Graph structure denoising method and device and storage medium

Abstract

The invention relates to a graph structure denoising method and device and a storage medium, and the method comprises the steps: correcting a first graph structure based on a graph structure learning model, and obtaining a second graph structure; processing the second graph structure based on a graph convolutional network model, and constructing a prediction confidence coefficient matrix of the nodes; generating a first mask matrix according to the predicted confidence coefficient matrix and a confidence coefficient threshold value, and adjusting the second graph structure by adopting the first mask matrix to obtain a third graph structure; retraining the graph structure learning model and the graph convolutional network model based on the third graph structure obtained after the confidence threshold is adjusted each time, and determining a target confidence threshold; generating a second mask matrix according to the target confidence threshold and the predicted confidence matrix; and adjusting the target graph structure by adopting the second mask matrix. According to the method and the device, connection (noise) is not additionally added to the adjacent matrix of the graph, and meanwhile, additional storage overhead is also prevented from being added to the adjacent matrix.

Description

Technical Field

[0001] The present application relates to the technical field of graph data processing, and in particular to a graph structure denoising method, device and storage medium. Background Art

[0002] Graph structure data is widely present in social networks and recommendation systems, and graph convolutional network models can be used to quickly perform user classification, community detection, and product recommendations. Before the graph structure is input into the graph convolutional network model, the graph structure is usually denoised, that is, the structured noise existing in social networks and recommendation systems is eliminated. The graph structure denoising method of the related art is based on the homogeneity assumption, that is, nodes with similar features are nodes of the same type. Specifically, it uses a variety of training methods, including co-training (including GRCN, SLAPS), iterative training (iter-training, including IDGL) and pre-training (including SUBLIME) to calculate the feature similarity matrix of the nodes and improve the quality of the graph structure data. However, this type of graph structure denoising method has a large storage overhead, and the denoising process will introduce new noise, resulting in a decrease in the prediction accuracy of the subsequent graph convolutional network model when aggregating nodes.

[0003] There is currently no effective solution to the problems of high storage overhead and introduction of new noise in graph structure denoising methods in related technologies. Summary of the invention

[0004] Based on this, it is necessary to provide a graph structure denoising method, device and storage medium that can reduce storage overhead and does not introduce new noise in order to address the above technical problems.

[0005] In a first aspect, the present application provides a graph structure denoising method, the method comprising:

[0006] The first graph structure is modified based on the graph structure learning model to obtain a second graph structure;

[0007] Processing the second graph structure based on a graph convolutional network model to construct a prediction confidence matrix of the node;

[0008] Generate a first mask matrix according to the prediction confidence matrix and the confidence threshold, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure;

[0009] Based on the third graph structure obtained after each adjustment of the confidence threshold, retraining the graph structure learning model and the graph convolutional network model to determine a target confidence threshold;

[0010] Generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix;

[0011] The target graph structure is adjusted using the second mask matrix.

[0012] In one embodiment, modifying the first graph structure based on the graph structure learning model includes:

[0013] Inputting the first graph structure and the corresponding feature matrix into the graph structure learning model, and calculating the similarity between nodes in the first graph structure;

[0014] Increase or decrease the connection between nodes according to the similarity between the nodes.

[0015] In one embodiment, the second graph structure is processed based on a graph convolutional network model to construct a prediction confidence matrix of a node, including:

[0016] Inputting the second graph structure and the corresponding feature matrix into the graph convolutional network model to obtain the predicted information entropy of the node;

[0017] The prediction confidence matrix is ​​obtained according to the prediction information entropy of the node.

[0018] In one embodiment, the prediction confidence matrix is ​​obtained according to the prediction information entropy of the node, including:

[0019] Normalizing the predicted information entropy to a confidence value between 0 and 1;

[0020] The prediction confidence matrix is ​​obtained according to the confidence values ​​after standardization.

[0021] In one embodiment, generating a first mask matrix according to the prediction confidence matrix and the confidence threshold includes:

[0022] The prediction confidence matrix is ​​input into a RAD function to generate the first mask matrix; wherein the RAD function is constructed based on the confidence threshold and an activation function.

[0023] In one embodiment, the formula of the RAD function is as follows:

[0024]

[0025] Wherein, M represents the first mask matrix, φ(·) represents the activation function, C represents the prediction confidence matrix, Δ represents the confidence threshold, and e represents a natural constant.

[0026] In one embodiment, based on the third graph structure obtained after adjusting the confidence threshold each time, retraining the graph structure learning model and the graph convolutional network model includes:

[0027] In the current training, the third graph structure generated in the previous training is used as the first graph structure of the current training, and the current first graph structure is input into the graph structure learning model for correction to obtain the current second graph structure;

[0028] Generate a current mask matrix according to the prediction confidence matrix and the currently adjusted confidence threshold;

[0029] Adjusting the current second graph structure according to the current mask matrix to obtain the current third graph structure;

[0030] The current third graph structure is input into the graph convolutional network model to obtain the classification prediction value of the node.

[0031] In one embodiment, based on the third graph structure obtained after each adjustment of the confidence threshold, retraining the graph structure learning model and the graph convolutional network model to determine the target confidence threshold includes:

[0032] Compare the classification prediction value of each training with the classification label of the node to obtain a classification error;

[0033] When the retraining reaches a preset number of training times, the minimum classification error is determined, and the confidence threshold corresponding to the minimum classification error is used as the target confidence threshold.

[0034] In one embodiment, before inputting the first graph structure into the graph structure learning model for correction, the method further includes:

[0035] Acquiring interaction data, wherein the interaction data includes identity identification information of an interaction object;

[0036] Elements for graph structure learning are constructed with the interactive objects as nodes, wherein the elements include an adjacency matrix constituting the first graph structure.

[0037] In a second aspect, this embodiment provides a graph structure denoising device, including:

[0038] A correction module, used for correcting the first graph structure based on the graph structure learning model to obtain a second graph structure;

[0039] A construction module, used to process the second graph structure based on a graph convolutional network model to construct a prediction confidence matrix of the node;

[0040] A first adjustment module, configured to generate a first mask matrix according to the prediction confidence matrix and the confidence threshold Δ, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure;

[0041] A retraining module, configured to retrain the graph structure learning model and the graph convolutional network model based on a third graph structure obtained after each adjustment of the confidence threshold, and determine a target confidence threshold;

[0042] A generating module, configured to generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix;

[0043] The second adjustment module is used to adjust the target graph structure by using the second mask matrix.

[0044] In a third aspect, this embodiment provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps of the method described in the first aspect above.

[0045] In a fourth aspect, this embodiment provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect are performed.

[0046] The above-mentioned graph structure denoising method, device and storage medium construct the prediction confidence matrix of the node, optimize the confidence threshold through model retraining, determine the optimal mask matrix according to the prediction confidence matrix of the node and the determined optimal confidence threshold, and then adjust the graph structure based on the optimal mask matrix to obtain the final denoised graph structure. Among them, the mask matrix takes into account the prediction confidence of the node, that is, the prediction confidence of the node is incorporated into the graph structure denoising process. The generation process of the mask matrix is ​​lightweight and plug-and-play, and can be directly used in any graph structure learning algorithm. It will not add additional connections (noise) to the adjacency matrix, but directly modify the existing edge weights, while also avoiding adding additional storage overhead to the adjacency matrix. Only a small number of optimization parameters can significantly improve the prediction effect of the graph convolutional network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 FIG. 1 is an application environment diagram of a graph structure denoising method in an embodiment;

[0048] Figure 2 A schematic diagram of a flow chart of a method for denoising a graph structure in one embodiment;

[0049] Figure 3 A schematic diagram of a flow chart of a method for denoising a graph structure in another embodiment;

[0050] Figure 4 It is a structural schematic diagram of a denoising device of a middle structure in one embodiment;

[0051] Figure 5 is a hardware structure block diagram of an electronic device in one embodiment;

[0052] Figure 6 FIG. 4 is a block diagram of the hardware structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the", "these" and the like in this application do not represent quantitative restrictions, and they can be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether directly or indirectly. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. Usually, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0055] In one embodiment, the graph structure denoising method can be applied to Figure 1In the application environment shown. The application environment is a recommendation system, which includes a terminal 101, a server 102 and a data storage system 103. Among them, the terminal 101 communicates with the server 102 through the network. The data storage system 103 can store the data that the server 102 needs to process. The data storage system 103 can be integrated on the server 102, or it can be placed on the cloud or other network servers. Among them, the terminal 101 can be but is not limited to various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 102 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0056] The server is deployed with a recommendation engine, a data collection layer, and a business logic layer. The recommendation engine is used to generate a personalized recommendation list based on the user's browsing history, purchase history, click behavior and other information. The data collection layer is used to collect the user's real-time behavior data and other related information, such as clicks on recommended content, likes, comments, shares, geographic location, device type, etc. These data will be transmitted to the recommendation engine, business logic layer and data storage system. The recommendation engine can perform instant calculations or offline analysis on user data. The business logic layer can process user data, such as updating long-term models such as user portraits and product features. The data storage system can clean, parse, and format user data for storage. The graph structure denoising method can be executed in the server, and the server can convert user data into a graph structure and denoise the graph structure.

[0057] The graph structure learning algorithm of the related art adds node connections to the original graph by calculating the feature similarity matrix. This method greatly increases the storage overhead of the adjacency matrix and limits the computing performance of the model. The graph structure learning algorithm of the related art only considers feature similarity and ignores the prediction confidence of the node. During the aggregation process, low-confidence neighbor nodes will cause a large deviation in node embedding, thereby introducing new noise and reducing the accuracy of the prediction.

[0058] Based on the research and discovery of the above problems, a graph structure denoising method is provided in this embodiment. Figure 2 is a flow chart of the graph structure denoising method in this embodiment. Figure 2 As shown, the method comprises the following steps:

[0059] Step S101, modifying the first graph structure A based on the graph structure learning model to obtain a second graph structure A*.

[0060] Graph structure: A data structure consisting of a set of nodes and edges between nodes. The first graph structure A is represented and stored in the computer in the form of an adjacency matrix, which reflects the connection information between nodes. When using a recommendation system, interaction data (or historical interaction data) can be collected from the system. The interaction data includes the identity information of the interaction object; the interaction object is used as a node to construct elements for graph structure learning. The elements include the adjacency matrix that constitutes the first graph structure. In addition, the elements can also include a node set, an edge set, and a feature matrix. Specifically, the graph G = (V, E, A, X), where V is a set of N nodes, E is an edge set, A∈R^(N×N) is an adjacency matrix, and X∈R^(N×d) is a node feature matrix with a dimension of d.

[0061] Graph structure learning model: a type of machine learning method (also called graph structure learning algorithm, graph structure learner) used to learn useful patterns or structures from graph structure data, such as the relationship between nodes, community structure, etc. In this step, the modified graph structure is to remove the noise connection information in the graph structure, and the graph structure learning model can be used to generate the modified graph structure. For example, the first graph structure A and the corresponding feature matrix X are input into the graph structure learning model, and the similarity between the nodes in the first graph structure A is calculated; the connection between the nodes is increased or decreased according to the similarity between the nodes.

[0062] Step S102: Process the second graph structure based on the graph convolutional network model to construct a prediction confidence matrix for the node.

[0063] Graph Convolutional Networks (GCNs): It is a specific type of graph structure learning model. It is a deep learning model based on graph structure data, which is used to process tasks such as node classification and link prediction in graph structure data. In this step, the second graph structure A* and the corresponding feature matrix X can be input into the graph convolutional network model to obtain the predicted information entropy S of the node; according to the predicted information entropy S of the node, the predicted confidence matrix C is obtained. Among them, the graph convolutional network model includes an encoder and a classifier, wherein the encoder is used to extract the feature vector of the node, and the classifier is used to classify the feature vector of the node. When the second graph structure and the corresponding feature matrix are input into the graph convolutional network model, the feature vector of the node is extracted by the encoder, and the feature vector is processed by the classifier to obtain the predicted probability vector for the label. Based on the predicted probability vector, the predicted information entropy S, that is, the Shannon information entropy, can be calculated, and the predicted confidence matrix C can be calculated based on the predicted information entropy S. Specifically, the predicted information entropy S is standardized to a confidence value between 0 and 1; according to the confidence value after standardization, the predicted confidence matrix C is obtained. The formula is as follows:

[0064] C=e -S

[0065] Step S103: Generate a first mask matrix M1 according to the prediction confidence matrix C and the confidence threshold Δ, and use the first mask matrix M1 to adjust the second graph structure A* to obtain a third graph structure

[0066] The first mask matrix M1 can change the connection information of the graph structure. According to the prediction confidence matrix C and the confidence threshold Δ, the first mask matrix M1 is generated. This process can be achieved by designing a function, which is temporarily defined as the RAD (Reliable Aggregation Denoising) function. Specifically, the RAD function can be constructed based on the confidence threshold Δ and the activation function φ(·), and the prediction confidence matrix C is input into RAD to generate the first mask matrix M1. The RAD function is as follows:

[0067]

[0068] Wherein, e represents a natural constant. In this embodiment, the confidence threshold Δ∈R N×1 , which can be continuously optimized through retraining, and is used to divide node neighbors into high-confidence neighbors and low-confidence neighbors, that is, those above the confidence threshold are trusted neighbors, and those below the confidence threshold are untrusted neighbors. The activation function φ(·) is used to increase the connection weights with high-confidence nodes and weaken the influence of low-confidence nodes on embedding. In this embodiment, the RAD function takes the adjacency matrix storing user interaction information as input to generate a high-confidence denoised adjacency matrix, which can generate a mask matrix based on the confidence of the node, and the mask matrix can change the connection information of the graph structure.

[0069] In this embodiment, the RAD function (mask matrix generation process) is lightweight and plug-and-play, and can be directly used in any graph structure learning algorithm. It does not add additional connections (noise) to the adjacency matrix, but directly modifies the existing edge weights to avoid adding additional storage overhead to the adjacency matrix. Only a small number of optimization parameters can significantly improve the prediction effect of the graph convolutional network model.

[0070] Step S104: based on the third graph structure obtained after each adjustment of the confidence threshold, retrain the graph structure learning model and the graph convolutional network model to determine the target confidence threshold.

[0071] Since the change of the mask matrix M will affect the denoising effect on the graph structure, and the mask matrix M is related to the confidence threshold Δ, which can be learned and optimized, the confidence threshold Δ can be optimized through retraining, thereby optimizing the mask matrix M, and finally using the optimal mask matrix to adjust the graph structure to obtain a better denoising effect. In this step, retraining involves multiple trainings, each of which includes training the graph structure learning model and the graph convolutional network model, that is, each training includes the correction of the graph structure and the mask adjustment. First, the graph structure learning model is used to calculate the feature similarity between nodes to increase or decrease the connection to obtain the corrected graph structure; then the mask matrix is ​​generated to adjust the corrected graph structure. These two steps together constitute one training.

[0072] Step S105: Generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix.

[0073] After determining the target confidence threshold, the second mask matrix M2 is generated according to the target confidence threshold and the prediction confidence matrix C, that is, the optimal mask matrix is ​​obtained, which is also a denoising adjacency matrix with high confidence.

[0074] Step S106: Use the second mask matrix to adjust the target graph structure.

[0075] By multiplying the second mask matrix M2 and the adjacency matrix of the target graph structure, the structural noise in the graph structure (adjacency matrix) can be reduced to obtain the final denoised graph structure, thereby effectively improving the prediction accuracy of the graph convolutional network model for downstream tasks.

[0076] In the above steps S101 to S106, the prediction confidence matrix of the node is constructed, and the confidence threshold is optimized by retraining the model. The optimal mask matrix is ​​determined according to the prediction confidence matrix of the node and the determined optimal confidence threshold, and then the graph structure is adjusted based on the optimal mask matrix to obtain the final denoised graph structure. Among them, the mask matrix takes into account the prediction confidence of the node, that is, the prediction confidence of the node is included in the graph structure denoising process. The generation process of the mask matrix is ​​lightweight and plug-and-play, and can be directly used in any graph structure learning algorithm. It will not add additional connections (noise) to the adjacency matrix, but directly modify the existing edge weights, and also avoid adding additional storage overhead to the adjacency matrix. Only a small number of optimization parameters can significantly improve the prediction effect of the graph convolutional network model.

[0077] Although the above embodiment introduces a graph structure denoising method for a recommendation system, based on the same inventive concept, the graph structure denoising method can also be applied to other systems, such as a social network.

[0078] In one embodiment, based on the third graph structure obtained after adjusting the confidence threshold each time, retraining the graph structure learning model and the graph convolution network model includes the following steps:

[0079] In the current training, the third graph structure generated in the previous training is used as the first graph structure of the current training, and the current first graph structure is input into the graph structure learning model for correction to obtain the current second graph structure;

[0080] Generate a current mask matrix according to the prediction confidence matrix and the currently adjusted confidence threshold;

[0081] Adjusting the current second graph structure according to the current mask matrix to obtain the current third graph structure;

[0082] The current third graph structure is input into the graph convolutional network model to obtain the classification prediction value of the node.

[0083] In this embodiment, a graph structure learning model is used to calculate the feature similarity between nodes to increase or decrease the connection, thereby obtaining a modified graph structure; and a mask matrix is ​​generated to adjust the obtained modified graph structure. These two steps together constitute one training, that is, each training includes the modification of the graph structure and the mask adjustment.

[0084] Furthermore, the classification prediction value of each training is compared with the classification label of the node to obtain the classification error; when the retraining reaches the preset number of training times, the minimum classification error is determined, and the confidence threshold corresponding to the minimum classification error is used as the target confidence threshold.

[0085] For easier understanding, the retraining process is further introduced below.

[0086] First training: Initialize the first graph structure A, modify the first graph structure A (update the connection of the graph by calculating the feature similarity between nodes) to obtain the second graph structure A*. Adjust the confidence threshold Δ in the RAD function, generate the mask matrix M based on the adjusted confidence threshold Δ to adjust the modified second graph structure A*, and obtain the third graph structure use It is trained as the input of the graph convolutional network model to obtain the classification prediction value, compare the error between the classification prediction value and the classification label, and obtain the classification error.

[0087] Subsequent training: Repeat the above process, and adjust the confidence threshold Δ to generate a mask matrix based on the results of the previous training in each iteration, and then adjust the graph structure. After each round of training, the classification error can be calculated by comparing the model's classification prediction value with the actual classification label. Then adjust the confidence threshold Δ based on the classification error, generate a new mask matrix, and then adjust the graph structure, so that it can better fit the data in the next iteration.

[0088] Termination condition: When the preset number of training times N is reached, the result with the smallest classification error is selected from the classification results obtained in all training processes, that is, the optimal confidence threshold is found.

[0089] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0090] In one embodiment, another graph structure denoising method is provided. Figure 3 This is the flowchart of the graph structure denoising method, as shown in Figure 3 As shown, the process includes the following steps:

[0091] Step S301, collect interaction data in the application system and construct a graph structure. The application system can be a social network system, a recommendation system, etc. Use the interaction data (or historical interaction data) collected in the social network or recommendation system, the interaction data includes the identity information of the interaction object; use the interaction object as a node to construct elements for graph structure learning, and the elements include an adjacency matrix that constitutes the first graph structure. In addition, the elements may also include a node set, an edge set, and a feature matrix. Specifically, the graph G = (V, E, A, X), where V is a set of N nodes, E is an edge set, A∈R^(N×N) is an adjacency matrix, and X∈R^(N×d) is a node feature matrix with a dimension of d.

[0092] Step S302, based on the graph structure learning model, the first graph structure A is modified to obtain a second graph structure A*, the second graph structure A* is processed based on the graph convolutional network model, and the prediction confidence matrix C of the node is constructed. Specifically, firstly, the original first graph structure A and the feature matrix X are input into the graph structure model to obtain the modified second graph structure A*, and then the second graph structure A* and the feature matrix X are input into the graph convolutional network model to obtain the classification probability of the node and calculate the information entropy matrix S accordingly, and the prediction confidence matrix C of the node is constructed based on the predicted information entropy S of the node.

[0093] Step S303, use RAD to retrain the graph structure learning model and the graph convolution network model to determine the optimal mask matrix. Construct the RAD function based on the confidence threshold Δ and the activation function φ(·), input the prediction confidence matrix C into RAD, and generate the mask matrix M. Optimize the confidence threshold Δ by retraining the graph structure learning model and the graph convolution network model to obtain the optimal mask matrix.

[0094] Step S304, adjusting the target graph structure based on the optimal mask matrix to obtain a denoised graph structure.

[0095] Step S305: input the denoised graph structure into the graph convolutional network model, learn the feature vectors of the nodes and perform classification prediction.

[0096] Table 1 shows the accuracy experimental results of the graph structure denoising method provided in this embodiment. This embodiment uses multiple node classification task data sets, including academic network data sets Cora, Citeseer and social network data set Blogcatalog, and tests the ACC index performance improvement of RAD relative to different graph structure learning algorithms (including GRCN, SLAPS, SUBLIME) based on these data sets. It can be seen from the experimental results that when RAD is combined with each graph structure learning algorithm, the ACC index performance is improved compared to using each graph structure learning algorithm alone. Therefore, RAD can effectively improve the prediction accuracy of different graph structure learning algorithms on the tested data sets.

[0097] Table 1 Accuracy experiment

[0098]

[0099] Based on the same inventive concept, the embodiment of the present application also provides a graph structure denoising device for implementing the above method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more graph structure denoising device embodiments provided below can refer to the limitations of the graph structure denoising method above, and will not be repeated here.

[0100] Figure 4 is a schematic diagram of the structure of the graph structure denoising device of this embodiment, as shown in Figure 4 As shown, the device includes: a correction module, a construction module, a first adjustment module, a retraining module, a generation module and a second adjustment module.

[0101] A correction module, used for correcting the first graph structure based on the graph structure learning model to obtain a second graph structure;

[0102] A first construction module is used to process the second graph structure based on a graph convolutional network model to construct a prediction confidence matrix of the node;

[0103] A first adjustment module is used to generate a first mask matrix according to the prediction confidence matrix and the confidence threshold, and use the mask matrix to adjust the second graph structure to obtain a third graph structure;

[0104] A retraining module, used for retraining the graph structure learning model and the graph convolution network model based on the third graph structure obtained after each adjustment of the confidence threshold, to determine the target confidence threshold;

[0105] A generating module, used for generating a second mask matrix according to a target confidence threshold and a prediction confidence matrix;

[0106] The second adjustment module is used to adjust the target graph structure by using a second mask matrix.

[0107] In one embodiment, the correction module includes:

[0108] A first input unit, used to input the first graph structure and the corresponding feature matrix into the graph structure learning model, and calculate the similarity between nodes in the first graph structure;

[0109] The connection unit is used to increase or decrease the connection between nodes according to the similarity between the nodes.

[0110] In one embodiment, the building blocks include:

[0111] A second input unit is used to input the second graph structure and the corresponding feature matrix into the graph convolutional network model to obtain the predicted information entropy of the node;

[0112] The calculation unit is used to obtain a prediction confidence matrix according to the prediction information entropy of the node.

[0113] In one embodiment, the computing unit comprises:

[0114] Normalization unit, used to normalize the prediction information entropy to a confidence value between 0 and 1;

[0115] The first determining unit is used to obtain a prediction confidence matrix according to the standardized confidence values.

[0116] In one embodiment, the first adjustment module includes:

[0117] The third input unit is used to input the prediction confidence matrix into the RAD function to generate a first mask matrix; wherein the RAD function is constructed based on the confidence threshold and the activation function.

[0118] In one embodiment, the formula of the RAD function is as follows:

[0119]

[0120] Among them, M represents the first mask matrix, φ(·) represents the activation function, C represents the prediction confidence matrix, Δ represents the confidence threshold, and e represents the natural constant.

[0121] In one embodiment, the retraining module includes:

[0122] A correction unit is used to use the third graph structure generated in the previous training as the first graph structure of the current training in the current training, input the current first graph structure into the graph structure learning model for correction, and obtain the current second graph structure;

[0123] A generating unit, used for generating a current mask matrix according to the prediction confidence matrix and the currently adjusted confidence threshold;

[0124] An adjustment unit, used for adjusting the current second graph structure according to the current mask matrix to obtain the current third graph structure;

[0125] The fourth input unit is used to input the current third graph structure into the graph convolutional network model to obtain the classification prediction value of the node.

[0126] In one embodiment, the retraining module further includes:

[0127] A comparison unit is used to compare the classification prediction value of each training with the classification label of the node to obtain the classification error;

[0128] The second determination unit is used to determine the minimum classification error when the retraining reaches a preset number of training times, and use the confidence threshold corresponding to the minimum classification error as the target confidence threshold.

[0129] In one embodiment, the apparatus further comprises:

[0130] An acquisition module, used to acquire interaction data, the interaction data including identity identification information of the interaction object;

[0131] The second construction module is used to construct elements for graph structure learning using interactive objects as nodes, where the elements include an adjacency matrix constituting the first graph structure.

[0132] Each module in the above device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module above.

[0133] In one embodiment, an electronic device is also provided in this embodiment, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0134] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0135] Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0136] The first graph structure is modified based on the graph structure learning model to obtain a second graph structure;

[0137] The second graph structure is processed based on the graph convolutional network model to construct a prediction confidence matrix of the node;

[0138] Generate a first mask matrix according to the prediction confidence matrix and the confidence threshold, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure;

[0139] Based on the third graph structure obtained after each adjustment of the confidence threshold, the graph structure learning model and the graph convolutional network model are retrained to determine the target confidence threshold;

[0140] Generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix;

[0141] The second mask matrix is ​​used to adjust the target graph structure.

[0142] Figure 5 is the hardware structure block diagram of the electronic device. Figure 5 As shown, the electronic device may include one or more ( Figure 5 Only one is shown in the figure) processor 501 and memory 502 for storing data, wherein processor 501 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA. The above electronic device may also include a transmission device 503 and an input and output device 504 for communication functions. It can be understood by those skilled in the art that Figure 5 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 5 More or fewer components as shown, or with Figure 5 Different configurations shown.

[0143] The memory 502 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the graph structure denoising method in this embodiment. The processor 501 executes various functional applications and data processing by running the computer program stored in the memory 502, that is, to implement the above method. The memory 502 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 502 may further include a memory remotely arranged relative to the processor 501, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0144] The transmission device 503 is used to receive or send data via a network. The above network includes a wireless network provided by a communication provider of the electronic device. In one example, the transmission device 503 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 503 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

[0145] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and will not be repeated in this embodiment.

[0146] In addition, in combination with the graph structure denoising method provided in the above embodiments, a storage medium may be provided in this embodiment to implement the method. The storage medium stores a computer program; when the computer program is executed by a processor, any graph structure denoising method in the above embodiments is implemented.

[0147] Optionally, in this embodiment, when the computer program is executed by the processor, the following steps are performed:

[0148] The first graph structure is modified based on the graph structure learning model to obtain a second graph structure;

[0149] The second graph structure is processed based on the graph convolutional network model to construct a prediction confidence matrix of the node;

[0150] Generate a first mask matrix according to the prediction confidence matrix and the confidence threshold, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure;

[0151] Based on the third graph structure obtained after each adjustment of the confidence threshold, the graph structure learning model and the graph convolutional network model are retrained to determine the target confidence threshold;

[0152] Generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix;

[0153] The second mask matrix is ​​used to adjust the target graph structure.

[0154] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store graph structure data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a graph structure denoising method is implemented.

[0155] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0156] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0157] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0158] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A graph structure denoising method, characterized in that: include: The first graph structure is modified based on the graph structure learning model to obtain a second graph structure; Processing the second graph structure based on a graph convolutional network model to construct a prediction confidence matrix of the node; Generate a first mask matrix according to the prediction confidence matrix and the confidence threshold, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure; Based on the third graph structure obtained after each adjustment of the confidence threshold, retraining the graph structure learning model and the graph convolutional network model to determine a target confidence threshold; Generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix; The target graph structure is adjusted using the second mask matrix.

2. The graph structure denoising method according to claim 1, characterized in that: The first graph structure is modified based on the graph structure learning model, including: Inputting the first graph structure and the corresponding feature matrix into the graph structure learning model, and calculating the similarity between nodes in the first graph structure; Increase or decrease the connection between nodes according to the similarity between the nodes.

3. The graph structure denoising method according to claim 1, characterized in that: The second graph structure is processed based on a graph convolutional network model to construct a prediction confidence matrix of the node, including: Inputting the second graph structure and the corresponding feature matrix into the graph convolutional network model to obtain the predicted information entropy of the node; The prediction confidence matrix is ​​obtained according to the prediction information entropy of the node.

4. The graph structure denoising method according to claim 3, characterized in that: According to the prediction information entropy of the node, the prediction confidence matrix is ​​obtained, including: Normalizing the predicted information entropy to a confidence value between 0 and 1; The prediction confidence matrix is ​​obtained according to the confidence values ​​after standardization.

5. The graph structure denoising method according to claim 1, characterized in that: Generating a first mask matrix according to the prediction confidence matrix and the confidence threshold includes: The prediction confidence matrix is ​​input into a RAD function to generate the first mask matrix; wherein the RAD function is constructed based on the confidence threshold and an activation function.

6. The graph structure denoising method according to claim 5, characterized in that: The formula of the RAD function is as follows: Wherein, M represents the first mask matrix, φ(·) represents the activation function, C represents the prediction confidence matrix, Δ represents the confidence threshold, and e represents a natural constant.

7. The graph structure denoising method according to claim 1, characterized in that: Based on the third graph structure obtained after adjusting the confidence threshold each time, retraining the graph structure learning model and the graph convolution network model, including: In the current training, the third graph structure generated in the previous training is used as the first graph structure of the current training, and the current first graph structure is input into the graph structure learning model for correction to obtain the current second graph structure; Generate a current mask matrix according to the prediction confidence matrix and the currently adjusted confidence threshold; Adjusting the current second graph structure according to the current mask matrix to obtain the current third graph structure; The current third graph structure is input into the graph convolutional network model to obtain the classification prediction value of the node.

8. The graph structure denoising method according to claim 7, characterized in that: Based on the third graph structure obtained after each adjustment of the confidence threshold, retraining the graph structure learning model and the graph convolution network model to determine the target confidence threshold, including: Compare the classification prediction value of each training with the classification label of the node to obtain a classification error; When the retraining reaches a preset number of training times, the minimum classification error is determined, and the confidence threshold corresponding to the minimum classification error is used as the target confidence threshold.

9. The graph structure denoising method according to claim 1, characterized in that: Before inputting the first graph structure into the graph structure learning model for correction, the method further includes: Acquiring interaction data, wherein the interaction data includes identity identification information of an interaction object; Elements for graph structure learning are constructed with the interactive objects as nodes, wherein the elements include an adjacency matrix constituting the first graph structure.

10. A graph structure denoising device, characterized in that: include: A correction module, used for correcting the first graph structure based on the graph structure learning model to obtain a second graph structure; A construction module, used to process the second graph structure based on a graph convolutional network model to construct a prediction confidence matrix of the node; A first adjustment module, configured to generate a first mask matrix according to the prediction confidence matrix and the confidence threshold Δ, and use the first mask matrix to adjust the second graph structure to obtain a third graph structure; A retraining module, configured to retrain the graph structure learning model and the graph convolutional network model based on a third graph structure obtained after each adjustment of the confidence threshold, and determine a target confidence threshold; A generating module, configured to generate a second mask matrix according to the target confidence threshold and the prediction confidence matrix; The second adjustment module is used to adjust the target graph structure by using the second mask matrix.

11. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: The processor is configured to run the computer program to perform the steps of the method of any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images