The central node acquires risk feature vectors uploaded by three types of data source nodes. Each data source node, relying on its local risk control analysis sub-model, processes risk control data for corporate credit transactions, entity registration, and tax-related invoicing, and outputs corresponding risk feature vectors for transaction risk control, entity compliance, and tax compliance. Each data source node does not upload the original business data of the enterprise.

[0070] The central node aligns the risk feature vectors of each data source node with the same vector format. It uses three types of institutions as graph nodes and their corresponding risk feature vectors as node attributes. Combined with the business relationships of enterprises, it constructs a cross-node risk association graph.

[0071] The central node combines business time-series attribute parameters and time decay function to determine the risk contribution of each vector. Business vectors from older time periods have lower contributions, while real-time business vectors have higher contributions. Feature weights are assigned based on the contribution, and the process generates enterprise credit fusion feature vectors.

[0072] The central node will integrate feature vectors into the risk collaborative analysis model. The model will then conduct risk scoring for multi-dimensional decision-making nodes, including anti-fraud, credit assessment, and post-loan management. A comprehensive risk score will be obtained through weighted voting rules, and the risk threshold range will be compared to determine the three target disposal strategies: normal loan disbursement, credit limit control, and qualification review.

[0073] The central node issues intervention instructions, which are then coordinated with the bank's data source nodes to execute corresponding credit business intervention operations, thus completing the risk control loop for this corporate financial business.

[0074] As can be seen, in the above scheme, the risk feature vectors output by the risk control analysis sub-models of at least two data source nodes are obtained through the central node. The original business risk control data of each node is not acquired or stored; the local sub-models of the data source nodes perform the pre-analysis of the data, and the original data is retained locally, effectively avoiding the leakage of original business privacy data during cross-node transmission. By aligning the risk feature vectors of at least two data source nodes, a cross-node risk correlation graph is constructed. This graph can be used to sort out the relationships between various data source nodes, breaking down the problem of fragmented risk control data from multiple parties, and helping to identify hidden risk correlations between nodes. Furthermore, by analyzing the risk feature vectors of at least two data source nodes through feature alignment, a cross-node risk correlation graph can be constructed. The risk contribution of risk feature vectors is assessed, and feature weights are assigned to each risk feature vector. By combining the risk feature vectors and feature weights, a fused feature vector is generated. This fused feature vector can be generated by assigning corresponding feature weights based on the risk contribution of each vector, integrating multi-dimensional risk control information, and improving the accuracy of risk feature assessment. By inputting the fused feature vector into the risk collaborative analysis model, the optimal handling strategy is selected through a weighted voting mechanism, avoiding errors in judgment based on a single risk control dimension and improving the credibility of risk assessment results. Based on the target risk handling strategy, corresponding business intervention operations are executed, realizing multi-node linkage risk control, timely blocking the spread of risks across institutions, and preventing regional chain financial risks.

[0075] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0076] In one embodiment, a risk collaborative control device is provided, which corresponds one-to-one with the risk collaborative control method described in the above embodiments. For example... Figure 6 As shown, the risk collaborative control device includes a feature acquisition module 101, a map construction module 102, a weight allocation module 103, a strategy determination module 104, and a strategy execution module 105. Detailed descriptions of each functional module are as follows: The feature acquisition module 101 is used to acquire risk feature vectors output by the risk control analysis sub-model deployed locally by the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes; The graph construction module 102 is used to perform feature alignment on the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph, wherein the cross-node risk association graph uses the data source nodes as graph nodes and the risk feature vectors as node attributes of the graph nodes. The weight allocation module 103 is used to allocate feature weights to each risk feature vector according to the risk contribution of each risk feature vector in the cross-node risk association graph, and generate a fused feature vector by combining the risk feature vector and the feature weights. The strategy determination module 104 is used to input the fused feature vector into a pre-trained risk collaborative analysis model and determine the target risk disposal strategy from a preset set of disposal strategies using a weighted voting mechanism. The strategy execution module 105 is used to execute corresponding business intervention operations according to the target risk management strategy.

[0077] In one embodiment, the map construction module 102 is specifically used for: The network acquisition submodule is used to acquire a pre-trained feature alignment mapping subnetwork, which is trained using the risk collaborative analysis model as the teacher model and historical risk feature samples through knowledge distillation. The vector aggregation submodule is used to classify and aggregate the risk feature vectors of the at least two data source nodes according to the node identifier; The vector alignment submodule is used to input the aggregated risk feature vectors into the feature alignment mapping subnet in sequence, and perform dimension mapping and distribution normalization on the aggregated risk feature vectors according to the standard feature rules learned by the teacher model to obtain a structure-aligned risk feature vector group. The parameter determination submodule is used to perform graph association operation processing on the risk feature vector group to obtain vector attribute association parameters, wherein the vector attribute association parameters include node attribute parameters and association edge parameters; The graph encapsulation submodule is used to encapsulate the graph based on the node attribute parameters and the associated edge parameters to obtain a cross-node risk association graph.

[0078] In one embodiment, the vector alignment submodule is specifically used for: The collected risk feature vectors are sorted and organized to obtain an ordered set of risk feature vectors. Sequentially extract a single set of risk feature vectors from the ordered set of risk feature vectors and perform admission verification processing according to the verification rules of the feature alignment mapping subnet to obtain the risk feature vector to be aligned; Obtain the standard feature rules learned by the teacher model, and perform dimensional mapping on the risk feature vector to be aligned according to the standard feature rules to obtain a risk feature vector with unified dimensions. The risk feature vector with the unified dimension is subjected to distribution normalization to eliminate the difference in numerical distribution and obtain the normalized risk feature vector. By summarizing and integrating all normalized risk feature vectors, a structure-aligned risk feature vector group is obtained.

[0079] In one embodiment, the parameter determination submodule is specifically used for: Obtain the preset subject matching rules, perform subject attribution matching on all vectors in the risk feature vector group, divide the vector set corresponding to the same business subject, and obtain the subject grouping result; Traverse any two risk feature vectors in the main grouping results to form a pairing vector, and use a preset similarity measurement function to solve the similarity of the pairing vectors to obtain the initial association score; The initial association scores are filtered using a preset association threshold. First pairing vectors that are lower than the association threshold are removed, while second pairing vectors that are higher than the association threshold and their corresponding initial association scores are retained. Construct the associated edges of the second pairing vector, and map the initial associated scores of the second pairing vector to the associated edge parameters of the second pairing vector; The self-attribute features of all vectors in the risk feature vector group are mapped to node attribute parameters, and the node attribute parameters are integrated with the associated edge parameters to obtain vector attribute association parameters.

[0080] In one embodiment, the weight allocation module 103 is specifically used for: From the risk feature vector of the cross-node risk association graph, business time series attribute parameters are extracted. These business time series attribute parameters are used to characterize the business occurrence time of the business risk control data. Obtain a preset time decay function, perform time decay calculation based on the time decay function and the business time sequence attribute parameters to obtain a risk decay factor, and assign the risk decay factor to the risk contribution of the risk feature vector in the time dimension. The risk feature vector is globally reduced in weight based on the risk attenuation factor to obtain the reduced-weight risk feature vector. The longer the business occurred, the larger the corresponding risk attenuation factor and the greater the reduction in vector weight. The reduced-weight risk feature vector is normalized using a normalization function to obtain a normalized risk feature vector. The normalized risk feature vector is subjected to feature dimensionality reduction processing to obtain the fused feature vector.

[0081] In one embodiment, the weight allocation module 103 is further configured to: Obtain the features and weight parameters of each dimension within the risk feature vector; The risk attenuation factor is synchronously superimposed on the weight parameters of each dimension feature, and the weight parameters are subjected to a global scaling operation to obtain the adjusted weight parameters. The reduced-weight risk feature vector is obtained by weighting and summing the features of each dimension and the adjusted weight parameters.

[0082] In one embodiment, the aforementioned risk collaborative analysis model internally deploys risk decision nodes with at least two dimensions, and the strategy determination module 104 is specifically used for: The fused feature vectors are input into the risk decision nodes of the at least two dimensions respectively, and risk assessment is performed through the at least two risk decision nodes to obtain at least two risk scores; Using a pre-defined weighted voting rule, at least two risk scores are weighted and calculated to determine the overall risk score; Obtain a preset risk threshold range and determine whether the comprehensive risk score falls within the risk threshold range; If the comprehensive risk score is within the risk threshold range, then the risk warning intervention method is determined to be the target risk management strategy; If the comprehensive risk score is not within the risk threshold range, then the risk-free intervention method is determined as the target risk management strategy.

[0083] This invention provides a risk collaborative control device applied to a central node. The central node communicates with at least two data source nodes. By acquiring risk feature vectors output by locally deployed risk control analysis sub-models from at least two data source nodes, it does not acquire or store the original business risk control data of each node. Instead, the local sub-models of the data source nodes perform pre-analysis of the data, and the original data is retained locally, effectively avoiding the leakage of original business privacy data during cross-node transmission. By aligning the risk feature vectors of at least two data source nodes, a cross-node risk correlation graph is constructed. This graph can be used to sort out the correlation between various data source nodes, breaking the problem of fragmented risk control data from multiple parties and helping to identify hidden risk correlations between nodes. By assigning feature weights to each risk feature vector based on its risk contribution in the cross-node risk correlation graph, and combining the risk feature vectors and feature weights to generate a fused feature vector, the system can allocate corresponding feature weights based on the risk contribution of each vector, integrate multi-dimensional risk control information, and generate a unified feature vector, thereby improving the accuracy of risk feature analysis. By inputting the fused feature vector into the risk collaborative analysis model, the optimal handling strategy is selected through a weighted voting mechanism, avoiding errors in judgment based on a single risk control dimension and improving the credibility of risk assessment results. Based on the target risk handling strategy, corresponding business intervention operations are executed, achieving multi-node linked risk control, timely preventing the spread of risk across institutions, and preventing regional chain financial risks.

[0084] Specific limitations regarding the risk collaborative control device can be found in the limitations of the intelligent question-and-answer method described above, and will not be repeated here. Each module in the aforementioned risk collaborative control device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0085] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a risk collaborative control method on the server side.

[0086] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a risk collaborative control method on the client side.

[0087] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain risk feature vectors output by a locally deployed risk control analysis sub-model from at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected from the data source nodes; The risk feature vectors of the at least two data source nodes are aligned to construct a cross-node risk association graph, wherein the data source nodes are the graph nodes and the risk feature vectors are the node attributes of the graph nodes. Based on the risk contribution of each risk feature vector in the cross-node risk association graph, feature weights are assigned to each risk feature vector, and a fused feature vector is generated by combining the risk feature vectors and the feature weights. The fused feature vector is input into a pre-trained risk collaborative analysis model, and a weighted voting mechanism is used to determine the target risk management strategy from a preset set of management strategies. Execute corresponding business intervention operations based on the target risk management strategy.

[0088] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain risk feature vectors output by a locally deployed risk control analysis sub-model from at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected from the data source nodes; The risk feature vectors of the at least two data source nodes are aligned to construct a cross-node risk association graph, wherein the data source nodes are the graph nodes and the risk feature vectors are the node attributes of the graph nodes. Based on the risk contribution of each risk feature vector in the cross-node risk association graph, feature weights are assigned to each risk feature vector, and a fused feature vector is generated by combining the risk feature vectors and the feature weights. The fused feature vector is input into a pre-trained risk collaborative analysis model, and a weighted voting mechanism is used to determine the target risk management strategy from a preset set of management strategies. Execute corresponding business intervention operations based on the target risk management strategy.

[0089] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0090] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0091] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0092] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0093] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A risk collaborative control method, characterized in that, Applied to a central node, which communicates with at least two data source nodes, the method includes: Obtain risk feature vectors output by the risk control analysis sub-model deployed locally on the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes; The risk feature vectors of the at least two data source nodes are aligned to construct a cross-node risk association graph, wherein the data source nodes are the graph nodes and the risk feature vectors are the node attributes of the graph nodes. Based on the risk contribution of each risk feature vector in the cross-node risk association graph, feature weights are assigned to each risk feature vector, and a fused feature vector is generated by combining the risk feature vectors and the feature weights. The fused feature vector is input into a pre-trained risk collaborative analysis model, and a weighted voting mechanism is used to determine the target risk management strategy from a preset set of management strategies. Execute corresponding business intervention operations based on the target risk management strategy.

2. The risk collaborative control method as described in claim 1, characterized in that, The step of aligning the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph includes: A pre-trained feature alignment mapping subnet is obtained, which is trained using the risk collaborative analysis model as the teacher model and historical risk feature samples through knowledge distillation. The risk feature vectors of the at least two data source nodes are classified and aggregated according to the node identifier; The aggregated risk feature vectors are sequentially input into the feature alignment mapping subnet. According to the standard feature rules learned by the teacher model, the aggregated risk feature vectors are subjected to dimension mapping and distribution normalization to obtain a structure-aligned risk feature vector group. The risk feature vector group is subjected to graph association operation to obtain vector attribute association parameters, wherein the vector attribute association parameters include node attribute parameters and association edge parameters; The graph is encapsulated based on the node attribute parameters and the associated edge parameters to obtain a cross-node risk association graph.

3. The risk collaborative control method as described in claim 2, characterized in that, The process involves sequentially inputting the aggregated risk feature vectors into the feature alignment mapping subnet, and performing dimensional mapping and distribution normalization on the aggregated risk feature vectors according to the standard feature rules learned by the teacher model, to obtain a structure-aligned risk feature vector group, including: The collected risk feature vectors are sorted and organized to obtain an ordered set of risk feature vectors. Sequentially extract a single set of risk feature vectors from the ordered set of risk feature vectors and perform admission verification processing according to the verification rules of the feature alignment mapping subnet to obtain the risk feature vector to be aligned; Obtain the standard feature rules learned by the teacher model, and perform dimensional mapping on the risk feature vector to be aligned according to the standard feature rules to obtain a risk feature vector with unified dimensions. The risk feature vector with the unified dimension is subjected to distribution normalization to eliminate the difference in numerical distribution and obtain the normalized risk feature vector. By summarizing and integrating all normalized risk feature vectors, a structure-aligned risk feature vector group is obtained.

4. The risk collaborative control method as described in claim 2, characterized in that, The graph association operation on the risk feature vector group to obtain vector attribute association parameters includes: Obtain the preset subject matching rules, perform subject attribution matching on all vectors in the risk feature vector group, divide the vector set corresponding to the same business subject, and obtain the subject grouping result; Traverse any two risk feature vectors in the main grouping results to form a pairing vector, and use a preset similarity measurement function to solve the similarity of the pairing vectors to obtain the initial association score; The initial association scores are filtered using a preset association threshold. First pairing vectors that are lower than the association threshold are removed, while second pairing vectors that are higher than the association threshold and their corresponding initial association scores are retained. Construct the associated edges of the second pairing vector, and map the initial associated scores of the second pairing vector to the associated edge parameters of the second pairing vector; The self-attribute features of all vectors in the risk feature vector group are mapped to node attribute parameters, and the node attribute parameters are integrated with the associated edge parameters to obtain vector attribute association parameters.

5. The risk collaborative control method as described in claim 1, characterized in that, The step of assigning feature weights to each risk feature vector based on its risk contribution in the cross-node risk association graph, and generating a fused feature vector by combining the risk feature vectors and the feature weights, includes: From the risk feature vector of the cross-node risk association graph, business time series attribute parameters are extracted. These business time series attribute parameters are used to characterize the business occurrence time of the business risk control data. Obtain a preset time decay function, perform time decay calculation based on the time decay function and the business time sequence attribute parameters to obtain a risk decay factor, and assign the risk decay factor to the risk contribution of the risk feature vector in the time dimension. The risk feature vector is globally reduced in weight based on the risk attenuation factor to obtain the reduced-weight risk feature vector. The longer the business occurred, the larger the corresponding risk attenuation factor and the greater the reduction in vector weight. The reduced-weight risk feature vector is normalized using a normalization function to obtain a normalized risk feature vector. The normalized risk feature vector is subjected to feature dimensionality reduction processing to obtain the fused feature vector.

6. The risk collaborative control method as described in claim 5, characterized in that, The step of performing global weight reduction processing on the risk feature vector according to the risk attenuation factor to obtain the weighted risk feature vector includes: Obtain the features and weight parameters of each dimension within the risk feature vector; The risk attenuation factor is synchronously superimposed on the weight parameters of each dimension feature, and the weight parameters are subjected to a global scaling operation to obtain the adjusted weight parameters. The reduced-weight risk feature vector is obtained by weighting and summing the features of each dimension and the adjusted weight parameters.

7. The risk collaborative control method as described in claim 1, characterized in that, The risk collaborative analysis model internally deploys risk decision nodes with at least two dimensions. The process of inputting the fused feature vector into the pre-trained risk collaborative analysis model and using a weighted voting mechanism to determine the target risk management strategy from a pre-set set of management strategies includes: The fused feature vectors are input into the risk decision nodes of the at least two dimensions respectively, and risk assessment is performed through the at least two risk decision nodes to obtain at least two risk scores; Using a pre-defined weighted voting rule, at least two risk scores are weighted and calculated to determine the overall risk score; Obtain a preset risk threshold range and determine whether the comprehensive risk score falls within the risk threshold range; If the comprehensive risk score is within the risk threshold range, then the risk warning intervention method is determined to be the target risk management strategy; If the comprehensive risk score is not within the risk threshold range, then the risk-free intervention method is determined as the target risk management strategy.

8. A risk collaborative control device, characterized in that, Applied to a central node, which communicates with at least two data source nodes, including: The feature acquisition module is used to acquire risk feature vectors output by the risk control analysis sub-model deployed locally by the at least two data source nodes, wherein the risk feature vectors are obtained by the risk control analysis sub-model after analyzing and processing the business risk control data collected by the data source nodes; The graph construction module is used to perform feature alignment on the risk feature vectors of the at least two data source nodes to construct a cross-node risk association graph, wherein the cross-node risk association graph uses the data source nodes as graph nodes and the risk feature vectors as the node attributes of the graph nodes. The weight allocation module is used to allocate feature weights to each risk feature vector based on the risk contribution of each risk feature vector in the cross-node risk association graph, and to generate a fused feature vector by combining the risk feature vector and the feature weights. The strategy determination module is used to input the fused feature vector into a pre-trained risk collaborative analysis model and determine the target risk disposal strategy from a preset set of disposal strategies using a weighted voting mechanism. The strategy execution module is used to execute corresponding business intervention operations according to the target risk management strategy.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the risk collaborative control method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the risk collaborative control method as described in any one of claims 1 to 7.

Images