Noise identification method, device and computer-readable storage medium for knowledge graph

Abstract

The present application provides a method, device and computer-readable storage medium for identifying noise in a knowledge graph, wherein the method comprises: splitting the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all nodes in the knowledge graph; converting each sub-graph into an adjacency matrix, inputting each adjacency matrix into a preset robust principal component analysis model, and obtaining a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship of the node number in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edge in the corresponding sub-graph; performing weighted summation on all sparse matrices to obtain a noise matrix, and determining the noise edge in the knowledge graph according to the numerical value of each element in the noise matrix. The method can effectively identify noise in the knowledge graph and has strong interpretability.

Description

Technical Field

[0001] The present application belongs to the field of knowledge graphs, and specifically relates to a noise identification method, device and computer-readable storage medium for knowledge graphs. Background Art

[0002] This section is intended to provide a background or context for the embodiments of the present application. The description herein is not admitted to be prior art by virtue of its inclusion in this section.

[0003] In the knowledge graphs of many real scenarios, a large amount of noise information will inevitably appear.

[0004] For example, in social graphs, friend relationships are a typical type of relationship. Taking WeChat as an example, there are always some contacts in the user's friend list who are rarely in contact, have no common friends, and are unlikely to contact in the future. These "unfamiliar friends" are the noise in the social graph. If the WeChat operator needs to recommend advertisements to users based on friend relationships, then this noise will interfere with the recommendation model and reduce the recommendation effect of the recommendation model. Removing this noise will help optimize the recommendation effect of the recommendation model.

[0005] For another example, in the communication network graph, the call relationships between people constitute the relationships in the communication network graph. If two people who have never talked to each other suddenly talk to each other, the probability that this call is suspected of sales promotion or fraud will be greater. These call relationships can also be regarded as a kind of noise in the communication network graph. Extracting this noise will help to discover potential sales promotion and fraud behaviors.

[0006] In the above two examples, the former requires the removal of noise from the knowledge graph, while the latter requires further analysis of the noise in the knowledge graph. Both require the identification of noise from the knowledge graph. Furthermore, most of the existing methods for identifying noise from knowledge graphs assume that the relationships in the knowledge graph are noise-free, which is not in line with the actual situation. In addition, the current classification-based knowledge graph denoising method relies on the training set. Generally speaking, a batch of training sets need to be constructed to train the classifier to determine whether the edges in the knowledge graph are noise, which increases the annotation cost. Summary of the invention

[0007] In view of the above problems in the prior art, the embodiments of the present application propose a method, device and computer-readable storage medium for identifying noise in a knowledge graph. By using this method and device, the above problems can be at least partially solved.

[0008] The following scheme is provided in an embodiment of the present application: a noise identification method for a knowledge graph, comprising:

[0009] Splitting the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all the nodes of the knowledge graph;

[0010] Convert each sub-graph into an adjacency matrix, input each adjacency matrix into a preset robust principal component analysis model, and obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship between the node numbers in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edges in the corresponding sub-graph;

[0011] All sparse matrices are weighted summed to obtain a noise matrix, and the noise edges in the knowledge graph are determined according to the values ​​of each element in the noise matrix.

[0012] The following scheme is provided in an embodiment of the present application: a noise recognition device for a knowledge graph, comprising:

[0013] A graph splitting module, used to split the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all the nodes of the knowledge graph;

[0014] A matrix decomposition module, used to convert each sub-graph into an adjacency matrix, and input each adjacency matrix into a preset robust principal component analysis model to obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship between the node numbers in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edges in the corresponding sub-graph;

[0015] The noise edge screening module is used to perform weighted summation on all sparse matrices to obtain a noise matrix, and determine the noise edges in the knowledge graph according to the values ​​of each element in the noise matrix.

[0016] The following scheme is provided in an embodiment of the present application: a noise identification device for a knowledge graph, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute: the aforementioned noise identification method for the knowledge graph.

[0017] The following solution is provided in an embodiment of the present application: a computer-readable storage medium, wherein the computer-readable storage medium stores a program, and when the program is executed by a processor, the processor executes: the aforementioned noise identification method of the knowledge graph.

[0018] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: 1. The technical solution proposed in the present application believes that the noise in the knowledge graph is objectively present and conforms to the real scene. 2. The technical solution proposed in the present application is different from the classification-based knowledge graph denoising method that relies on the training set in the prior art. It is not dependent on the training data set, and the recognition result has stronger credibility and consistency. 3. The technical solution proposed in the present application takes into account the global denoising effect, and the method has strong interpretability.

[0019] It should be understood that the above description is only an overview of the technical solution of the present application, so that the technical means of the present application can be more clearly understood and implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically described below by way of example. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] By reading the detailed description of the exemplary embodiments below, the advantages and benefits described herein and other advantages and benefits will be apparent to those skilled in the art. The accompanying drawings are only used for the purpose of illustrating exemplary embodiments and are not to be considered as limitations of the present application. In addition, the same reference numerals are used throughout the accompanying drawings to represent the same components.

[0021] Figure 1 Schematic diagram of the process of the noise identification method of the knowledge graph according to one embodiment of the present application.

[0022] Figure 2 This is a numerical example of a noise identification method for a knowledge graph according to an embodiment of the present application.

[0023] Figure 3 Schematic diagram of the structure of a noise recognition device for a knowledge graph according to an embodiment of the present application.

[0024] Figure 4 It is a structural schematic diagram of a noise recognition device for a knowledge graph according to another embodiment of the present application.

[0025] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0026] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0027] In the present application, it should be understood that terms such as “including” or “having” are intended to indicate the presence of features, numbers, steps, behaviors, components, parts, or a combination thereof disclosed in the present specification, and do not exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof.

[0028] It should also be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] Figure 1 This is a flow chart of a noise identification method for a knowledge graph according to an embodiment of the present application. The method is used to: in this flow, from the device perspective, the execution subject can be one or more electronic devices, more specifically, the processing modules therein; from the program perspective, the execution subject can correspondingly be the programs installed on these electronic devices.

[0030] Figure 1 The process in may include the following steps 101 to 103.

[0031] Step 101: split the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond one-to-one to the relationship types of the connection relationships in the knowledge graph, and the sub-graphs contain all the nodes of the knowledge graph.

[0032] In other words, the sub-graph retains the node information of the original knowledge graph and only reflects one type of connection relationship of the original knowledge graph.

[0033] For example, there are two types of connection relationships in a knowledge graph. One type of connection relationship is the consumption behavior of bank cards at merchants, which is split into sub-graph A; the other type of connection relationship is the transfer behavior between bank cards, which is split into sub-graph B.

[0034] Since a subgraph only contains connection relationships of one connection type, isolated nodes may appear in the subgraph.

[0035] Step 102: Convert each sub-graph into an adjacency matrix, input each adjacency matrix into a preset robust principal component analysis model, and obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship of the node numbers in the knowledge graph, and the low-rank matrix represents the correct connection relationship in the corresponding sub-graph. The sparse matrix represents the potential noise edge in the corresponding sub-graph.

[0036] In this step, it is considered that there are high probability noise edges in the knowledge graph, and the robust principal component analysis model is used to select potential noise edges from each sub-graph. Although there is a certain probability that the sparse matrix is ​​an all-zero matrix, in actual scenarios, this probability is very low.

[0037] Step 103: Perform weighted summation on all sparse matrices to obtain a noise matrix, and determine the noise edges in the knowledge graph according to the values ​​of each element in the noise matrix.

[0038] If two nodes have had abnormal connections (i.e., potential noise edges) multiple times before, then the probability that the relationship between the two nodes is abnormal will be greater. Therefore, it is necessary to perform statistical analysis on these potential abnormal connections to more accurately identify the truly abnormal connections.

[0039] For a certain scenario, abnormal connection relationships of different connection types should have different importance. When screening noise edges, the information expressed by all types of potential noise edges between all nodes is comprehensively evaluated.

[0040] In this way, the technical solution proposed in this application believes that noise in the knowledge graph is an objective existence and conforms to the real scene. The existing knowledge graph construction technology generally requires the construction of a batch of training sets to train the classifier to determine whether the edges in the knowledge graph are noise. The technical solution proposed in this application does not rely on the training data set, and the recognition results have stronger credibility and consistency. The technical solution proposed in this application takes into account the global denoising effect and splits the adjacency matrix of the original knowledge graph into a low-rank matrix and a sparse matrix, that is, the denoised adjacency matrix and the noise matrix. The splitting process is a denoising process for the original adjacency matrix on the one hand, which helps to improve the generation quality of the knowledge graph. On the other hand, it is also a noise matrix recognition process, which can quickly and accurately identify abnormal relationships in the original knowledge graph. This method has strong interpretability.

[0041] based on Figure 1 Some embodiments of the present application also provide some specific implementation schemes and extension schemes of the knowledge graph noise identification method. Figure 2 Provide explanation. Figure 2 Both subgraphs shown are undirected graphs.

[0042] First, we construct a knowledge graph G = (V, E) based on multiple data sources, where V is the point set of the knowledge graph and E is the edge set of the knowledge graph. For any two nodes v1, v2∈V, if there is a certain relationship r between them, then (v1, v2, r)∈E. Figure 2 There are two types of relationships in the CCP, represented by two different line types (solid line and dashed line).

[0043] Then, for each relationship r, the corresponding adjacency matrix X is calculated. r For example, in Figure 2 In , the adjacency matrices of the subgraphs represented by the two relationships are:

[0044]

[0045]

[0046] In each adjacency matrix, if the i-th row and j-th column are 1, it means that the i-th node is connected to the j-th node in the knowledge graph. Of course, the connection between the i-th node and the j-th node can also be assigned a value.

[0047] This method assumes that the knowledge graph is likely to contain noise. The adjacency matrix, such as the adjacency matrix X1, can be split into a low-rank matrix L1 and a sparse matrix S1. Since both the sparse matrix and the low-rank matrix describe undirected subgraphs, their transposed matrices should be equal to their own.

[0048] The low-rank matrix L1 is the adjacency matrix after denoising. Its low rank reflects the strong correlation between nodes. For example Figure 2 In the subgraph corresponding to the adjacency matrix X1, nodes 1, 2, and 3 are all connected, while node 4 is only connected to node 3. It is obvious that the correlation between node 4 and other nodes is far less strong than that between nodes 1, 2, and 3. Therefore, if the relationship between nodes 3 and 4 is removed, the overall rank of the adjacency matrix X1 will be reduced. By calculation, the rank of the adjacency matrix X1 is 3, and the rank of the adjacency matrix L1 is 2. By removing the relationship (noise) between nodes 3 and 4, the rank of the adjacency matrix can be reduced by 1.

[0049] The above method assumes that only a small number of relations in the knowledge graph are abnormal, so the matrix S1 is a sparse matrix.

[0050] The goals of noise stripping are:

[0051]

[0052] Where λ is a penalty factor for controlling the sparsity of S1, for example, λ = 0.01. rank(L1) represents the rank of matrix L1, ||S1||0 represents the l0 norm of matrix S1, that is, the number of non-zero elements in matrix S1. The content after st represents the constraint condition. However, the above formula is non-convex and difficult to optimize. Therefore, consider using the nuclear norm ||L1|| * and L1 norm ||S1||1 to relax the rank(L1) and ||S1||0 in the original target respectively. The new target is expressed as:

[0053]

[0054] in It is also the sum of the eigenvalues ​​of the matrix L1, and tr(·) represents the trace of the matrix (the sum of the diagonals). ||S1||1 is the sum of all elements in the matrix S1.

[0055] The Lagrange multiplier method (which can also be replaced by the iterative threshold method, the accelerated proximal gradient method, or the alternating direction method) is used to optimize the above formula to obtain the final matrix L1 and matrix S1.

[0056] The method for splitting the matrix X2 into L2 and S2 is the same as in the previous example.

[0057] The edges in the corresponding sub-graphs represented by matrix S1 and matrix S2 are potential noise edges, which need to be further confirmed in subsequent steps.

[0058] In the above example, both sub-graphs are undirected graphs, so the low-rank matrix obtained after denoising is also a mathematical expression of an undirected graph. Therefore, it is necessary to add constraints that the matrix L is equal to its transposed matrix and the matrix S is equal to its transposed matrix. If a sub-graph is a directed graph, there is no such constraint.

[0059] Finally, the sparse matrices are weighted summed. For example, α i is the weight of the sparse matrix Si corresponding to each relationship, representing the importance of the relationship for abnormal judgment; r represents the total number of relationship types. Generally,

[0060] For example, in Figure 2 In the example shown, S′=α1S1+α2S2. For example, let α1=α2=0.5 (i.e., the importance of the two types of relationships is equal), then

[0061] Set the screening threshold δ, when S′ ij When δ > δ, the edge connecting the i-th node and the j-th node is considered to be a noise edge, otherwise it is a normal edge. In the above formula, for example, if δ = 0.2, the edge connecting node 1 and node 4, node 3 and node 4 is a noise edge.

[0062] Two specific examples of noise merging are provided below.

[0063] In the Feima scenario, there is such a situation: a payment code of an offline merchant in Beijing is uploaded to the Internet. A consumer in Shanghai uses the payment code to make a purchase, which constitutes a Feima anomaly. In this scenario, the defined nodes are bank cards and merchants, and the defined relationships are the behavior of bank cards spending at merchants and the behavior of bank card-to-bank card transfers. The above two relationships can be represented as two adjacency matrices X1 and X2, respectively. The two adjacency matrices can be decomposed into X1=L1+S1 and X2=L2+S2 through the above method steps.

[0064] In this scenario, if there is no foreknowledge, we believe based on experience that consumption behavior is more important than transfer behavior, so we assign weights of 0.8 and 0.2 to each, respectively. Then the final combined noise matrix is ​​S′=0.8×S1+0.2×S2.

[0065] In this scenario, if some abnormal relationships are already known, let S′=α1×S1+(1-α1)×S2, and use the grid search method for α1 and the screening threshold δ to find the α1 and δ that make the final judgment result optimal.

[0066] In the telecom fraud scenario, the nodes are defined as individuals, and the relationships are defined as the transfer relationship and call relationship between people. The above two relationships can be represented as two adjacency matrices X3 and X4 respectively. Through the above method steps, these two adjacency matrices can be decomposed into X3=L3+S3 and X4=L4+S4.

[0067] In this scenario, if there is no foreknowledge, we believe based on experience that the transfer relationship is slightly more important than the call relationship, so we assign weights of 0.4 and 0.6 to the two respectively. Then the final combined noise matrix is ​​S'=0.4×S3+0.6×S4.

[0068] In this scenario, if some abnormal relationships have been predicted, let S″=α3×S3+(1-α3)×S4, and use the grid search method for the proportional coefficient α3 and the screening threshold δ to find the α3 and δ that make the final judgment result optimal.

[0069] Of course, the search method for the proportional coefficient and the screening threshold may also be other methods, such as the gradient descent method. Although it cannot ensure the best search results, at least relatively good search results can be obtained. Of course, the number of sub-graphs can be more than 2.

[0070] Based on the same technical concept, an embodiment of the present application also provides a noise recognition device for a knowledge graph, which is used to execute the noise recognition method for a knowledge graph provided by any of the above embodiments. Figure 3 A schematic diagram of the structure of a noise recognition device for a knowledge graph provided in an embodiment of the present application.

[0071] like Figure 3 As shown, the noise recognition device of the knowledge graph includes:

[0072] A graph splitting module 1 is used to split the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all the nodes of the knowledge graph;

[0073] Matrix decomposition module 2, used to convert each sub-graph into an adjacency matrix, and input each adjacency matrix into a preset robust principal component analysis model to obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship of the node numbers in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edges in the corresponding sub-graph;

[0074] The noise edge screening module 3 is used to perform weighted summation on all sparse matrices to obtain a noise matrix, and determine the noise edges in the knowledge graph according to the values ​​of each element in the noise matrix.

[0075] In some embodiments, the robust principal component analysis model is:

[0076]

[0077] Where X represents the adjacency matrix, L represents the low-rank matrix decomposed from the adjacency matrix X, S represents the sparse matrix decomposed from the adjacency matrix X, and the content after st represents the constraint condition, ||L|| * represents the nuclear norm of L, ||S||1 represents the L1 norm of S, λ is the penalty factor, It means that ||L|| * +λ||S||1 when L and S are minimized.

[0078] In some embodiments, when the sub-graph is an undirected graph, the robust principal component analysis model also has constraints: matrix L is equal to the transposed matrix of matrix L, and matrix S is equal to the transposed matrix of matrix S.

[0079] In some embodiments, the robust principal component analysis model uses a Lagrange multiplier method, an iterative threshold method, an accelerated proximal gradient method, or an alternating direction method to search for an optimal solution.

[0080] In some embodiments, the noise edge screening module 3 is specifically used to: perform weighted summation on all sparse matrices according to a preset proportional coefficient, and use the connection relationships corresponding to the elements in the obtained noise matrix whose values ​​are greater than a set screening threshold as noise edges.

[0081] In some embodiments, the noise edge screening module 3 is specifically used to: search for the proportional coefficient and the screening threshold so that the noise edge obtained according to the noise matrix is ​​consistent with the previously known noise edge, wherein in each search operation, the weighted summation of all sparse matrices is performed according to the current proportional coefficient, and the connection relationship corresponding to the elements in the current noise matrix whose values ​​are greater than the current screening threshold is taken as the noise edge.

[0082] It should be noted that the device in the embodiment of the present application can implement the various processes of the embodiment of the aforementioned method and achieve the same effects and functions, which will not be repeated here.

[0083] Figure 4 A noise recognition device for a knowledge graph according to an embodiment of the present application, for performing Figure 1 The noise identification method of the knowledge graph shown in the figure, the noise identification device of the knowledge graph includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute: the noise identification method of the aforementioned knowledge graph.

[0084] According to some embodiments of the present application, a non-volatile computer storage medium is provided, on which computer executable instructions are stored, and the computer executable instructions are configured to execute, when executed by a processor: the aforementioned noise identification method of the knowledge graph.

[0085] Each embodiment in the present application 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. In particular, for the device and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, their descriptions are simplified, and the relevant parts can be referred to the partial description of the method embodiments.

[0086] The device and computer-readable storage medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the device and computer-readable storage medium also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the device and computer-readable storage medium will not be repeated here.

[0087] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0088] The present application is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses, and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0089] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0091] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0092] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0093] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. In addition, although the operations of the method of the present application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple steps.

[0094] Although the spirit and principle of the present application have been described with reference to several specific embodiments, it should be understood that the present application is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present application is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A noise identification method for a knowledge graph, characterized in that: Applied to the flying code scenario or telecommunication fraud scenario, the method includes: The knowledge graph is split into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all the nodes of the knowledge graph; when the knowledge graph is applied to the Feima scenario, the nodes include bank cards and merchants, and the connection relationships include the behavior of bank card consumption at merchants and the transfer behavior between bank cards; when the knowledge graph is applied to the telecommunications fraud scenario, the nodes include individuals, and the connection relationships include the transfer relationship and call relationship between people; Convert each sub-graph into an adjacency matrix, input each adjacency matrix into a preset robust principal component analysis model, and obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship between the node numbers in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edges in the corresponding sub-graph; All sparse matrices are weightedly summed to obtain a noise matrix, and the noise edges in the knowledge graph are determined according to the values ​​of each element in the noise matrix. When applied to the flying code scenario, the noise edges in the knowledge graph are determined based on the weights of the sparse matrices corresponding to the two relationships of consumption behavior and transfer behavior; when applied to the telecommunications fraud scenario, the noise edges in the knowledge graph are determined based on the weights of the sparse matrices corresponding to the two relationships of transfer relationship and call relationship.

2. The method according to claim 1, characterized in that The robust principal component analysis model is: Where X represents the adjacency matrix, L represents the low-rank matrix decomposed from the adjacency matrix X, S represents the sparse matrix decomposed from the adjacency matrix X, and the content after st represents the constraint condition, ‖L‖ * represents the nuclear norm of L, ‖S‖1 represents the L1 norm of S, λ is the penalty factor, Denotes that ‖L‖ * +λ‖S‖1 when L and S are minimized.

3. The method according to claim 2, characterized in that The robust principal component analysis model uses Lagrange multiplier method, iterative threshold method, accelerated proximal gradient method, or alternating direction method to search for the optimal solution.

4. The method according to claim 2, characterized in that: When the sub-graph is an undirected graph, the robust principal component analysis model also has constraints: the matrix L is equal to its transposed matrix, and the matrix S is equal to its transposed matrix.

5. The method according to claim 1, characterized in that All sparse matrices are weighted summed to obtain a noise matrix, and the noise edges in the knowledge graph are determined according to the values ​​of each element in the noise matrix, including: All sparse matrices are weighted summed according to a preset proportional coefficient, and the connection relationships corresponding to the elements in the obtained noise matrix whose values ​​are greater than the set screening threshold are taken as noise edges.

6. The method according to claim 1, characterized in that All sparse matrices are weighted summed to obtain a noise matrix, and the noise edges in the knowledge graph are determined according to the values ​​of each element in the noise matrix, including: The proportional coefficient and the screening threshold are searched so that the noise edge obtained according to the noise matrix is ​​consistent with the previously known noise edge, wherein in each search operation, all sparse matrices are weighted summed according to the current proportional coefficient, and the connection relationship corresponding to the elements in the current noise matrix whose values ​​are greater than the current screening threshold is taken as the noise edge.

7. A noise recognition device for a knowledge graph, characterized in that: Applied to the scenario of flying code or telecommunication fraud, the device includes: A graph splitting module is used to split the knowledge graph into multiple sub-graphs, wherein the sub-graphs correspond to the relationship types of the connection relationships in the knowledge graph one by one, and the sub-graphs contain all the nodes of the knowledge graph; when the knowledge graph is applied to the Feima scenario, the nodes include bank cards and merchants, and the connection relationships include the behavior of bank card consumption at merchants and the transfer behavior between bank cards; when the knowledge graph is applied to the telecommunications fraud scenario, the nodes include individuals, and the connection relationships include the transfer relationship and call relationship between people; A matrix decomposition module, used to convert each sub-graph into an adjacency matrix, and input each adjacency matrix into a preset robust principal component analysis model to obtain a low-rank matrix and a sparse matrix corresponding to each adjacency matrix, wherein the coordinates of the elements in each adjacency matrix are consistent with the corresponding relationship between the node numbers in the knowledge graph, the low-rank matrix represents the correct connection relationship in the corresponding sub-graph, and the sparse matrix represents the potential noise edges in the corresponding sub-graph; The noise edge screening module is used to perform weighted summation on all sparse matrices to obtain a noise matrix, and determine the noise edges in the knowledge graph according to the values ​​of each element in the noise matrix. When applied to the flying code scenario, the noise edges in the knowledge graph are determined based on the weights of the sparse matrices corresponding to the two relationships of consumption behavior and transfer behavior; when applied to the telecommunications fraud scenario, the noise edges in the knowledge graph are determined based on the weights of the sparse matrices corresponding to the two relationships of transfer relationship and call relationship.

8. The device according to claim 7, characterized in that The robust principal component analysis model is: Where X represents the adjacency matrix, L represents the low-rank matrix decomposed from the adjacency matrix X, S represents the sparse matrix decomposed from the adjacency matrix X, and the content after st represents the constraint condition, ‖L‖ * represents the nuclear norm of L, ‖S‖1 represents the L1 norm of S, λ is the penalty factor, Denotes that ‖L‖ * +λ‖S‖1 when L and S are minimized.

9. The device according to claim 8, characterized in that The robust principal component analysis model uses Lagrange multiplier method, iterative threshold method, accelerated proximal gradient method, or alternating direction method to search for the optimal solution.

10. The device according to claim 8, characterized in that When the sub-graph is an undirected graph, the robust principal component analysis model also has constraints: the matrix L is equal to its transposed matrix, and the matrix S is equal to its transposed matrix.

11. The device according to claim 7, characterized in that The noise edge screening module is specifically used for: All sparse matrices are weighted summed according to a preset proportional coefficient, and the connection relationships corresponding to the elements in the obtained noise matrix whose values ​​are greater than the set screening threshold are taken as noise edges.

12. The device according to claim 7, characterized in that The noise edge screening module is specifically used for: The proportional coefficient and the screening threshold are searched so that the noise edge obtained according to the noise matrix is ​​consistent with the previously known noise edge, wherein in each search operation, all sparse matrices are weighted summed according to the current proportional coefficient, and the connection relationship corresponding to the elements in the current noise matrix whose values ​​are greater than the current screening threshold is taken as the noise edge.

13. A noise recognition device for a knowledge graph, characterized in that: include: at least one processor; And, a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor so that at least one processor can execute: the noise recognition method for the knowledge graph according to any one of claims 1 to 6.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, and when the program is executed by a processor, the processor executes: the noise recognition method for the knowledge graph according to any one of claims 1 to 6.

Images