A data processing method, device and equipment

Abstract

The embodiments of this specification disclose a data processing method, apparatus and equipment, which include: obtaining graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features; locating the privacy data contained in the graph structure data according to a preset privacy data forgetting requirement, and determining the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the locating result; adjusting the data in the target subgraph data to generate the hedge subgraph data corresponding to the target subgraph data through a hedging rule that matches the type corresponding to the privacy data forgetting requirement; and training the graph neural network model through comparative learning based on the target subgraph data and obtaining the trained graph neural network model.

Description

Technical Field

[0001] This document relates to the field of computer technology, and in particular to a data processing method, device and equipment. Background Art

[0002] The Graph Neural Network (GNN) model is a new machine learning model proposed to utilize the rich information contained in graph structured data (data with a discrete structure consisting of vertices and edges connecting the vertices), and it has great prospects.

[0003] Graph neural network models can convert graph structure data into low-dimensional vectors by aggregating the feature information of adjacent nodes. Similar to other machine learning models, graph neural network models can be trained on graph structure data of private data such as traffic interaction networks, where the data subject may need to forget some private data. Common methods for forgetting private data include grouping the data and then retraining the model. However, the training of graph neural network models depends on the structural information in the graph structure data, and the random division of subgraphs may seriously damage the practicality of the model. Therefore, new methods are urgently needed to eliminate the impact of private data that needs to be forgotten on graph neural network models. To this end, it is necessary to provide a private data forgetting method for graph neural network models that is independent of the model, has low training cost, and has little impact on the model. Summary of the invention

[0004] The purpose of the embodiments of this specification is to provide a privacy data forgetting method for a graph neural network model that is independent of the model, has low training cost, and has little impact on the model.

[0005] In order to implement the above technical solution, the embodiments of this specification are implemented as follows:

[0006] A data processing method provided in an embodiment of the present specification comprises: obtaining graph structure data for model training of a graph neural network model, wherein the graph structure data comprises nodes, edges and node features. According to a pre-set privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined according to the location result. By adjusting the data in the target subgraph data through a hedging rule that matches the type corresponding to the privacy data forgetting requirement, the hedging subgraph data corresponding to the target subgraph data is generated. Based on the target subgraph data and the hedging subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0007] A data processing device provided in an embodiment of the present specification comprises: a data acquisition module, which acquires graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features. A target subgraph determination module, which locates the privacy data contained in the graph structure data according to a preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. A hedge subgraph determination module, which adjusts the data in the target subgraph data to generate hedge subgraph data corresponding to the target subgraph data by using a hedging rule that matches the type corresponding to the privacy data forgetting requirement. A model training module, which trains the graph neural network model by comparative learning based on the target subgraph data and the hedge subgraph data to obtain a trained graph neural network model.

[0008] A data processing device provided in an embodiment of the present specification includes: a processor; and a memory arranged to store computer executable instructions, wherein when the executable instructions are executed, the processor: obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features. According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined according to the location result. Through a hedging rule that matches the type corresponding to the privacy data forgetting requirement, the data in the target subgraph data is adjusted to generate the hedge subgraph data corresponding to the target subgraph data. Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0009] The embodiments of this specification also provide a storage medium, which is used to store computer executable instructions, and the executable instructions implement the following process when executed by the processor: obtain graph structure data for model training of the graph neural network model, and the graph structure data includes nodes, edges and node features. According to the preset privacy data forgetting requirements, the privacy data contained in the graph structure data is located, and the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined according to the positioning result. Through the hedging rules that match the type corresponding to the privacy data forgetting requirements, the data in the target subgraph data is adjusted to generate the hedge subgraph data corresponding to the target subgraph data. Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain the trained graph neural network model.

[0010] The embodiments of this specification also provide a computer program product, including a computer program, which implements the following process when executed by a processor: obtaining graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features. According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined according to the location result. Through a hedging rule that matches the type corresponding to the privacy data forgetting requirement, the data in the target subgraph data is adjusted to generate the hedge subgraph data corresponding to the target subgraph data. Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings required for use in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative labor.

[0012] Figure 1 This is an embodiment of a data processing method of this specification;

[0013] Figure 2 A schematic diagram of a data processing process of this specification;

[0014] Figure 3 Another data processing method embodiment of this specification;

[0015] Figure 4 This is another data processing method embodiment of the present specification;

[0016] Figure 5 This is another data processing method embodiment of the present specification;

[0017] Figure 6 This is another data processing method embodiment of the present specification;

[0018] Figure 7 This is an embodiment of a data processing device of the present specification;

[0019] Figure 8 This is an embodiment of a data processing device in this specification. DETAILED DESCRIPTION

[0020] The embodiments of this specification provide a data processing method, device and equipment.

[0021] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0022] The embodiments of this specification provide a privacy data forgetting solution for graph neural network models. The current large models, deep learning and machine learning methods use a large amount of user information for training, and then provide personalized results for users in specific task scenarios. However, there is often privacy data in user information, and for models trained using user privacy data, attackers can steal the privacy data through deliberately designed attack methods, or can make the trained model exhibit certain dangerous behaviors when performing reasoning. In order to meet the requirements of privacy protection regulations, artificial intelligence models constructed by large models, deep learning and machine learning need to be forgotten (also called machine forgetting (MU)) according to user needs, that is, the influence of certain user data (especially privacy data) on the artificial intelligence model (the influence here refers to the changes brought to the artificial intelligence model by these privacy data, including positive and negative effects) is eliminated.

[0023] The Graph Neural Network (GNN) model is a new machine learning model proposed to utilize the rich information contained in graph structured data (data with discrete structures consisting of vertices and edges connecting vertices), and it has great prospects. The core idea of ​​the graph neural network model is to convert graph structured data into low-dimensional vectors by aggregating the feature information of adjacent nodes. Similar to other machine learning models, the graph neural network model can be trained on graph structured data such as traffic interaction networks, where the data subject may need to forget some private data. Common forgetting methods include grouping the data and then retraining the corresponding model. However, the training of the graph neural network model depends on the structural information in the graph structured data, and the random division of subgraphs may seriously damage the practicality of the model. Therefore, new methods are urgently needed to eliminate the impact of the private data that needs to be forgotten on the graph neural network model.

[0024] In order to meet the needs of privacy protection, the simplest and most direct way is to delete the private data from the training sample data and then retrain the graph neural network model. However, the problems with this method are also obvious: in actual situations, the amount of data or the scale of the model is often too large, and retraining the graph neural network model will be very costly. Secondly, due to storage costs, many training sample data will not be retained indefinitely. Finally, some data is constantly updated, and the information contained in it may change over time and cannot be deleted.

[0025] At present, there are three main methods for solving the forgetting problem on graph neural network models: shard retraining, impact estimation, and external models. Among them, the shard retraining method is to first confirm the shard where the private data is located when the forgetting requirement is obtained, then remove the private data on the shard and retrain the corresponding model using the processed training sample data. The impact estimation method does not require data division, but calculates the impact function by confirming the impact range of the private data, and finally uses the impact function to directly update the parameters of the model to meet the forgetting requirement. The external model method is to use independent modules or models to help the graph neural network model meet the forgetting requirement. The main methods include knowledge distillation and third-party modules. Among them, the knowledge distillation method uses a teacher model trained on the remaining data other than the private data to guide the graph neural network model to perform forgetting operations, while the third-party module method introduces independent operators to meet the forgetting requirement.

[0026] There are many ways to try to forget private data in graph neural network models from different angles, but these methods often have various problems. The sharding retraining method is inefficient and its accuracy will be reduced to a certain extent when faced with a large amount of forgotten data. The impact estimation method is often designed for a specific graph neural network model, cannot be effectively applied to other models, and requires high computing resources. The external model method requires the introduction of additional models or operators, resulting in additional training costs. Therefore, it is necessary to design a forgetting method for graph neural network models that is independent of the model, has low training cost, and has little impact on the model.

[0027] This specification proposes a privacy data forgetting method for graph neural network models to eliminate the influence of specified privacy data (privacy data refers to data samples to be forgotten or the correlations, data features, etc.) on graph neural network models as required. For situations where the impact estimation method requires high computing resources, this specification effectively reduces the consumption of computing resources by generating hedge subgraph data and calculating the corresponding training loss instead of calculating the influence function. For situations where shard retraining and external model methods require high computing costs, this specification uses a comparative learning method to optimize the loss information of the hedge subgraph data and the original data, avoids the introduction of third-party modules and avoids complete retraining, thereby effectively reducing training costs. According to different forgetting requirements, this specification first determines the forgetting scope involved in the forgetting requirement and delineates the target subgraph data; then, for the privacy data in the target subgraph data, the corresponding hedge subgraph data is generated by adjusting the structural data and node features therein; finally, the privacy data and the hedge subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for forward training to obtain the corresponding loss information, and the difference in loss information between the hedge subgraph data and the privacy data is optimized by contrastive learning, so as to finally achieve the purpose of replacing the privacy data with the hedge data and eliminating the influence of the privacy data on other nodes in the target subgraph data, thereby completing the forgetting requirement. For specific processing, please refer to the specific content in the following embodiments.

[0028] like Figure 1 As shown, an embodiment of this specification provides a data processing method, and the execution subject of the method can be a terminal device or a server, etc., wherein the terminal device can be a mobile terminal device such as a mobile phone, a tablet computer, or a computer device such as a laptop or a desktop computer, or an IoT device (specifically such as a smart watch, a car-mounted device, etc.), etc., wherein the server can be an independent server, or a server cluster composed of multiple servers, etc., and the server can be a background server for financial services or online shopping services, or a background server for an application, etc. In this embodiment, the execution subject is taken as an example for detailed description. For the case where the execution subject is a terminal device, please refer to the following server situation processing, which will not be repeated here. The method can specifically include the following steps:

[0029] In step S102, graph structure data for model training of the graph neural network model is obtained, and the graph structure data includes nodes, edges and node features.

[0030] Among them, the graph neural network model (i.e., GNN model) may refer to a model composed of algorithms that use neural networks to learn graph structure data, extract and discover features and patterns in graph structure data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. The graph neural network model may specifically include graph convolutional neural network models, graph autoencoders, graph generation network models, graph recurrent network models, and graph attention network models, etc., which can be set according to actual conditions, and the embodiments of this specification do not limit this. The graph structure data may be any data related to the graph, for example, transaction-based graph structure data composed of transaction relationships between different users (transaction accounts as nodes, transaction relationships, etc. as edges, etc.), or communication interaction-based graph structure data composed of communication relationships between different users (communication accounts as nodes, interaction relationships, etc. as edges, etc.), etc. In addition, other types of graph structure data may also be included, which will not be repeated here. Generally, graph structure data can be composed of nodes and edges. In addition, for the nodes in the graph structure data, it can also include node attribute information (such as the node identification, the size of the node, the location of the node, etc., which can be set according to actual conditions), node characteristics (such as historical transaction characteristics of a trading account, historical transaction time characteristics of a trading account, historical transaction location characteristics of a trading account, or, friend characteristics of a communication account, language characteristics of a communication account, usage characteristics of a communication type of a communication account, etc.) and other data. For the edges in the graph structure data, it can also include edge attribute information (such as edge identification, relevant information of the two nodes connected by the edge, etc., which can be set according to actual conditions), edge weights and other data, which can be set according to actual conditions, and the embodiments of this specification do not limit this.

[0031] In implementation, graph structure data for model training of graph neural network model can be obtained in a variety of different ways. For example, graph structure data that can be used for model training of graph neural network model can be obtained from graph structure data stored in a specified database, or various graph structure data can be crawled from the Internet through a web crawler, and graph structure data that can be used for model training of graph neural network model can be selected from the crawled graph structure data. Alternatively, the server can record relevant data of the process of the server executing the corresponding business when the user executes the corresponding business, and can construct corresponding graph structure data based on the relevant data, and can select graph structure data that can be used for model training of graph neural network model from the constructed graph structure data. In addition to the above methods, graph structure data for model training of graph neural network model can be obtained in a variety of different ways. For example, the specified graph structure data can be purchased from the user, or the user can be invited to use it through internal testing or trial and obtain the user's consent to obtain the corresponding graph structure data, etc. The specific setting can be based on the actual situation, and the embodiments of this specification are not limited to this.

[0032] In step S104, the private data contained in the graph structure data is located according to the preset private data forgetting requirement, and the target subgraph data in the graph structure data that is within the forgetting range corresponding to the private data is determined according to the location result.

[0033] Among them, the privacy data forgetting requirement can be a requirement for forgetting privacy data according to actual needs, wherein the privacy data is located in the graph structure data, and the privacy data can be located in the nodes in the graph structure data, or in the edges in the graph structure data, or in the node features in the graph structure data. Based on this, the privacy data forgetting requirement can include multiple types, that is, the privacy data forgetting requirement can include one or more of the privacy data forgetting requirements of the node type, the privacy data forgetting requirements of the edge type, and the privacy data forgetting requirements of the node feature type. Among them, the privacy data forgetting requirement of the node type can be the privacy data forgetting requirement of the privacy data located in the node, and the privacy data forgetting requirement of the edge type can be the privacy data forgetting requirement of the privacy data located in the edge. The privacy data forgetting requirement of the node feature type may be the privacy data forgetting requirement where the privacy data is located in the node feature. The privacy data forgetting requirement may record which data or data related to which data needs to be forgotten (or forgotten). For example, the node type privacy data forgetting requirement may record the node identifier of the node that needs to be forgotten (or forgotten) (that is, it may record the node identifier of a node where the privacy data is located or the node identifier of the node where the privacy data is located). In addition to dividing and setting different types of privacy data forgetting requirements in the above manner, the division and / or setting of different types of privacy data forgetting requirements may also be achieved in other ways, which may be set specifically according to actual conditions, and the embodiments of this specification do not limit this. The forgetting range may be the range that the privacy data can affect. The forgetting range may be set based on expert experience or the actual conditions of the business corresponding to the graph structure data, or may be set based on relevant information of the graph neural network model (such as model parameters contained in the graph neural network model or relevant information of the network layer contained in the graph neural network model, etc.). In addition, in another embodiment, the forgetting range may also be set in the above privacy data forgetting requirement, which may be set specifically according to actual conditions. In practical applications, in graph structured data, when some of the data is private data, due to the association relationship in the graph structured data, the data that has a certain association relationship with the private data will also have a certain degree of privacy or sensitivity. For example, if a node in the graph structured data involves the user's private data, then other nodes directly connected to the node, and even designated nodes indirectly connected to the node may have a certain degree of privacy or sensitivity. For this reason, the forgetting range can be determined based on the scope that the private data can affect. For example, in the above example, the node and other nodes directly connected to the node can construct the forgetting range, or the node and other nodes directly connected to the node, as well as designated nodes indirectly connected to the node can construct the forgetting range, etc. The specific settings can be based on actual conditions.

[0034] In implementation, after obtaining the graph structure data in the above manner, a privacy data forgetting requirement can be set according to the actual situation. The privacy data forgetting requirement may include relevant information of the data to be forgotten (or relevant information of the privacy data). In addition, it may also include relevant information of the forgetting scope. Based on the relevant information of the data to be forgotten (or relevant information of the privacy data) recorded in the privacy data forgetting requirement, the privacy data to be forgotten can be searched from the graph structure data to locate the privacy data contained in the graph structure data. Specifically, if the graph structure data is ,in Represents a collection of nodes in graph structure data, Represents the set of edges in graph structure data, is a collection of node features. First, according to different privacy data forgetting requirements, the privacy data that needs to be forgotten can be located and the corresponding positioning results can be determined. Then, the forgetting scope can be clarified based on the positioning results. ,in Represents the relevant information of the data that needs to be forgotten (or the relevant information of the private data) recorded in different privacy data forgetting requirements, that is, it may be the node in the graph structure data Or side After that, the corresponding target subgraph data can be circled based on the forgetting range ,in Represents the set of nodes within the forgotten range, represents the nodes in the collection, Forgotten range of nodes and The set of edges between The data of node features within the forgotten range.

[0035] In step S106, the data in the target subgraph data is adjusted to generate the hedging subgraph data corresponding to the target subgraph data by using the hedging rule that matches the type corresponding to the privacy data forgetting requirement.

[0036] Among them, the hedging rules can be rules for generating hedging data, and the hedging data can be data used to replace privacy data. In this embodiment, the hedging data can be nodes, edges, node features, etc. The hedging rules can include multiple types, for example, adjusting the specified node features (specifically, the specified node features can be multiplied by a specified multiple, or the value corresponding to the specified node features minus a preset value, etc., which can be set according to actual conditions), or adjusting the weight of the specified edge (specifically, the weight of the specified edge can be multiplied by a specified multiple, or the weight of the specified edge minus a preset value, etc., which can be set according to actual conditions). It can be set according to actual conditions, and the embodiments of this specification do not limit this.

[0037] In implementation, corresponding hedging rules can be set according to actual conditions for different types of privacy data forgetting requirements. Hedging data corresponding to the privacy data can be generated through the hedging rules, and the corresponding privacy data can be replaced by the hedging data. Specifically, for example, the privacy data forgetting requirement includes the privacy data forgetting requirement of the node type, and the type corresponding to the privacy data forgetting requirement is the node type. At this time, the hedging rule corresponding to the node type can be obtained, and the preset adjustment method in the obtained hedging rule can be used to adjust the specified data in the target subgraph data. For example, the hedging rule corresponding to the node type is to multiply the node feature of the node corresponding to the node identifier recorded in the privacy data forgetting requirement by a preset multiple, then the node feature of the node corresponding to the above-mentioned node identifier can be obtained from the target subgraph data, and the obtained node feature can be multiplied by the preset multiple, and the result is used as the corresponding hedging data. Then, the hedging data can be used to replace the node feature of the node corresponding to the above-mentioned node identifier in the target subgraph data, thereby generating the hedging subgraph data corresponding to the target subgraph data.

[0038] It should be noted that the above hedging rules are only an optional implementation method. In actual applications, a variety of different hedging rules can be set. For example, hedging rules including adjusting the specified node characteristics and adjusting the weights of specified edges can be set for node types. The specific settings can be based on actual conditions, and the embodiments of this specification do not limit this.

[0039] In step S108, based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0040] In implementation, in order to replace the impact of hedging data on privacy data and achieve the purpose of forgetting privacy data, the target subgraph data and the hedging subgraph data can be used to train the graph neural network model through comparative learning. Specifically, the target subgraph data containing privacy data can be input into the graph neural network model to obtain the corresponding output result, and the hedging subgraph data that replaces the privacy data in the target subgraph data can be input into the graph neural network model to obtain the corresponding output result. Then, the similarity between the two output results can be determined through the above two output results, and the corresponding loss information can be obtained based on the obtained similarity. The model parameters in the graph neural network model can be adjusted through the loss information with the goal of maximizing the difference between the two output results. Then, the above model parameter adjustment process is repeated to iteratively train the graph neural network model until the specified number of iterations is reached. Finally, the trained graph neural network model can be obtained. The hedging data generated in the above manner can replace the privacy data in the graph structure data without having a significant impact on the output results of the graph neural network model, thereby protecting the security of user privacy data in the graph structure data.

[0041] In addition, after the graph neural network model is trained through the graph structure data as mentioned above, the graph structure data can be modified according to different privacy data forgetting requirements to complete the deletion requirements of the user's privacy data. Specifically, if the privacy data forgetting requirement is a node-type privacy data forgetting requirement, the corresponding node can be deleted from the graph structure data. If the privacy data forgetting requirement is an edge-type privacy data forgetting requirement, the corresponding edge can be deleted from the graph structure data. If the privacy data forgetting requirement is a node feature-type privacy data forgetting requirement, the node feature of the corresponding node in the graph structure data can be deleted, or the node feature of the corresponding node in the graph structure data can be set to a specified value (such as 0, etc.) or an invalid value, etc.

[0042] The embodiments of the present specification provide a data processing method, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0043] In practical applications, such as Figure 2 As shown, the privacy data forgetting requirement may include a node type privacy data forgetting requirement, and the node type privacy data forgetting requirement includes a node identifier of the first node. Based on this, the specific processing method for determining the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the positioning result in the above step S104 can be various. An optional processing method is provided below, such as Figure 3 As shown, the processing may specifically include the following steps S10402 and S10404.

[0044] In step S10402, according to the positioning result and the above-mentioned node identifier, a second node whose shortest path to the first node is within a first preset threshold is obtained in the graph structure data, and the forgetting range corresponding to the privacy data is determined based on the second node.

[0045] The first preset threshold value may be set according to actual conditions, such as 5 or 10.

[0046] In implementation, based on the node identifier in the privacy data forgetting requirement of the node type, the node corresponding to the node identifier (i.e., the first node) can be found in the graph structure data, thereby completing the location of the privacy data contained in the graph structure data and obtaining the corresponding location result. Based on the location result, the node (i.e., the second node) whose shortest path to the first node is within the first preset threshold can be found in the graph structure data, and the forgetting range corresponding to the privacy data can be determined based on the second node. Specifically, the forgetting range is ,in Represents the first node The shortest path between the first preset threshold The second node within.

[0047] In step S10404, the subgraph formed by the second nodes within the forgotten range corresponding to the privacy data is used as target subgraph data.

[0048] In implementation, based on the above content, the target subgraph data The subgraph structure data of the second node in the forgotten range, where Indicates the second node in the forgotten range.

[0049] Based on the processing of the above steps S10402 and S10404, Figure 2 As shown, the specific processing method of the above step S106 can be various. An optional processing method is provided below, which may specifically include the following contents: according to the hedging rule matching the node type, the node feature corresponding to the first node contained in the target subgraph data is converted into a negative node feature that is negatively correlated with the node feature, and the converted target subgraph data is used as the hedging subgraph data.

[0050] In implementation, based on the above content, the above privacy data forgetting requirement may include a privacy data forgetting requirement of a node type, and the privacy data forgetting requirement of the node type includes a node identifier of the first node. The first node ,in Represents the first node In the target subgraph data Structural data in and feature data Nodes can be assigned to different nodes according to the hedging rules that match the node type. (i.e. the first node ) into corresponding hedging data , the first node In the target subgraph data Feature data in Convert to the node feature Negatively correlated negative node features , specifically, copy the first node Corresponding feature data and structure data ,in Represents the first node In the target subgraph data The set of edges in , Represents the first node The corresponding node feature set, then the feature data can be Multiply all elements of - m (in m is a positive number. In practical applications, in order to simplify the processing, m Can be etc.) to obtain the corresponding hedging characteristic data , you can use the original structure data As the corresponding hedging structure data , thereby generating hedging data , finally, the hedge data can be used Replace the node where the private data in the target subgraph data is located Get the hedge subgraph data .

[0051] In practical applications, such as Figure 2 As shown, the privacy data forgetting requirement includes the edge type privacy data forgetting requirement, and the edge type privacy data forgetting requirement includes the edge identifier of the first edge. Based on this, the specific processing method for determining the target subgraph data in the graph structure data that is located within the forgetting range corresponding to the privacy data according to the positioning result in the above step S104 can be various. An optional processing method is provided below, such as Figure 4 As shown, the processing may specifically include the following steps S10406 ​​and S10408.

[0052] In step S10406, based on the positioning result and the above-mentioned edge identifier, two nodes connected by the first edge are obtained in the graph structure data, and a third node whose shortest path between at least one of the two nodes is within a second preset threshold is obtained in the graph structure data, and the forgetting range corresponding to the above-mentioned privacy data is determined based on the third node.

[0053] Among them, the second preset threshold can be set according to actual conditions, such as 8 or 15, etc. The second preset threshold can be the same as or different from the above-mentioned first preset threshold. In this embodiment, the second preset threshold and the above-mentioned first preset threshold can be the same as an example for explanation.

[0054] In implementation, based on the edge identifier of the first edge in the edge type privacy data forgetting requirement, the edge corresponding to the edge identifier (i.e., the first edge) can be found in the graph structure data, thereby completing the positioning of the privacy data contained in the graph structure data and obtaining the corresponding positioning result. Based on the positioning result, two nodes connected by the first edge can be found in the graph structure data, and a node (i.e., the third node) whose shortest path between at least one of the two nodes is within a second preset threshold can be obtained in the graph structure data, that is, the third node whose shortest path between any one of the two nodes is within the second preset threshold can be obtained in the graph structure data, or the third node whose shortest path between each of the two nodes is within the second preset threshold can be obtained in the graph structure data, and the forgetting range corresponding to the above privacy data can be further determined based on the third node. Specifically, the forgetting range , indicating that at least the first side A node of and / or ) the shortest path at the second preset threshold The third node within.

[0055] In step S10408, the subgraph formed by the third node within the forgotten range corresponding to the privacy data is used as the target subgraph data.

[0056] In implementation, based on the above content, the target subgraph data ,in It is the third node in the forgotten range.

[0057] Based on the processing of the above steps S10406 ​​and S10408, as Figure 2 As shown, the specific processing method of the above step S106 can be various. An optional processing method is provided below, which may specifically include the following contents: according to the hedging rule matching the edge type, the weight of the first edge in the structural data corresponding to each of the two nodes in the target subgraph data is converted into a negative weight negatively correlated with the weight, and the converted target subgraph data is obtained, and the converted target subgraph data is used as the hedging subgraph data.

[0058] In implementation, based on the above content, the above privacy data forgetting requirement may include an edge type privacy data forgetting requirement, and the edge type privacy data forgetting requirement includes the edge identifier of the first edge, then the node where the privacy data is located Contains two nodes connected by the first edge, and can determine the corresponding hedging data based on the two nodes connected by the first edge according to the hedging rules matching the edge type , that is, convert the weight of the first edge in the structure data corresponding to each of the two nodes in the target subgraph data into a negative weight negatively correlated with the weight. Specifically, copy the two nodes connected by the first edge respectively. and Corresponding structure data , , then multiply the weight of the first edge by (in, is a positive number. In practical applications, in order to simplify the processing, Can be etc.), with nodes For example, at node Corresponding structure data The weight of the first leg of the general becomes , the corresponding hedging structure data can be obtained through the above method , and use the original feature data ( ) as the hedging characteristic data ( ), and finally, the hedging data can be generated , finally, the hedge data can be used Replace the node where the private data in the target subgraph data is located Get the hedge subgraph data .

[0059] In practical applications, such as Figure 2 As shown, the privacy data forgetting requirement includes the privacy data forgetting requirement of the node feature type, and the privacy data forgetting requirement of the node feature type includes the identifier of the node feature of the fourth node. Based on this, the specific processing method for determining the target subgraph data in the graph structure data that is located within the forgetting range corresponding to the privacy data according to the positioning result in the above step S104 can be various. An optional processing method is provided below, such as Figure 5 As shown, the processing may specifically include the following steps S10410 and S10412.

[0060] In step S10410, according to the positioning result and the identification of the above node features, a fifth node whose shortest path to the fourth node is within a third preset threshold is obtained in the graph structure data, and the forgetting range corresponding to the above privacy data is determined based on the fifth node.

[0061] Among them, the third preset threshold can be set according to actual conditions, such as 7 or 13, etc. The third preset threshold, the above-mentioned first preset threshold and the second preset threshold can be the same as each other or different. In this embodiment, the third preset threshold, the above-mentioned first preset threshold and the second preset threshold are taken as an example to illustrate.

[0062] In implementation, based on the identifier of the node feature in the privacy data forgetting requirement of the node feature type, the node feature corresponding to the identifier of the node feature can be found in the graph structure data, and then the fourth node can be obtained, thereby completing the positioning of the privacy data contained in the graph structure data and obtaining the corresponding positioning result. Based on the positioning result, the fourth node can be found in the graph structure data, and the node (i.e., the fifth node) whose shortest path between the fourth node and the node feature corresponding to the identifier of the node feature is within the third preset threshold can be obtained in the graph structure data, that is, the fifth node whose shortest path between the fourth node and the fifth node is within the third preset threshold can be obtained in the graph structure data, and the forgetting range corresponding to the above privacy data can be determined based on the fifth node. Specifically, the forgetting range , which is the target node that needs to forget the feature The shortest path is within the third preset threshold The fifth node within.

[0063] In step S10412, the subgraph formed by the fifth node within the forgotten range corresponding to the privacy data is used as the target subgraph data.

[0064] In implementation, based on the above content, the target subgraph data is the subgraph structure data of the fifth node in the forgotten range, where It is the fifth node in the forgotten range.

[0065] Based on the processing of the above steps S10410 and S10412, as Figure 2 As shown, the specific processing method of the above step S106 can be various. An optional processing method is provided below, which may specifically include the following contents: according to the hedging rules matching the node feature type, the node features corresponding to the first four nodes contained in the target subgraph data are set as preset features to obtain the converted target subgraph data, and the converted target subgraph data is used as the hedging subgraph data.

[0066] In implementation, based on the above content, the above privacy data forgetting requirement may include a privacy data forgetting requirement of a node feature type, and the privacy data forgetting requirement of the node feature type includes an identifier of the node feature of the fourth node, then the node where the privacy data is located The fourth node ,in Represents the fourth node In the target subgraph data Structural data in and feature data Nodes can be assigned according to the hedging rules that match the node feature type. (i.e. the fourth node ) into corresponding hedging data , that is, the node feature corresponding to the fourth node contained in the target subgraph data is set to the preset feature. Specifically, copy the fourth node Corresponding feature data and structure data ,in Represents the fourth node In the target subgraph data The set of edges in , Represents the fourth node The corresponding node feature set, then the feature data corresponding to the fourth node can be Multiply all elements of by p to get the corresponding hedging feature data , you can use the original structure data As the corresponding hedging structure data , thereby generating hedging data ,in , ultimately, the hedge data can be used Replace the node where the private data in the target subgraph data is located Get the hedge subgraph data .

[0067] In practical applications, in order to simplify the processing and reduce resource consumption, the first preset threshold, the second preset threshold and the third preset threshold can be determined according to the number of network layers included in the graph neural network model. For example, the first preset threshold, the second preset threshold and the third preset threshold are the same, which are the number of network layers included in the graph neural network model. In practical applications, the first preset threshold, the second preset threshold and the third preset threshold can also be preset multiples of the number of network layers included in the graph neural network model (such as 2 times or 5 times, etc.). In addition, in order to simplify the processing and reduce resource consumption, the above preset feature can be 0.

[0068] In practical applications, the specific processing methods of the above step S108 can be varied. An optional processing method is provided below, such as Figure 6 As shown, the processing may specifically include the following steps S1082 to S1086.

[0069] In step S1082, the target subgraph data is input into the graph neural network model to obtain a first output result. Based on the first output result, the loss information corresponding to the first output result is determined through a preset first loss function. The first loss function is constructed through a cross entropy function or a KL divergence loss function.

[0070] In the implementation, the target subgraph data and the hedge subgraph data are used for model training respectively, and then the difference in loss information between the two batches of subgraph data is minimized by contrastive learning, so as to replace the influence of the hedge data on the private data and achieve the purpose of forgetting the private data. The iterative training process is as follows: Step S1082 to Step S1086. Specifically, the target subgraph data Input into the graph neural network model to obtain a first output result, and based on the first output result, determine the loss information corresponding to the first output result through a preset first loss function ,in, Represents different types of loss information such as cross entropy and KL divergence calculated based on model parameters and target subgraph data. Represents model parameters.

[0071] In step S1084, the hedge subgraph data is input into the graph neural network model to obtain a second output result. Based on the second output result, the loss information corresponding to the second output result is determined by a preset second loss function. The second loss function is constructed by a cross entropy function or a KL divergence loss function.

[0072] In implementation, the hedge subgraph data can be Input into the graph neural network model to obtain a second output result, and based on the second output result, determine the loss information corresponding to the second output result through a preset second loss function ,in, Represents different types of loss information such as cross entropy and KL divergence calculated based on model parameters and hedge subgraph data.

[0073] In step S1086, the total loss information is determined based on the loss information corresponding to the first output result and the loss information corresponding to the second output result, and the model parameters of the graph neural network model are updated by a preset stochastic gradient optimization algorithm and the total loss information to train the graph neural network model and obtain a trained graph neural network model.

[0074] In the implementation, the total loss information of this round of training process is calculated , where the total loss information It can be the difference between the loss information corresponding to the hedge subgraph data and the target subgraph data, that is, , and use parameter optimization algorithms such as stochastic gradient optimization to minimize the total loss information As the target, update the model parameters , to train the graph neural network model, and finally, the trained graph neural network model can be obtained.

[0075] The embodiments of the present specification provide a data processing method, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0076] The following is a detailed description of a data processing method provided by an embodiment of this specification in conjunction with a specific application scenario, wherein the graph structured data is a data transmission network constituted by sensor network traffic, the nodes in the graph structured data represent different sensors in the data transmission network constituted by sensors, the edges in the graph structured data represent the data transmission relationship between different sensors in the data transmission network, the node features in the graph structured data represent the data packets sent by the sensors, and the privacy data forgetting requirements include one or more of the requirements for the sensor to exit the data transmission network (i.e., the privacy data forgetting requirements of the node type), the requirements for clearing historical data packets in the sensor (i.e., the privacy data forgetting requirements of the node feature type), and the requirements for stopping the data transmission relationship between sensors (i.e., the privacy data forgetting requirements of the edge type).

[0077] In the above application scenario, the specific processing can also refer to the specific processing process provided in the above related content. This embodiment can be briefly described. For the related content or specific processing process not mentioned, refer to the above related content. Figure 2 , as follows:

[0078] In general, in the privacy data forgetting requirements of node type and node feature type, the nodes that need to be forgotten can be obtained, and in the privacy data forgetting requirements of edge type, the two nodes connected by the edge that needs to be forgotten can be obtained, so as to determine the corresponding forgetting range and target subgraph data. The goal of the forgetting processing corresponding to the privacy data forgetting requirements is to remove the influence of the information transmitted by the sensor (i.e., the node) on the graph neural network model, including the influence of the sensor itself and the influence on other sensors associated with it, and to achieve an effect as close as possible to that of a fully retrained model (i.e., completely retrained).

[0079] First, according to different types of privacy data forgetting requirements, the corresponding forgetting scope is clarified, and then the target subgraph data is circled. For the privacy data forgetting requirements of node types, the forgetting scope is the shortest path of the target node corresponding to the node identifier recorded in the privacy data forgetting requirements of the node type. (i.e., the number of network layers included in the graph structure data), the target subgraph data is the subgraph structure data composed of the nodes in the forgetting range. For edge-type privacy data forgetting requirements, the forgetting range is the shortest path of at least one node of the target edge corresponding to the edge identifier recorded in the edge-type privacy data forgetting requirements. For nodes within the range, the target subgraph data is delineated in the same way as the privacy data forgetting requirement of the node type. For the privacy data forgetting requirement of the node feature type, the forgetting range is the node shortest path of the node feature corresponding to the identifier of the node feature recorded in the privacy data forgetting requirement of the node feature type. Nodes within.

[0080] Then, the private data can be rewritten according to the different types corresponding to the privacy data forgetting requirements to generate the corresponding hedging data. The generation method includes negating all the values ​​of the feature data, negating the target edge weights, and returning the feature values ​​in the feature data to 0, which are respectively applicable to the privacy data forgetting requirements of the node type, the privacy data forgetting requirements of the edge type, and the privacy data forgetting requirements of the node feature type. After that, the hedging data is used to replace the private data in the target subgraph data to obtain the hedging subgraph data. Finally, the target subgraph data and the hedging subgraph data are added to the training process to calculate the corresponding loss information, and the contrastive learning method is used to optimize the interval between the loss information of the two batches of subgraph data to complete the information forgetting process. That is, after obtaining the loss information of the target subgraph data and the hedging subgraph data respectively, the difference between the two is used as the total loss information, and the parameter optimization method such as stochastic gradient optimization is used to minimize the total loss information for model training to complete the model parameter update process.

[0081] From the above specific implementation methods, this specification uses hedging data as a way to remove the impact of private data, so that the forgetting process can be completed with the help of the original model training process. Through comparative learning, the purpose of removing the impact of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently achieved. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0082] The embodiments of the present specification provide a data processing method, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0083] The above is a data processing method provided in the embodiment of this specification. Based on the same idea, the embodiment of this specification also provides a data processing device, such as Figure 7 shown.

[0084] The data processing device comprises: a data acquisition module 701, a target subgraph determination module 702, a hedge subgraph determination module 703 and a model training module 704, wherein:

[0085] A data acquisition module 701 acquires graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features;

[0086] The target subgraph determination module 702 locates the private data contained in the graph structure data according to the preset private data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the private data according to the location result;

[0087] A hedge subgraph determination module 703 adjusts the data in the target subgraph data to generate hedge subgraph data corresponding to the target subgraph data by using a hedge rule that matches the type corresponding to the privacy data forgetting requirement;

[0088] The model training module 704 performs model training on the graph neural network model by comparative learning based on the target subgraph data and the hedge subgraph data to obtain a trained graph neural network model.

[0089] In the embodiment of this specification, the privacy data forgetting requirement includes the privacy data forgetting requirement of the node type, and the privacy data forgetting requirement of the node type includes the node identifier of the first node, and the target subgraph determination module 702 includes:

[0090] A first range determination unit, which obtains, in the graph structure data, a second node whose shortest path to the first node is within a first preset threshold according to the positioning result and the node identifier, and determines a forgetting range corresponding to the privacy data based on the second node;

[0091] The first target subgraph determining unit uses the subgraph formed by the second nodes within the forgotten range corresponding to the privacy data as the target subgraph data.

[0092] In an embodiment of the present specification, the hedge subgraph determination module 703, according to a hedging rule matching the node type, converts the node feature corresponding to the first node contained in the target subgraph data into a negative node feature negatively correlated with the node feature, and uses the converted target subgraph data as the hedge subgraph data.

[0093] In the embodiment of this specification, the privacy data forgetting requirement includes an edge type privacy data forgetting requirement, and the edge type privacy data forgetting requirement includes an edge identifier of the first edge, and the target subgraph determination module 702 includes:

[0094] a second range determination unit, which obtains, in the graph structure data, two nodes connected by the first edge according to the positioning result and the edge identifier, and obtains, in the graph structure data, a third node whose shortest path between the two nodes and a node is within a second preset threshold, and determines a forgetting range corresponding to the privacy data based on the third node;

[0095] The second target subgraph determining unit uses the subgraph formed by the third node within the forgotten range corresponding to the privacy data as the target subgraph data.

[0096] In an embodiment of the present specification, the hedge subgraph determination module 703, according to a hedging rule matching the edge type, converts the weight of the first edge in the structural data corresponding to each of the two nodes in the target subgraph data into a negative weight negatively correlated with the weight, obtains the converted target subgraph data, and uses the converted target subgraph data as the hedge subgraph data.

[0097] In the embodiment of this specification, the privacy data forgetting requirement includes a privacy data forgetting requirement of a node feature type, and the privacy data forgetting requirement of the node feature type includes an identifier of a node feature of a fourth node. The target subgraph determination module 702 includes:

[0098] a third range determination unit, which obtains, in the graph structure data, a fifth node whose shortest path to the fourth node is within a third preset threshold according to the positioning result and the identifier of the node feature, and determines a forgetting range corresponding to the privacy data based on the fifth node;

[0099] The third target subgraph determining unit uses the subgraph formed by the fifth node within the forgotten range corresponding to the privacy data as the target subgraph data.

[0100] In the embodiment of the present specification, the hedge subgraph determination module 703 sets the node feature corresponding to the fourth node contained in the target subgraph data as a preset feature according to the hedging rule matching the node feature type, obtains the converted target subgraph data, and uses the converted target subgraph data as the hedge subgraph data.

[0101] In an embodiment of the present specification, the third preset threshold is determined according to the number of network layers included in the graph neural network model, and the preset feature is 0.

[0102] In the embodiment of this specification, the model training module 704 includes:

[0103] A first processing unit inputs the target subgraph data into the graph neural network model to obtain a first output result, and based on the first output result, determines loss information corresponding to the first output result through a preset first loss function, where the first loss function is constructed through a cross entropy function or a KL divergence loss function;

[0104] A second processing unit inputs the hedge subgraph data into the graph neural network model to obtain a second output result, and based on the second output result, determines the loss information corresponding to the second output result through a preset second loss function, where the second loss function is constructed through a cross entropy function or a KL divergence loss function;

[0105] The model training unit determines the total loss information based on the loss information corresponding to the first output result and the loss information corresponding to the second output result, and updates the model parameters of the graph neural network model through a preset stochastic gradient optimization algorithm and the total loss information to perform model training on the graph neural network model to obtain a trained graph neural network model.

[0106] In an embodiment of the present specification, the nodes in the graph structure data represent different sensors in a data transmission network formed by sensors, the edges in the graph structure data represent data transmission relationships between different sensors in the data transmission network, the node features in the graph structure data represent data packets sent by sensors, and the privacy data forgetting requirements include one or more of the requirements for sensors to exit the data transmission network, the requirements for clearing historical data packets in sensors, and the requirements for stopping data transmission relationships between sensors.

[0107] The embodiment of the present specification provides a data processing device, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training. Finally, the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0108] The above is a data processing device provided in the embodiment of this specification. Based on the same idea, the embodiment of this specification also provides a data processing device, such as Figure 8 shown.

[0109] The data processing device may provide a terminal device or a server, etc. for the above-mentioned embodiments.

[0110] The data processing device may have relatively large differences due to different configurations or performances, and may include one or more processors 801 and memory 802, and the memory 802 may store one or more storage applications or data. Among them, the memory 802 may be a short-term storage or a persistent storage. The application stored in the memory 802 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the data processing device. Furthermore, the processor 801 may be configured to communicate with the memory 802 and execute a series of computer executable instructions in the memory 802 on the data processing device. The data processing device may also include one or more power supplies 803, one or more wired or wireless network interfaces 804, one or more input and output interfaces 805, and one or more keyboards 806.

[0111] Specifically in this embodiment, the data processing device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions in the data processing device, and the one or more programs are configured to be executed by one or more processors, including computer executable instructions for performing the following:

[0112] Acquire graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features;

[0113] According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and according to the location result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined;

[0114] By using a hedging rule that matches the type corresponding to the privacy data forgetting requirement, adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data;

[0115] Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0116] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the data processing device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0117] The embodiment of the present specification provides a data processing device, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0118] Furthermore, based on the above Figures 1 to 6 In one embodiment, the present specification further provides a storage medium for storing computer executable instruction information. In a specific embodiment, the storage medium may be a USB flash drive, an optical disk, a hard disk, etc. When the computer executable instruction information stored in the storage medium is executed by the processor, the following process can be implemented:

[0119] Acquire graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features;

[0120] According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and according to the location result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined;

[0121] By using a hedging rule that matches the type corresponding to the privacy data forgetting requirement, adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data;

[0122] Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0123] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above-mentioned storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0124] The embodiment of the present specification provides a storage medium, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0125] Furthermore, based on the above Figures 1 to 6 In one or more embodiments of the present specification, a computer program product is provided, including a computer program. When the computer program in the computer program product is executed by a processor, the following process can be implemented:

[0126] Acquire graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features;

[0127] According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and according to the location result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined;

[0128] By using a hedging rule that matches the type corresponding to the privacy data forgetting requirement, adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data;

[0129] Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

[0130] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above-mentioned computer program product embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0131] The embodiments of the present specification provide a computer program product, which obtains graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges and node features, and then locates the privacy data contained in the graph structure data according to the preset privacy data forgetting requirement, and determines the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data according to the location result. Then, the data in the target subgraph data can be adjusted by a hedging rule that matches the type corresponding to the privacy data forgetting requirement to generate the hedging subgraph data corresponding to the target subgraph data. Finally, the graph neural network model can be trained by comparative learning based on the target subgraph data and the hedging subgraph data to obtain the trained graph neural network model. In this way, according to the forgetting requirement, the forgetting range involved in the forgetting requirement is first determined, and the target subgraph data is circled. Then, for the privacy data in the target subgraph data, the corresponding hedging subgraph data is generated by adjusting the structure data and node features therein. Finally, the privacy data and the hedging subgraph data are respectively input into the graph neural network model in combination with the target subgraph data for training, and finally the hedging data is used as a way to remove the influence of the privacy data, so that the forgetting process can be completed with the help of the original model training process. Through contrastive learning, the purpose of removing the influence of private data on the graph neural network model is achieved, and the forgetting of private data is efficiently realized. At the same time, this method is compatible with different graph neural network models and encourages the use of different hedging data generation technologies in different scenarios, so that the forgetting method can effectively adapt to more scenarios.

[0132] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0133] In the 1990s, it was very clear whether the improvement of a technology was hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one kind of HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0134] The controller may be implemented in any suitable manner, for example, the controller may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (e.g., software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, and the memory controller may also be implemented as part of the control logic of the memory. It is also known to those skilled in the art that, in addition to implementing the controller in a purely computer-readable program code manner, the controller may be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, such a controller may be considered as a hardware component, and the devices for implementing various functions included therein may also be considered as structures within the hardware component. Or even, the devices for implementing various functions may be considered as both software modules for implementing the method and structures within the hardware component.

[0135] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0136] For the convenience of description, the above devices are described in terms of functions and are divided into various units. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0137] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, one or more embodiments of this specification may take 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.) containing computer-usable program codes.

[0138] The embodiments of this specification are described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes 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 fraud case serial and parallel device to produce a machine, so that the instructions executed by the processor of the computer or other programmable fraud case serial and parallel device generate instructions for implementing the processes in the process. 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.

[0139] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable fraud case serial and parallel device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including 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.

[0140] These computer program instructions may also be loaded onto a computer or other programmable device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide 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.

[0141] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0142] 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. Memory is an example of a computer-readable medium.

[0143] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules 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, 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. As defined in this article, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0144] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0145] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, one or more embodiments of this specification may be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Furthermore, one or more embodiments of this specification may be in 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.) containing computer-usable program code.

[0146] One or more embodiments of the present specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0147] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0148] The above description is only an embodiment of this specification and is not intended to limit this document. For those skilled in the art, this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this specification should be included in the scope of the claims of this specification.

Claims

1. A data processing method, the method comprising: Acquire graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features, and the graph structure data includes data consisting of transaction relationships between different users; According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and according to the location result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined; By using a hedging rule that matches the type corresponding to the privacy data forgetting requirement, adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data; Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

2. According to the method of claim 1, the privacy data forgetting requirement includes a privacy data forgetting requirement of a node type, and the privacy data forgetting requirement of the node type includes a node identifier of the first node, and the determining, according to the positioning result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data comprises: According to the positioning result and the node identifier, a second node whose shortest path to the first node is within a first preset threshold is obtained in the graph structure data, and a forgetting range corresponding to the private data is determined based on the second node; The subgraph formed by the second nodes within the forgotten range corresponding to the privacy data is used as the target subgraph data.

3. The method according to claim 2, wherein adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data by using the hedging rule that matches the type corresponding to the privacy data forgetting requirement comprises: According to the hedging rule matching the node type, the node feature corresponding to the first node contained in the target subgraph data is converted into a negative node feature negatively correlated with the node feature, and the converted target subgraph data is used as the hedging subgraph data.

4. According to the method of claim 1, the privacy data forgetting requirement includes an edge-type privacy data forgetting requirement, and the edge-type privacy data forgetting requirement includes an edge identifier of a first edge, and determining the target subgraph data in the graph structure data that is within a forgetting range corresponding to the privacy data according to the positioning result comprises: According to the positioning result and the edge identifier, two nodes connected by the first edge are obtained in the graph structure data, and a third node whose shortest path between the two nodes and a node is within a second preset threshold is obtained in the graph structure data, and a forgetting range corresponding to the private data is determined based on the third node; The subgraph formed by the third node within the forgotten range corresponding to the privacy data is used as the target subgraph data.

5. The method according to claim 4, wherein adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data by using the hedging rule that matches the type corresponding to the privacy data forgetting requirement comprises: According to the hedging rule matching the edge type, the weight of the first edge in the structural data corresponding to each of the two nodes in the target subgraph data is converted into a negative weight negatively correlated with the weight to obtain the converted target subgraph data, and the converted target subgraph data is used as the hedging subgraph data.

6. According to the method of claim 1, the privacy data forgetting requirement includes a privacy data forgetting requirement of a node feature type, and the privacy data forgetting requirement of the node feature type includes an identifier of a node feature of a fourth node, and the determining, according to the positioning result, target subgraph data in the graph structure data that is within a forgetting range corresponding to the privacy data comprises: According to the positioning result and the identifier of the node feature, a fifth node whose shortest path to the fourth node is within a third preset threshold is obtained in the graph structure data, and a forgetting range corresponding to the privacy data is determined based on the fifth node; The subgraph formed by the fifth node within the forgotten range corresponding to the privacy data is used as the target subgraph data.

7. The method according to claim 6, wherein adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data by using the hedging rule that matches the type corresponding to the privacy data forgetting requirement comprises: According to the hedging rule matching the node feature type, the node feature corresponding to the fourth node contained in the target subgraph data is set as a preset feature to obtain the converted target subgraph data, and the converted target subgraph data is used as the hedging subgraph data.

8. According to the method of claim 7, the third preset threshold is determined according to the number of network layers included in the graph neural network model, and the preset feature is 0.

9. According to the method of claim 1, the graph neural network model is trained by contrastive learning based on the target subgraph data and the hedge subgraph data to obtain a trained graph neural network model, comprising: Input the target subgraph data into the graph neural network model to obtain a first output result, and based on the first output result, determine the loss information corresponding to the first output result through a preset first loss function, where the first loss function is constructed through a cross entropy function or a KL divergence loss function; Input the hedge subgraph data into the graph neural network model to obtain a second output result, and based on the second output result, determine the loss information corresponding to the second output result through a preset second loss function, where the second loss function is constructed through a cross entropy function or a KL divergence loss function; Based on the loss information corresponding to the first output result and the loss information corresponding to the second output result, the total loss information is determined, and the model parameters of the graph neural network model are updated through a preset stochastic gradient optimization algorithm and the total loss information to perform model training on the graph neural network model to obtain a trained graph neural network model.

10. According to the method according to claim 1, the nodes in the graph structure data represent different sensors in the data transmission network formed by sensors, the edges in the graph structure data represent the data transmission relationship between different sensors in the data transmission network, the node features in the graph structure data represent the data packets sent by the sensors, and the privacy data forgetting requirements include one or more of the requirements for the sensor to exit the data transmission network, the requirements for clearing historical data packets in the sensor, and the requirements for stopping the data transmission relationship between sensors.

11. A data processing device, comprising: A data acquisition module, which acquires graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features, and the graph structure data includes data consisting of transaction relationships between different users; a target subgraph determination module, which locates the private data contained in the graph structure data according to a preset private data forgetting requirement, and determines the target subgraph data in the graph structure data that is within a forgetting range corresponding to the private data according to the location result; A hedge subgraph determination module, which adjusts the data in the target subgraph data to generate the hedge subgraph data corresponding to the target subgraph data by using a hedge rule that matches the type corresponding to the privacy data forgetting requirement; The model training module performs model training on the graph neural network model by comparative learning based on the target subgraph data and the hedge subgraph data to obtain a trained graph neural network model.

12. A data processing device, comprising: processor; as well as a memory arranged to store computer executable instructions which, when executed, cause the processor to: Acquire graph structure data for model training of a graph neural network model, wherein the graph structure data includes nodes, edges, and node features, and the graph structure data includes data consisting of transaction relationships between different users; According to a preset privacy data forgetting requirement, the privacy data contained in the graph structure data is located, and according to the location result, the target subgraph data in the graph structure data that is within the forgetting range corresponding to the privacy data is determined; By using a hedging rule that matches the type corresponding to the privacy data forgetting requirement, adjusting the data in the target subgraph data to generate the hedging subgraph data corresponding to the target subgraph data; Based on the target subgraph data and the hedge subgraph data, the graph neural network model is trained by comparative learning to obtain a trained graph neural network model.

Images