A method, device and storage medium for identifying risk causes

Abstract

The embodiments of this specification relate to the field of big data technology, and specifically disclose a risk cause identification method, device and storage medium, the method comprising: obtaining a specified risk control object set corresponding to a specified risk level; merging the risk control objects whose object labels meet the label consistency requirements in the specified risk control object set into one node, and determining the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set, to obtain a specified risk relationship map corresponding to the specified risk level; based on the edge importance factor and the node feature importance factor, screening the edges or nodes whose importance meets the specified importance requirements from the specified risk relationship map, and constructing a risk pattern recognition map corresponding to the specified risk level, so as to identify the risk causes of the risk control objects based on the risk pattern recognition map, thereby further improving the accuracy of risk cause identification.

Description

Technical Field

[0001] The present specification relates to the field of big data technology, and in particular, to a risk cause identification method, device and storage medium. Background Art

[0002] Identifying customer risk levels is an important part of risk management. By connecting the relationships between customers in a graph format and using a graph neural network model to identify risk levels, we can make better use of customer-related information and obtain better customer risk level identification results. However, in actual application scenarios, customer characteristics and relationships between customers are complex and changeable, and how to better analyze the causes of risk has become a technical problem that needs to be solved in this field. Summary of the invention

[0003] The purpose of the embodiments of this specification is to provide a risk cause identification method, device and storage medium, which can identify the influencing factors that cause the risk control object to be at different risk levels, and further improve the accuracy of risk cause identification.

[0004] This specification provides a risk cause identification method, device and storage medium which are implemented in the following ways:

[0005] A method for identifying risk causes, applied to a server, comprises: obtaining a specified risk control object set corresponding to a specified risk level; the specified risk control object set comprises risk control objects corresponding to the specified risk level; merging the risk control objects whose object labels in the specified risk control object set meet label consistency requirements into one node, and determining the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set, to obtain a specified risk relationship map corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship map are determined based on the object features of the risk control objects merged into the corresponding nodes; based on the edge importance factor and the node feature importance factor, edges or nodes whose importance meets the specified importance requirements are screened from the specified risk relationship map, to construct a risk pattern recognition map corresponding to the specified risk level, so as to identify the risk causes of risk control objects based on the risk pattern recognition map.

[0006] On the other hand, an embodiment of the present specification provides a risk cause identification device, which is applied to a server, and the device includes: an acquisition module, which is used to obtain a specified risk control object set corresponding to a specified risk level; the specified risk control object set includes the risk control objects corresponding to the specified risk level; a graph construction module, which is used to merge the risk control objects whose object labels in the specified risk control object set meet the label consistency requirements into one node, and determine the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set, so as to obtain a specified risk relationship graph corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship graph are determined based on the object features of the risk control objects merged into the corresponding nodes; a risk pattern determination module, which is used to screen edges or nodes whose importance meets the specified importance requirements from the specified risk relationship graph based on the edge importance factor and the node feature importance factor, and construct a risk pattern identification graph corresponding to the specified risk level, so as to identify the risk causes of the risk control objects based on the risk pattern identification graph.

[0007] On the other hand, an embodiment of the present specification provides a computer-readable storage medium having computer instructions stored thereon, and when the instructions are executed, the steps of the method described in any one or more of the above embodiments are implemented.

[0008] The risk cause identification method, device and storage medium provided by one or more embodiments of the present specification can more effectively characterize the impact of object labels, object characteristics of risk control objects with similar object labels and the relationship between risk control objects on the risk control objects falling into a certain risk level by merging risk control objects with similar object labels into one node. Furthermore, by determining the edge importance factor and the node feature importance factor in the risk relationship map obtained after the merger, the edges and nodes that have a greater impact on the risk control objects falling into a certain risk level are extracted based on the edge importance factor and the node feature importance factor, and a risk pattern recognition map is constructed based on the extracted edges and nodes, which can more concisely and accurately characterize the impact of object labels, object characteristics of risk control objects with similar object labels and the relationship between risk control objects on the risk control objects falling into a certain risk level, and further improve the accuracy of risk cause identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0010] Figure 1A schematic diagram of the risk pattern identification process provided for this manual;

[0011] Figure 2 A schematic diagram of the risk cause analysis process provided for this manual;

[0012] Figure 3 A schematic diagram of the implementation process of the risk cause identification method provided in this manual;

[0013] Figure 4 This is a schematic diagram of the module structure of the risk cause identification device provided in this manual. DETAILED DESCRIPTION

[0014] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below in conjunction with the drawings in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all of the embodiments. Based on one or more embodiments of the specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the embodiments of this specification.

[0015] In a scenario example provided in this specification, the risk cause identification method can be applied to a single server or a server cluster consisting of multiple servers. Figure 1 to Figure 2 As shown, the risk cause identification method may include at least two steps: risk pattern identification and risk cause analysis. Figure 1 As shown, the risk prediction model can be constructed using the following steps S101 to S103:

[0016] Step S101: constructing an initial risk control object relationship graph.

[0017] Define the graph structure G = (V, E), where V represents nodes and E represents the edges between nodes. Corresponding to the risk classification scenario of risk control objects, V represents the risk control object and E represents the relationship between risk control objects. Risk control objects can be individual users, enterprises, etc. Correspondingly, the relationship between risk control objects can be investment relationship, guarantee relationship, etc. Risk control object and risk control object (v i ,v j ) is related to e i,j , different e i,j Construct the adjacency matrix A∈R between the risk control object relationships n*n , initial state A i,j =e i,j When (v i ,v j ) When there is a direct association in the relationship graph i,j =1, otherwise ei,j =0.

[0018] Remember each risk control object v i There are K feature variables

[0019] For each risk control object v i Record its object tag Indicates the risk control object v i With this label, otherwise

[0020] Mark risk control object v i Risk Level When the risk control object v i When the risk level is c, on the contrary Each risk control object can only correspond to one risk level.

[0021] Step S102: Identify the risk level of the risk control object.

[0022] Let the model be represents the risk level, Φ(·) represents the graph neural network model, G(v) represents the graph structure relationship of v, X(v) represents the feature vector of v, and B(v) represents the label vector of v.

[0023] Step S103: Risk pattern identification.

[0024] For each risk control object set of risk level c, N typical risk control objects are taken respectively. The typical risk control objects in each risk control object set are integrated into a large graph to obtain a specified risk relationship graph corresponding to risk level c. The above-mentioned typical risk control objects can be all risk control objects under risk level c, or they can be some risk control objects extracted according to certain requirements. For example, risk control objects with more label types, more feature variable types, and more relationship information that can be obtained from other risk control objects can be extracted. By extracting this type of risk control objects for risk pattern recognition, the final risk pattern recognition results can be made more accurate and more representative. Of course, business personnel can set the typical risk control object selection method according to the needs of the business scenario, and there is no limitation here.

[0025] For risk control level c, the risk control object points of the same type can be merged into the same node first. The same type of risk control objects are mainly judged based on labels, and those with higher label consistency are merged into the same node. The relationship between nodes is marked according to the risk control objects associated with the nodes. Based on the above preliminary merger and node association on the server, business personnel adjust the point-edge relationship to make the point-edge relationship more accurate, and obtain the specified risk relationship map corresponding to risk level c. The mean of the corresponding characteristic variables of each risk control object merged into a node can be used as the value of the characteristic variable corresponding to the node, as the node feature vector of the node. It is also possible to record information such as the object labels involved on each node and the frequency of occurrence of each object label in the specified risk relationship map as the node label of the corresponding node.

[0026] The relevant symbols are defined as follows:

[0027] The designated risk relationship graph corresponding to the risk control level c is marked as G c (v) The adjacency matrix is ​​A c (v), the node feature vector is X c (v) = {x i |v i ∈G c (v)}, label vector B c (v) = {b i |v i ∈G c (v)}、Prediction result P Φ (Y|G c ,X c ).

[0028] The sub-graph of the specified risk relationship graph can be extracted to reduce the amount of data to be calculated and reduce the resource consumption of the system calculation. c (v) Within the range, extract subgraph G c,S (v). The system defaults to risk control object v i As a starting point, extend the relationship outward for 3 steps. Other subgraph extraction methods can also include: limiting a number of label risk control objects, limiting the number of nodes in the subgraph, limiting the number of edges in the subgraph, etc. The subgraph extraction method can be set according to actual business needs. Of course, if the amount of data is not large, it is also possible to not extract the subgraph and directly perform subsequent processing based on the specified risk relationship map. The processing method is the same as the processing method based on the extracted subgraph, which will not be repeated here.

[0029] Assume that the extracted subgraph is labeled G c,S (v) The adjacency matrix is ​​A c,S (v), the node feature vector is X c,S (v) = {x i |vi ∈G c,S (v)}, node label vector B c,S (v) = {b i |v i ∈G c,S (v)}、Prediction result P Φ (Y|G c,S ,X c,S ). Of course, if the subsequent analysis is performed directly based on the specified risk relationship map without sub-graph extraction, it can be understood that the above sub-graph label is the same as the G of the specified risk relationship map. c (v) A c (v) X c (v) = {x i |v i ∈G c (v)}, B c (v) = {b i |v i ∈G c (v)}, P Φ (Y|G c ,X c ) have the same meaning, but are expressed in different ways.

[0030] The recognition mode is obtained by solving the first objective function as follows.

[0031]

[0032]

[0033] in, is an n-dimensional vector, each element of which takes the value of 1; y represents the target variable, Y represents the vector composed of the target variable, c represents the risk level, C represents the upper limit of the risk level, P Φ represents the risk prediction result, G represents the graph, A c,S represents the adjacency matrix of the specified risk relationship graph, represents the dot product of the matrix, σ(M) represents the sigmoid function, σ(M)∈[0,1] n*n , n represents the adjacency matrix A c,S The number of rows or columns, M represents the edge mask function, H(·) represents entropy, X represents the feature vector, X c,S G represents the feature vector formed by the node features of the specified risk relationship graph; c,S Represents the specified risk relationship graph; F represents the mask function of the node feature, F∈{0,1} K , K represents the number of characteristic variables of node features.

[0034] According to the above objective function, A c,S·σ(M) will A c,S In the algorithm, important edge relationships that affect the prediction results are retained, and minor edge relationships are weakened. K It is a mask function of the feature. It takes values ​​from 0 to 1 to retain the features that are important to the prediction results and weaken the minor features. Of course, if a feature or edge is very important and must be retained, the value of the corresponding element in M ​​or F can also be fixed.

[0035] The mask functions of the edges and the mask functions of the node features obtained by solving the above objective function can be marked as M * 、F * Correspondingly, the relational model of the risk pattern recognition graph can be obtained And characteristic patterns

[0036] Elements of Represents node v i With v j The strength of the relationship. The larger the value, the more important the relationship corresponding to this edge is to the prediction result. By setting a certain threshold, the edges with lower importance are screened out and the edges with higher importance are retained.

[0037] Indicates the importance of the node feature of a node. Retaining the node indicates that it is important for the prediction result, and not retaining the node indicates that it is less important. Remove the nodes with lower importance and retain the nodes with higher importance to obtain

[0038] use Constructing a risk pattern recognition map in, That is the adjacency matrix of the risk pattern recognition graph. Indicates the labels on the retained nodes and the relationship identifiers corresponding to the edges, and uses text to describe the risk control object v i Factors that affect different risk levels are analyzed to make risk interpretation closer to business scenarios and easier to understand.

[0039] Accordingly, the server can generate a risk description of the risk control object based on the risk pattern recognition graph. The risk description of the risk control object may include at least one of the node label, node features, and risk conduction mode of the node where the risk control object is located. The risk conduction mode may be characterized by the edge between the node where the risk control object is located and other nodes and the edge importance factor of the edge. For example, the risk conduction mode may be expressed as:

[0040] “××× (risk level) risk control object v iThe first-level risk transmission mode is the edge relationship A1 (such as loan) (affecting importance, taking the value of the corresponding element of the neighboring matrix, the same below), A2 (such as guarantee) (affecting importance), A3 (affecting importance). The second-level risk transmission mode is the edge relationship B1 (affecting importance), B2 (affecting importance), B3 (affecting importance), B4 (affecting importance), .... "Among them, A1, A2, A3 are related to the risk control object v i The node relationship corresponding to the edge directly connected to the node Q, B1, B2, B3, B4 are the nodes with the risk control object v i The node relationship corresponding to the edge connected to the node Q that is separated by one node. Of course, it can also include the three-level risk transmission mode, the four-level risk transmission mode, etc., but generally, the farther the distance from the node Q, the smaller the risk impact of the edge relationship on the node Q. The corresponding risk transmission level can be selected for analysis as needed.

[0041] The above implementation method, by merging risk control objects with similar object labels into one node, can more effectively characterize the impact of object labels, object features of risk control objects with similar object labels, and the relationship between risk control objects on the risk control objects falling into a certain risk level. Furthermore, by determining the edge importance factor and the node feature importance factor in the risk relationship map obtained after the merger, the edges and nodes that have a greater impact on the risk control objects falling into a certain risk level are extracted based on the edge importance factor and the node feature importance factor, and a risk pattern recognition map is constructed based on the extracted edges and nodes, which can more concisely and accurately characterize the impact of object labels, object features of risk control objects with similar object labels, and the relationship between risk control objects on the risk control objects falling into a certain risk level.

[0042] The risk influencing factors can be analyzed based on the risk pattern recognition map corresponding to each risk level. The above risk pattern recognition map is constructed based on typical risk control objects, and typical risk control objects may not be able to fully characterize the characteristics of all risk control objects at each risk level. Accordingly, this scenario example further provides a general risk cause (i.e., risk influencing factors) analysis method based on the above risk pattern recognition map.

[0043] Figure 2 It is a flow chart of the risk cause analysis module.

[0044] Step S201: Risk level cause analysis of risk control objects

[0045] For the risk control object set of risk level c, for each risk control object v i Extract subgraphs. The subgraph extraction rules are the same as the previous steps. Mark the extracted risk control object relationship subgraph as G i,c,S (v) The adjacency matrix is ​​A i,c,S (v), the eigenvector is Xi,c,S (v) = {x i |v i ∈G i,c,S (v)}.

[0046] Solve the following second objective function:

[0047]

[0048] Get each risk control object v under risk level c i The corresponding P c T A i,c,S P c Among them, P c represents the adjustment coefficient matrix, Indicates P c The transposition of . Each risk control object v under risk level c can be i Corresponding Take the mean or median and get or You can As the benchmark adjacency matrix for risk level c, we use To characterize the risk transmission pattern of risk level c.

[0049] From the above process, we can see that It can also characterize the risk transmission mode of risk level c, but The risk objects used for calculation may not be comprehensive enough. Based on this, we further calculate use To characterize the risk transmission mode of risk level C, we can more accurately and comprehensively characterize the risk transmission characteristics of risk level C. Of course, by first conducting , and then calculate based on the risk control object set of risk level c to obtain use To characterize the risk transmission mode of risk level c, it can also reduce the overall calculation amount and improve the overall processing efficiency.

[0050] Of course, through the above objective function, we can also get each risk control object v under risk level c i Corresponding You can also perform the following operations on each risk control object v under risk level c: i Corresponding Take the mean or median and get You can As the benchmark node feature of risk level c, use To characterize the impact of object characteristics on the risk control object falling into risk level c.

[0051] Correspondingly, the server can also use any object tag of risk level c as the specified object tag, and extract any risk control object v of the specified object tag under risk level c. i Corresponding risk control object relationship subgraph G i,c,S The occurrence frequency in (v) is used as the specified label frequency; the relationship subgraph G corresponding to each risk control object under risk level c is calculated i,c,S (v) The mean or median of the specified label frequency is used to obtain the benchmark label frequency of the specified object label at risk level c, so as to analyze the risk level cause of the risk level c according to the benchmark label frequency. By calculating the benchmark label frequency in the above manner, the benchmark label frequency can be used to more accurately and comprehensively characterize the impact of the object label on the risk control object falling into risk level c.

[0052] Step S202: Scenario analysis of risk level changes of risk control objects

[0053] For the target risk control object whose risk cause is to be identified, the server can obtain the target risk control object relationship subgraph corresponding to the target risk control object; calculate the adjacency matrix of the target risk control object relationship subgraph as the target adjacency matrix; and compare the target adjacency matrix with the benchmark adjacency matrix under each risk level to determine the influencing factors of the target risk control object falling into each risk level.

[0054] For example, the target risk control object v i Extract according to the subgraph extraction method to obtain the target risk control object relationship subgraph corresponding to the target risk control object. And the target risk control object v can be calculated i The corresponding adjacency matrix A i,S (v i ). i,S (v i ) and A at different risk levels c Compare and analyze the difference between the adjacency matrices of the two. Business personnel can adjust the degree of association between any nodes in the target risk control object relationship subgraph, and the server can calculate the adjacency matrix of the adjusted target risk control object relationship subgraph and compare it with A. i,S (v i ) to the element corresponding to the node being adjusted The corresponding element value in the , if the element value is closer, it can be explained that the adjustment operation of the correlation degree between the two nodes makes the risk control object v i There is a greater possibility of falling into risk level C.

[0055] For example, A i,S (v i )[v p ,v q]=1, Description for v i , if the associated node v in its subgraph is reduced p With v q The degree of correlation, then v i The risk level may become c. Further analysis, if the risk control object v i The risk level (assuming the calculation result is B) changes to B+ / B-, which transmission factors affect its change, which nodes’ correlation degree needs to be changed to increase / decrease the risk level, and generate a risk analysis report.

[0056] It can be seen that identifying the causes of risks through the above method can greatly improve the accuracy and comprehensiveness of risk analysis and investigation, and can facilitate business personnel to more clearly and intuitively determine the risk influencing factors that cause the risk control object to fall into a certain risk level, thereby greatly improving the accuracy, comprehensiveness and efficiency of risk cause identification in complex business scenarios.

[0057] Based on the above scenario examples, this specification also provides a method for identifying risk causes. Figure 3 This is a flow chart of an embodiment of the risk cause identification method provided in this specification. Figure 3 As shown, in one embodiment of the risk cause identification method provided in this specification, the method can be applied to a server. The method can include the following steps.

[0058] S30: Obtain a specified risk control object set corresponding to a specified risk level; the specified risk control object set includes the risk control objects corresponding to the specified risk level.

[0059] The specified risk level can be any risk level. The specified risk control object set can include all risk control objects under the specified risk level, or can only include extracted typical risk control objects, which is not limited here. The method of extracting typical risk control objects can refer to the above scenario examples. By extracting typical risk control objects, the amount of data for subsequent processing can be greatly reduced, and the processing efficiency can be improved.

[0060] S32: Merge the risk control objects whose object labels meet the label consistency requirements in the specified risk control object set into one node, and determine the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set to obtain a specified risk relationship graph corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship graph are determined based on the object features of the risk control objects merged into the corresponding node.

[0061] The object labels can be configured as needed to meet the label consistency requirements, which are not limited here. By setting the label consistency requirements, object labels with high object label consistency can be extracted to merge the risk control objects into one node. Other implementation methods for specifying the construction of risk relationship maps can refer to the above scenario examples, which will not be repeated here.

[0062] S34: Based on the edge importance factor and the node feature importance factor, edges or nodes whose importance meets the specified importance requirement are selected from the specified risk relationship map, and a risk pattern recognition map corresponding to the specified risk level is constructed to identify the risk causes of the risk control object based on the risk pattern recognition map.

[0063] In some embodiments, the edge importance factor and the node feature importance factor may be determined in the following manner: Solve the following first objective function based on the specified risk relationship graph:

[0064]

[0065] in, is an n-dimensional vector, each element of which takes the value of 1; y represents the target variable, Y represents the vector composed of the target variable, c represents the risk level, C represents the upper limit of the risk level, P Φ represents the risk prediction result, G represents the graph, A c,S represents the adjacency matrix of the specified risk relationship graph, represents the dot product of the matrix, σ(M) represents the sigmoid function, σ(M)∈[0,1] n*n , n represents the adjacency matrix A c,S The number of rows or columns, M represents the edge mask function, H(·) represents entropy, X represents the feature vector, X c,S G represents the feature vector formed by the node features of the specified risk relationship graph; c,S Represents the specified risk relationship graph; F represents the mask function of the node feature, F∈{0,1} K , K represents the number of characteristic variables of node features;

[0066] The mask functions of the edges and the mask functions of the node features obtained by solving the above objective function are marked as M * 、F * , and respectively with σ(M * ), F * As edge importance factor and node feature importance factor.

[0067] Afterwards, the edges or nodes whose importance meets the specified importance requirements can be screened based on the edge importance factor and node feature importance factor. Feature Mode The extracted relationship patterns and feature patterns also indirectly characterize the importance of the edge relationships and node features on the risk control object falling into risk level c. At the same time, each node also corresponds to a node label, and the frequency of label occurrence in the node label can also intuitively determine the importance of the object label on the risk control object falling into risk level c, so that based on the risk pattern recognition map, the influencing factors of the risk control object falling into a certain risk level can be accurately and intuitively determined.

[0068] Of course, the above edge importance factors and node feature importance factors are only preferred examples. In specific implementations, appropriate variations of the above determination process are also included in the protection scope of the embodiments of this specification. Of course, in addition to the mask function, other functions with similar characteristics can also be used to characterize the impact of edges or features on the prediction results.

[0069] In some embodiments, the risk pattern identification map may also include a risk description of the risk control object; wherein, the risk description of the risk control object includes at least a node label, node characteristics and a risk transmission pattern of the node where the risk control object is located; wherein, the risk transmission pattern is characterized by the edges between the node where the risk control object is located and other nodes and the edge importance factor of the edges.

[0070] In some other embodiments, based on the determination of the risk pattern recognition graph based on the typical risk control object, the following processing is further performed. Based on the risk pattern recognition graph, the following second objective function is solved to obtain : Among them, P c represents the adjustment coefficient matrix, Indicates P c The transpose of Relational patterns representing risk pattern recognition graphs, The characteristic patterns representing the risk pattern recognition graph, A i,c,S Represents the risk control object v under risk level c i The adjacency matrix of the corresponding risk control object relationship subgraph, X i,c,S Indicates the risk control object v i According to the corresponding characteristics of each risk control object under risk level c. Determine the benchmark adjacency matrix of risk level c, and identify the risk causes of the risk control object according to the benchmark adjacency matrix. The mean or median of , etc., is used to obtain the benchmark neighbor matrix of risk level c.

[0071] Correspondingly, in some other embodiments, the server can obtain the target risk control object relationship subgraph corresponding to the target risk control object for which the risk cause is to be identified; calculate the adjacency matrix of the target risk control object relationship subgraph as the target adjacency matrix; and compare the target adjacency matrix with the benchmark adjacency matrix under each risk level to determine the influencing factors of the target risk control object falling into each risk level.

[0072] In some other embodiments, the server may also use any object tag of risk level c as a designated object tag, and extract any risk control object v of the designated object tag under risk level c. i Corresponding risk control object relationship subgraph G i,c,S The occurrence frequency in (v) is used as the specified label frequency. According to the risk level c, the corresponding subgraph G of each risk control object relationship i,c,S (v) The specified label frequency of the specified object label is determined to determine the benchmark label frequency of the specified object label at risk level c, so as to analyze the risk level cause of the risk level c according to the benchmark label frequency.

[0073] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. For details, please refer to the description of the above-mentioned related processing related embodiments, and no further description is given here.

[0074] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0075] like Figure 4As shown, based on the method provided in the above embodiment, the embodiment of this specification also provides a risk cause identification device, which is applied to a server, and the device includes: an acquisition module 40, which can be used to obtain a specified risk control object set corresponding to a specified risk level; the specified risk control object set includes the risk control objects corresponding to the specified risk level. A graph construction module 42, which can be used to merge the risk control objects whose object labels meet the label consistency requirements in the specified risk control object set into one node, and determine the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set, to obtain a specified risk relationship graph corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship graph are determined based on the object features of the risk control objects merged under the corresponding node. A risk pattern determination module 44, which can be used to screen edges or nodes whose importance meets the specified importance requirements from the specified risk relationship graph based on the edge importance factor and the node feature importance factor, and construct a risk pattern identification graph corresponding to the specified risk level, so as to identify the risk causes of the risk control objects based on the risk pattern identification graph.

[0076] In some other embodiments, the risk pattern determination module is further used to determine the edge importance factor and the node feature importance factor based on the following method: solving the following first objective function based on the specified risk relationship graph:

[0077]

[0078] in, is an n-dimensional vector, each element of which takes the value of 1; y represents the target variable, Y represents the vector composed of the target variable, c represents the risk level, C represents the upper limit of the risk level, P Φ represents the risk prediction result, G represents the graph, A c,S represents the neighbor matrix of the specified risk relationship graph, represents the dot product of the matrix, σ(M) represents the sigmoid function, n represents the adjacency matrix A c,S The number of rows or columns, M represents the edge mask function, H(·) represents entropy, X represents the feature vector, X c,S G represents the feature vector formed by the node features of the specified risk relationship graph; c,S Represents the specified risk relationship graph; F represents the mask function of the node feature, F∈{0,1} K , K represents the number of characteristic variables of node features; the mask function of the edge and the mask function of the node features obtained by solving the above objective function are marked as M * 、F * , and respectively with σ(M * ), F * As edge importance factor and node feature importance factor.

[0079] In some other embodiments, the device further comprises a benchmark risk model determination module, which is used to solve the following second objective function based on the risk pattern recognition graph to obtain Among them, P c represents the adjustment coefficient matrix, Indicates P c The transpose of Relational patterns representing risk pattern recognition graphs, The characteristic patterns representing the risk pattern recognition graph, A i,c,S Represents the risk control object v under risk level c i The adjacency matrix of the corresponding risk control object relationship subgraph, X i,c,S Indicates the risk control object v i object characteristics; and, also used to correspond to each risk control object under risk level c Determine a benchmark adjacency matrix for risk level c to identify risk causes of the risk control object based on the benchmark adjacency matrix.

[0080] It should be noted that the above-mentioned device may also include other implementations according to the description of the above-mentioned embodiment. The specific implementation methods may refer to the description of the relevant method embodiments, which will not be described one by one here.

[0081] This specification also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed, the steps of the method described in any one or more of the above embodiments are implemented. The storage medium may include a physical device for storing information, which is usually to digitize the information and then store it in a medium using electrical, magnetic or optical methods. The storage medium may include: devices that store information using electrical energy, such as various types of memories, such as RAM, ROM, etc.; devices that store information using magnetic energy, such as hard disks, floppy disks, magnetic tapes, magnetic core memories, bubble memories, and U disks; devices that store information optically, such as CDs or DVDs. Of course, there are other types of readable storage media, such as quantum memories, graphene memories, and so on.

[0082] It should be noted that the embodiments of this specification are not limited to the situations that must comply with the standard data model / template or the embodiments of this specification. Certain industry standards or slightly modified implementation plans based on the implementation described in the custom method or the embodiment can also achieve the same, equivalent or similar, or predictable implementation effects after deformation of the above-mentioned embodiments. The embodiments obtained by using these modified or deformed data acquisition, storage, judgment, processing methods, etc. can still fall within the scope of the optional implementation plans of this specification.

[0083] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. In the description of this specification, the description of the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of this specification. In this specification, the schematic representation of the above terms does not necessarily target the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, in the absence of contradiction, a person skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0084] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A method for identifying risk causes, characterized in that: Applied to a server, the method comprises: Obtain a specified risk control object set corresponding to a specified risk level; the specified risk control object set includes the risk control objects corresponding to the specified risk level; The risk control objects whose object labels in the specified risk control object set meet the label consistency requirements are merged into one node, and the edges between the nodes are determined based on the relationship between the risk control objects in the specified risk control object set to obtain a specified risk relationship graph corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship graph are determined based on the object features of the risk control objects merged into the corresponding node; Based on the edge importance factor and the node feature importance factor, edges or nodes whose importance meets the specified importance requirement are selected from the specified risk relationship graph, and a risk pattern recognition graph corresponding to the specified risk level is constructed to identify the risk causes of the risk control object based on the risk pattern recognition graph; The edge importance factor and node feature importance factor are determined in the following way: Based on the specified risk relationship graph, the following first objective function is solved: in, is an n-dimensional vector, each element of which takes the value of 1; y represents the target variable, Y represents the vector composed of the target variable, c represents the risk level, C represents the upper limit of the risk level, P Φ represents the risk prediction result, G represents the graph, A c,S represents the adjacency matrix of the specified risk relationship graph, represents the dot product of the matrix, σ(M) represents the sigmoid function, σ(M)∈[0,1] n*n , n represents the adjacency matrix A c,S The number of rows or columns, M represents the edge mask function, H(·) represents entropy, X represents the feature vector, X c,S G represents the feature vector formed by the node features of the specified risk relationship graph; c,S Represents the specified risk relationship graph; F represents the mask function of the node feature, F∈{0,1} K , K represents the number of characteristic variables of node features; The mask functions of the edges and the mask functions of the node features obtained by solving the above objective function are marked as M * 、F * , and respectively with σ(M * ), F * As edge importance factor and node feature importance factor; The risk cause identification of the risk control object based on the risk pattern identification map includes: Based on the risk pattern recognition graph, the following second objective function is solved to obtain P c T A i,c,S P c : Among them, P c represents the adjustment coefficient matrix, P c T Indicates P c The transpose of Relational patterns representing risk pattern recognition graphs, The characteristic patterns representing the risk pattern recognition graph, A i,c,S Represents the risk control object v under risk level c i The adjacency matrix of the corresponding risk control object relationship subgraph, X i,c,S Indicates the risk control object v i Object characteristics; According to the corresponding risk control objects under risk level c Determine a benchmark adjacency matrix for risk level c to identify risk causes of the risk control object based on the benchmark adjacency matrix.

2. The method according to claim 1, characterized in that The risk pattern identification map also includes a risk description of the risk control object; wherein, the risk description of the risk control object includes at least a node label, node characteristics and a risk transmission pattern of the node where the risk control object is located; wherein, the risk transmission pattern is characterized by the edges between the node where the risk control object is located and other nodes and the edge importance factor of the edges.

3. The method according to claim 1, characterized in that The step of identifying the risk causes of the risk control object according to the benchmark adjacency matrix includes: For the target risk control object to be identified for risk causes, obtain the target risk control object relationship subgraph corresponding to the target risk control object; Calculate the adjacency matrix of the target risk control object relationship subgraph as the target adjacency matrix; The target adjacency matrix is ​​compared with the benchmark adjacency matrix at each risk level to determine the influencing factors of the target risk control object falling into each risk level.

4. The method according to claim 1, characterized in that: The method further comprises: Take any object label of risk level c as the designated object label, and extract the risk control object relationship subgraph G corresponding to any risk control object vi under risk level c with the designated object label i,c,S The frequency of occurrence in (v) is used as the frequency of the specified label; According to the risk level c, the corresponding subgraph G of each risk control object i,c,S (v) The specified label frequency of the specified object label is determined to determine the benchmark label frequency of the specified object label at risk level c, so as to analyze the risk level cause of the risk level c according to the benchmark label frequency.

5. A risk cause identification device, characterized in that: Applied to a server, the device comprises: An acquisition module, used to acquire a specified risk control object set corresponding to a specified risk level; the specified risk control object set includes the risk control objects corresponding to the specified risk level; A graph construction module is used to merge the risk control objects whose object labels meet the label consistency requirements in the specified risk control object set into one node, and determine the edges between the nodes based on the relationship between the risk control objects in the specified risk control object set, so as to obtain a specified risk relationship graph corresponding to the specified risk level; wherein the node features corresponding to each node in the specified risk relationship graph are determined based on the object features of the risk control objects merged into the corresponding node; A risk pattern determination module is used to select edges or nodes whose importance meets the specified importance requirements from the specified risk relationship map based on edge importance factors and node feature importance factors, and to construct a risk pattern recognition map corresponding to the specified risk level, so as to identify the risk causes of the risk control object based on the risk pattern recognition map; The edge importance factor and node feature importance factor are determined in the following way: Based on the specified risk relationship graph, the following first objective function is solved: in, is an n-dimensional vector, each element of which takes the value of 1; y represents the target variable, Y represents the vector composed of the target variable, c represents the risk level, C represents the upper limit of the risk level, P Φ represents the risk prediction result, G represents the graph, A c,S represents the adjacency matrix of the specified risk relationship graph, represents the dot product of the matrix, σ(M) represents the sigmoid function, σ(M)∈[0,1] n*n , n represents the adjacency matrix A c,S The number of rows or columns, M represents the edge mask function, H(·) represents entropy, X represents the feature vector, X c,S G represents the feature vector formed by the node features of the specified risk relationship graph; c,S Represents the specified risk relationship graph; F represents the mask function of the node feature, F∈{0,1} K , K represents the number of characteristic variables of node features; The mask functions of the edges and the mask functions of the node features obtained by solving the above objective function are marked as M * 、F * , and respectively with σ(M * ), F * As edge importance factor and node feature importance factor; The risk cause identification of the risk control object based on the risk pattern identification map includes: Based on the risk pattern recognition graph, the following second objective function is solved to obtain P c T A i,c,S P c : Among them, P c represents the adjustment coefficient matrix, P c T Indicates P c The transpose of Relational patterns representing risk pattern recognition graphs, The characteristic patterns representing the risk pattern recognition graph, A i,c,S Represents the risk control object v under risk level c i The adjacency matrix of the corresponding risk control object relationship subgraph, X i,c,S Indicates the risk control object v i Object characteristics; According to the P corresponding to each risk control object under risk level c c T A i,c,S P c Determine a benchmark adjacency matrix for risk level c to identify risk causes of the risk control object based on the benchmark adjacency matrix.

6. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instructions are executed, the steps of the method according to any one of claims 1 to 4 are implemented.

Images