Training method and system of graph neural network and abnormal account identification method

Abstract

The disclosure provides a graph neural network training method, a training system and an abnormal account identification method. The graph neural network training method comprises: obtaining initial graph structure data corresponding to a terminal device; the initial graph structure data obtained by a plurality of distributed training terminals respectively is derived from the same sample graph structure data; the following graph structure data processing stage and graph neural network training stage are executed cyclically until a target neural network meeting the training requirement is obtained: determining a processing time of the current execution graph structure data processing stage according to historical execution data of the historical execution graph structure data processing stage and the historical execution graph neural network training stage; performing graph structure data processing on the initial graph structure data in the graph structure data processing stage according to the processing time to generate target graph structure data; the graph structure data processing comprises data sampling processing and feature extraction processing; and training the target neural network based on the target graph structure data in the graph neural network training stage.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more specifically, to a training method, training system, and abnormal account identification method for graph neural networks. Background Technology

[0002] With the development of artificial intelligence technology, deep learning algorithms have achieved great success in application fields such as image processing and text processing. However, many real-world data, due to their special data forms, such as social networks, knowledge graphs, and molecular structures, are often difficult to process using conventional deep learning algorithms. To address this, researchers have proposed various graph neural network (GNN) models for graph structure data, which have been widely applied in fields such as content recommendation, drug development, and chip design.

[0003] In related technologies, when training graph neural networks, the graph structure data is generally processed first, and then the processed graph structure data is used for training. The two steps are executed sequentially. However, since the time required for processing the graph structure data and the time required for training with the processed graph structure data may not be the same, it may result in the next round of graph structure data not being processed before the current round of graph structure data has been trained, or the current round of graph structure data has been processed before the previous round of graph structure data has not been trained. This causes one of the processing steps to enter a waiting state, but the waiting processing step still occupies computing resources, affecting the utilization rate of computing resources. Summary of the Invention

[0004] This disclosure provides at least one method for training a graph neural network, a training system, and a method for identifying abnormal accounts.

[0005] In a first aspect, embodiments of this disclosure provide a training method for a graph neural network, applied to any terminal device among multiple distributed training terminals, wherein the multiple distributed training terminals are used to train the same target neural network, and the method includes:

[0006] The initial graph structure data corresponding to the terminal device is obtained; the initial graph structure data obtained by the multiple distributed training terminals are derived from the same sample graph structure data.

[0007] The following graph structure data processing stage and graph neural network training stage are executed iteratively until the target neural network that meets the training requirements is obtained:

[0008] Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage;

[0009] According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing.

[0010] The target neural network is trained based on the target graph structure data during the graph neural network training phase.

[0011] In one possible implementation, the initial graph structure data is obtained based on the following allocation method:

[0012] The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data.

[0013] Based on the greedy algorithm, the number of distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

[0014] In one possible implementation, the graph structure data processing of the initial graph structure data includes:

[0015] Based on the execution logic of multiple sampling and multiple feature extraction processes, a second graph structure data is generated to characterize the dependencies between operators.

[0016] Based on the second graph structure data, determine the topological order of each operator;

[0017] Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling and feature extraction processes.

[0018] One possible implementation involves determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the previous execution graph structure data processing stage and the graph neural network training stage, including:

[0019] Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment.

[0020] When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

[0021] In one possible implementation, training the target neural network based on the target graph structure data includes:

[0022] The execution code corresponding to the training phase of the graph neural network is input into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code;

[0023] The target graph structure data is processed according to the third graph structure data to train the target neural network.

[0024] Secondly, embodiments of this disclosure provide a training system for a graph neural network, including multiple distributed training terminals and a database, wherein:

[0025] The distributed training terminal is used to determine the processing timing of the current graph structure data processing stage based on historical execution data from previous graph structure data processing stages and graph neural network training stages; based on the processing timing, graph structure data processing is performed on the initial graph structure data in the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing; and the target neural network is trained based on the target graph structure data in the graph neural network training stage.

[0026] The database is used to store the initial graph data structure corresponding to each distributed training terminal, and to send the initial graph data structure to each of the distributed training terminals.

[0027] Thirdly, this disclosure also provides a method for identifying abnormal accounts, including:

[0028] Retrieve the browsing history of the account to be verified;

[0029] Based on the historical browsing records and the trained target neural network, a target identification result is determined to characterize whether the account to be verified is an abnormal account; wherein, the target neural network is trained based on any of the graph neural network training methods described in the first aspect.

[0030] In one possible implementation, determining the target identification result for characterizing whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network includes:

[0031] Based on the historical browsing records, the browsing feature information corresponding to the account to be verified is determined;

[0032] Based on the browsing feature information, construct the unverified graph structure data corresponding to the unverified account;

[0033] The graph structure data to be verified is input into the target neural network to obtain the target recognition result output by the target neural network.

[0034] In one possible implementation, after determining the target identification result used to characterize whether the account to be verified is an abnormal account, the method further includes:

[0035] If the account to be verified is an abnormal account, adjust the permission information corresponding to the account to be verified in order to control the permissions of the account.

[0036] Fourthly, embodiments of this disclosure also provide a training apparatus for a graph neural network, applied to any terminal device among multiple distributed training terminals, wherein the multiple distributed training terminals are used to train the same target neural network, and the apparatus includes:

[0037] The first acquisition module is used to acquire the initial graph structure data corresponding to the terminal device; the initial graph structure data acquired by the multiple distributed training terminals respectively comes from the same sample graph structure data;

[0038] The loop module is used to repeatedly execute the following graph structure data processing stage and graph neural network training stage until the target neural network that meets the training requirements is obtained:

[0039] Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage;

[0040] According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing.

[0041] The target neural network is trained based on the target graph structure data during the graph neural network training phase.

[0042] In one possible implementation, the initial graph structure data is obtained based on the following allocation method:

[0043] The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data.

[0044] Based on the greedy algorithm, the number of distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

[0045] In one possible implementation, the loop module, when performing graph structure data processing on the initial graph structure data, is used to:

[0046] Based on the execution logic of multiple sampling and multiple feature extraction processes, a second graph structure data is generated to characterize the dependencies between operators.

[0047] Based on the second graph structure data, determine the topological order of each operator;

[0048] Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling and feature extraction processes.

[0049] In one possible implementation, the loop module, when determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the historical execution graph structure data processing stage and the graph neural network training stage, is used to:

[0050] Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment.

[0051] When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

[0052] In one possible implementation, the loop module, when training the target neural network based on the target graph structure data, is used to:

[0053] The execution code corresponding to the training phase of the graph neural network is input into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code;

[0054] The target graph structure data is processed according to the third graph structure data to train the target neural network.

[0055] Fifthly, embodiments of this disclosure also provide an abnormal account identification device, comprising:

[0056] The second acquisition module is used to acquire the historical browsing records corresponding to the account to be verified.

[0057] The determination module is used to determine, based on the historical browsing records and the trained target neural network, a target identification result that characterizes whether the account to be verified is an abnormal account; wherein the target neural network is trained based on any of the graph neural network training methods described in the first aspect.

[0058] In one possible implementation, the determining module, when determining the target identification result for characterizing whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network, is used to:

[0059] Based on the historical browsing records, the browsing feature information corresponding to the account to be verified is determined;

[0060] Based on the browsing feature information, construct the unverified graph structure data corresponding to the unverified account;

[0061] The graph structure data to be verified is input into the target neural network to obtain the target recognition result output by the target neural network.

[0062] In one possible implementation, after determining the target identification result used to characterize whether the account to be verified is an abnormal account, the determining module is further configured to:

[0063] If the account to be verified is an abnormal account, adjust the permission information corresponding to the account to be verified in order to control the permissions of the account.

[0064] In a sixth aspect, embodiments of this disclosure also provide a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, they perform the steps in any of the possible implementations of the first or third aspect described above.

[0065] In a seventh aspect, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps in any of the possible implementations of the first or third aspect described above.

[0066] The graph neural network training method, training system, and abnormal account identification method provided in this disclosure utilize a distributed training terminal for graph neural network training. During the training process, the timing of the current graph structure data processing stage is determined based on historical execution data of the previous graph structure data processing stage and the graph neural network training stage. The graph structure data processing stage and the graph neural network training stage are then performed according to the determined timing. By controlling the timing of the graph structure data processing stage, the allocation of computing resources for both stages can be adjusted in real time, achieving a reasonable allocation of computing resources. Furthermore, by reasonably allocating computing resources for the graph structure data processing stage and the graph neural network training stage, the computation time of the graph structure data processing stage can be the same as or similar to that of the graph neural network training stage, avoiding a waiting state in any training stage and thus improving the utilization rate of computing resources.

[0067] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0069] Figure 1 A flowchart illustrating a training method for a graph neural network provided in an embodiment of this disclosure is shown;

[0070] Figure 2 The diagram illustrates the architecture of the graph neural network training system in the graph neural network training method provided in this embodiment of the present disclosure.

[0071] Figure 3 A flowchart of an abnormal account identification method provided by an embodiment of this disclosure is shown;

[0072] Figure 4 A schematic diagram of the architecture of a graph neural network training device provided in an embodiment of this disclosure is shown;

[0073] Figure 5This diagram illustrates the architecture of an abnormal account identification device provided in an embodiment of the present disclosure.

[0074] Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0076] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0077] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0078] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0079] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution, based on the prompt message.

[0080] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0081] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0082] In related technologies, during graph neural network training, the graph structure data processing stage and the graph neural network training stage are often executed sequentially. That is, after the graph data processing is completed, the graph neural network is trained based on the target graph structure data obtained after the graph data processing. However, this can lead to idle computing resources on the terminal device. In addition, even if the graph structure data processing stage and the graph neural network training stage are executed in parallel, because the computing resources of the terminal device are limited, if the computing resources allocated to each processing stage are uneven, it may result in the next round of graph structure data not being processed before the current round of graph structure data has been trained, or the current round of graph structure data has been processed before the previous round of graph structure data has not been trained. This may cause one of the processing steps to enter a waiting state, which may also lead to idle computing resources.

[0083] Based on the above research, this disclosure provides a training method, training system, and abnormal account identification method for graph neural networks. The graph neural network is trained using a distributed training terminal. During the training process, the timing of the current graph structure data processing stage is determined based on historical execution data of the previous graph structure data processing stage and the graph neural network training stage. The graph structure data processing stage and the graph neural network training stage are then performed according to the determined timing. By controlling the timing of the graph structure data processing stage, the allocation of computing resources for both stages can be adjusted in real time, achieving a reasonable allocation of computing resources. Furthermore, by reasonably allocating computing resources for the graph structure data processing stage and the graph neural network training stage, the computation time of the graph structure data processing stage can be the same as or similar to that of the graph neural network training stage, avoiding a waiting state in one training stage and thus improving the utilization rate of computing resources.

[0084] To facilitate understanding of this embodiment, a detailed description of the graph neural network training method disclosed in this disclosure is provided first. The execution entity of the graph neural network training method provided in this disclosure is generally a computer device with certain computing capabilities. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), mobile device, user terminal, terminal, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, the graph neural network training method can be implemented by a processor calling computer-readable instructions stored in memory.

[0085] See Figure 1 The diagram shows a flowchart of a graph neural network training method provided in this embodiment. The method is applied to any terminal device among multiple distributed training terminals, which are used to train the same target neural network. The method includes steps S101 to S102, wherein:

[0086] S101: Obtain the initial graph structure data corresponding to the terminal device; the initial graph structure data obtained by the multiple distributed training terminals respectively comes from the same sample graph structure data.

[0087] S102: Repeatedly execute the graph structure data processing stage and the graph neural network training stage until the target neural network that meets the training requirements is obtained.

[0088] The graph structure data processing stage includes:

[0089] S1021. Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage.

[0090] S1022. According to the processing timing, the initial graph structure data is processed in the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing.

[0091] The training phase of the graph neural network includes:

[0092] S1023. During the graph neural network training phase, the target neural network is trained based on the target graph structure data.

[0093] The following is a detailed explanation of the steps described above.

[0094] Regarding S101, the sample graph structure data is graph structure data used to train the target neural network. The graph structure data consists of vertices and edges, and can be used to characterize the relationships between elements in fields such as protein analysis, chip design, and autonomous driving. The network type of the target neural network can be a neural network capable of processing graph structure data, such as a graph natural network (GNN) or a graph convolutional network (GCN).

[0095] In practical applications, in distributed training scenarios, the initial graph structure data is distributed to each distributed training terminal, and each distributed training terminal performs distributed training based on the distributed graph structure data.

[0096] In one possible implementation, the initial graph structure data can be allocated through the following steps A1-A2:

[0097] A1: The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data.

[0098] Here, the sample graph structure data can be composed of multiple vertices and multiple edges. Through the breadth-first search algorithm, the sample graph structure data containing multiple vertices and multiple edges can be divided into multiple first graph structure data.

[0099] A2: Based on the greedy algorithm, the number of terminals of the distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

[0100] Here, the greedy algorithm can be used to distribute multiple first graph structure data to each distributed training terminal to reduce the network data access overhead between distributed training terminals in subsequent data stages, thereby reducing the network resource consumption caused by data transmission and improving the training efficiency of the target neural network.

[0101] Specifically, a mathematical model can be constructed based on the number of distributed training terminals and the number of portions of the first graph structure data obtained from the segmentation, to obtain the mathematical model used by the greedy algorithm for allocation. Based on the greedy algorithm and the mathematical model, an allocation score is calculated for any first graph structure data, and the allocation score is used to determine which distributed training terminal to which the first graph structure data is allocated. When calculating the allocation score of the first graph structure data, it can be calculated according to the type of vertices it contains, based on the weight value corresponding to the type of the vertex. Different types of vertices may have different weight values, and the type of vertex may include training, verification, testing, etc.

[0102] For example, taking the allocation of the first graph structure data "graph structure data 1" as an example, the allocation scores of "graph structure data 1" for distributed training terminals 1 to 3 are score 1, score 2 and score 3 respectively, where score 1 < score 2 < score 3. Then, according to the allocation scores, the "graph structure data 1" can be allocated to the distributed training terminal 1 with the largest allocation score.

[0103] In this way, the mathematical problem of "how to allocate multiple first graph structure data obtained after the sample graph structure data is processed to each distributed training terminal to achieve the optimal effect (i.e., reduce the network resource consumption caused by data transmission)" can be transformed into "how to allocate any graph structure data to a suitable distributed training terminal to achieve the local optimum effect (i.e., reduce the network resource consumption caused by data transmission)" by using a greedy algorithm. By calculating the allocation score that represents the matching degree between each first graph structure data and each distributed training terminal, the first graph structure data can be allocated to the distributed training terminal with the highest allocation score to achieve the local optimum effect. By merging the local optimum terminals, the final allocation result can be obtained.

[0104] For S102, the graph structure data processing may include sampling processing and feature extraction processing; wherein, for any of the distributed training terminals, in addition to sampling the initial graph structure data acquired by the distributed training terminal, other distributed training terminals besides the one mentioned above may also be sampled according to the actual needs during training; in the feature extraction processing, the graph structure data obtained after sampling processing may be subjected to feature extraction processing to obtain target graph structure data that can be used to train the target neural network.

[0105] In practical applications, during a graph structure data processing process, sampling processing is required first, followed by feature extraction processing. Therefore, these two steps are often executed sequentially in related technologies. However, there may be no dependency between different sampling processes (or feature extraction processes) during graph structure data processing. Therefore, different sampling processes (or feature extraction processes) may also be executed simultaneously to improve the data processing efficiency of the graph structure data processing stage.

[0106] In one possible implementation, the graph structure data processing of the initial graph structure data can be performed through the following steps B1 to B3:

[0107] B1: Based on the execution logic of multiple sampling and multiple feature extraction processes, a second graph structure data is generated to represent the dependencies between operators.

[0108] B2: Based on the second graph structure data, determine the topological order of each operator.

[0109] B3: Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling processes and multiple feature extraction processes.

[0110] Here, the execution logic of the sampling process and feature extraction process can be represented by the execution code, and the programming language corresponding to the execution code can be, for example, the graph traversal language Gremlin.

[0111] Specifically, when generating the second graph structure data, the second graph structure data used to represent the dependencies between operators can be automatically constructed according to the sampling logic of the target neural network for graph data structure, and the topological order used to represent the target processing training can be determined according to the dependencies between operators in the second graph structure data.

[0112] In practical applications, since terminal devices can perform multiple graph structure data processing operations simultaneously, and the computing resources of terminal devices are limited, the number of graph structure data processing operations performed at the same time can be changed by adjusting the timing of the current graph structure data processing, thereby changing the computing resources allocated to the graph structure data processing stage.

[0113] In one possible implementation, when determining the timing of the current graph structure data processing stage based on historical execution data of the previous graph structure data processing stage and graph neural network training stage, it can be determined whether to execute the current graph structure data processing at the current moment based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage.

[0114] Here, the performance parameters of the terminal device may include at least one of CPU utilization, intrinsic parameter utilization, and GPU utilization; the first historical average duration of the previous executions of the graph structure data processing stage can be determined by the duration of each previous execution of the graph structure data processing stage before the current execution of the graph structure data processing stage during the current training process, and is used to measure the amount of first computing resources allocated to the graph structure data processing stage during the current training process. The first historical average duration is inversely proportional to the first computing resources, that is, the longer the first historical average duration, the less first computing resources are allocated to the graph structure data processing stage during the current training process; the second historical average duration of the previous executions of the graph neural network training stage can be determined by the duration of each previous execution of the graph neural network training stage before the current execution of the graph neural network training stage during the current training process, and is used to measure the amount of second computing resources allocated to the graph neural network training stage during the current training process. The second historical average duration is inversely proportional to the second computing resources, that is, the longer the second historical average duration, the less second computing resources are allocated to the graph neural network training stage during the current training process.

[0115] Specifically, when determining whether to execute the current graph structure data processing at the current moment based on the performance parameters of the terminal device, the first historical average duration of the previous executions of the graph structure data processing stage, and the second historical average duration of the previous executions of the graph neural network training stage, an execution score can be determined to characterize whether to execute the current graph structure data processing at the current moment. If the execution score is greater than a preset execution score threshold, it indicates that the current graph structure data processing needs to be executed at the current moment; if the execution score is less than the preset execution score threshold, it indicates that the current graph structure data processing does not need to be executed at the current moment.

[0116] Furthermore, if it is not necessary to perform the current graph structure data processing at the current moment, it can be determined again after a preset time interval, based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, whether to perform the current graph structure data processing at the current moment (i.e., after the preset time interval has elapsed).

[0117] In this way, by taking into account the performance parameters of the terminal device at the current moment, the first historical average duration of the first computing resources allocated to the graph structure data processing stage during this training process, and the second historical average duration of the second computing resources allocated to the graph neural network training stage during this training process, it is possible to comprehensively consider the performance of the terminal device and the allocation of computing resources in each processing stage to determine whether to execute the current graph structure data processing at the current moment. This allows for control over the computing resources allocated to each processing stage, thereby improving the training efficiency of the target neural network by allocating computing resources reasonably.

[0118] Below, we will use a specific calculation formula to illustrate the above-mentioned "determining whether to perform the current sub-graph structure data processing" in this embodiment of the disclosure:

[0119] Specifically, depending on the different tasks performed by the terminal device, the task processing module in the terminal device can be divided into a sampling module and a training module. The sampling module is used to perform data processing tasks in the graph structure data processing stage, and the training module is used to perform data processing tasks in the graph neural network training stage. The data processing tasks to be performed can be stored in the form of a queue after they are generated.

[0120] Assuming the performance parameter of the terminal device is A, the time interval from the last execution of graph structure data processing to the current moment is T1, the task queue length of the sampling module is S, the first historical average duration of the previous executions of the graph structure data processing phase is T2, and the second historical average duration of the previous executions of the graph neural network training phase is T3, then the formula for determining the execution score X, which characterizes whether the current graph structure data processing is performed at the current moment, can be:

[0121] χ=f1×f2×T1

[0122]

[0123] f2=α-e A / β

[0124] Where α and β are preset hyperparameters; f1 is used to characterize whether the computing resources currently allocated to each processing stage of the terminal device are balanced; f2 is used to characterize the remaining computing resources of the terminal device.

[0125] In this way, by combining the computing resources currently allocated to each processing stage by the terminal device and the remaining computing resources, an execution score can be determined to characterize whether to execute the current subgraph structure data processing at the current moment. By comparing the execution score with a preset execution score threshold, the timing of executing the current subgraph structure data processing stage can be determined, thereby balancing the computing resources currently allocated to each processing stage by the terminal device and improving the training efficiency of the target neural network.

[0126] In one possible implementation, when training the target neural network based on the target graph structure data, the following steps C1 to C2 can be used:

[0127] C1: Input the execution code corresponding to the training stage of the graph neural network into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code.

[0128] Here, the execution code corresponding to the training stage of the graph neural network can be input into the target code compiler TensorFlow for compilation to obtain the third graph structure data output by the target code compiler. The third graph structure data can be used to characterize the execution logic during the training stage of the graph neural network.

[0129] C2: Perform data processing on the target graph structure data according to the third graph structure data to train the target neural network.

[0130] Here, when training the target neural network, the target neural network can be trained according to the execution logic corresponding to the third graph structure data, the target graph structure data, and the pre-set graph neural network training algorithm. This embodiment of the disclosure does not limit the specific type of graph neural network training algorithm, as long as the target neural network that meets the preset training requirements can be obtained after implementation.

[0131] See Figure 2 The diagram shown is an architectural schematic of a graph neural network training system provided in this embodiment of the present disclosure, including multiple distributed training terminals and a database, wherein:

[0132] The distributed training terminal is used to determine the processing timing of the current graph structure data processing stage based on historical execution data from previous graph structure data processing stages and graph neural network training stages; based on the processing timing, graph structure data processing is performed on the initial graph structure data in the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing; and the target neural network is trained based on the target graph structure data in the graph neural network training stage.

[0133] The database is used to store the initial graph data structure corresponding to each distributed training terminal, and to send the initial graph data structure to each of the distributed training terminals.

[0134] Specifically, the distributed training terminal can consist of a sampling module and a training module; wherein, the sampling module is used to execute data processing tasks in the graph structure data processing stage, and the training module is used to execute data processing tasks in the graph neural network training stage. The data processing tasks to be executed can be stored in the form of a queue after they are generated.

[0135] See Figure 3 As shown, an abnormal account identification method provided in this embodiment includes steps S301 to S302, wherein:

[0136] S301: Retrieve the browsing history of the account to be verified.

[0137] Here, the browsing history corresponding to the account to be verified is the browsing history collected after the user's authorization.

[0138] S302: Based on the historical browsing records and the trained target neural network, determine the target identification result used to characterize whether the account to be verified is an abnormal account; wherein, the target neural network is trained based on the graph neural network training method described in any embodiment of this disclosure.

[0139] In one possible implementation, when determining the target identification result for characterizing whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network, the following steps D1 to D3 can be used:

[0140] D1: Based on the historical browsing records, determine the browsing feature information corresponding to the account to be verified.

[0141] Here, before determining the browsing characteristic information corresponding to the account to be verified, data cleaning operations can also be performed on the historical browsing records.

[0142] D2: Based on the browsing feature information, construct the graph structure data to be verified corresponding to the account to be verified.

[0143] D3: Input the graph structure data to be verified into the target neural network to obtain the target recognition result output by the target neural network.

[0144] In one possible implementation, after determining the target identification result used to characterize whether the account to be verified is an abnormal account, the permission information corresponding to the account to be verified can be adjusted to control the permissions of the account to be verified if the account to be verified is an abnormal account.

[0145] Here, adjusting the permission information corresponding to the account to be verified can mean canceling the target permissions corresponding to the account to be verified.

[0146] Furthermore, if the account to be verified is a normal account, it can be verified again after a preset verification interval.

[0147] Furthermore, the graph neural network training method provided in this disclosure can be used not only for identifying abnormal accounts, but also for applications such as chip design, scene analysis and problem reasoning, content recommendation, drug development, traffic flow prediction, and autonomous driving. This disclosure does not limit the specific application scenarios; the focus is on enabling data processing within the chosen application scenario using the graph neural network.

[0148] For example, taking content recommendation as an application scenario, the browsing feature information of the target user can be determined based on the target user's historical browsing records collected after the target user's authorization. Then, based on the trained target neural network and the browsing feature information, media content, advertisements, and other content can be recommended to the target user, so that the recommended content can meet the different needs of different users.

[0149] The graph neural network training method, training system, and abnormal account identification method provided in this disclosure utilize a distributed training terminal for graph neural network training. During the training process, the timing of the current graph structure data processing stage is determined based on historical execution data of the previous graph structure data processing stage and the graph neural network training stage. The graph structure data processing stage and the graph neural network training stage are then performed according to the determined timing. By controlling the timing of the graph structure data processing stage, the allocation of computing resources for both stages can be adjusted in real time, achieving a reasonable allocation of computing resources. Furthermore, by reasonably allocating computing resources for the graph structure data processing stage and the graph neural network training stage, the computation time of the graph structure data processing stage can be the same as or similar to that of the graph neural network training stage, avoiding a waiting state in any training stage and thus improving the utilization rate of computing resources.

[0150] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0151] Based on the same inventive concept, this disclosure also provides a graph neural network training device corresponding to the graph neural network training method. Since the principle of the device in this disclosure for solving the problem is similar to the graph neural network training method described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0152] Reference Figure 4 The diagram shown is an architectural schematic of a graph neural network training device provided in an embodiment of this disclosure. The device includes: a first acquisition module 401 and a loop module 402; wherein,

[0153] The first acquisition module 401 is used to acquire the initial graph structure data corresponding to the terminal device; the initial graph structure data acquired by the multiple distributed training terminals respectively comes from the same sample graph structure data;

[0154] Loop module 402 is used to repeatedly execute the following graph structure data processing stage and graph neural network training stage until the target neural network that meets the training requirements is obtained:

[0155] Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage;

[0156] According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing.

[0157] The target neural network is trained based on the target graph structure data during the graph neural network training phase.

[0158] In one possible implementation, the initial graph structure data is obtained based on the following allocation method:

[0159] The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data.

[0160] Based on the greedy algorithm, the number of distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

[0161] In one possible implementation, the loop module 402, when performing graph structure data processing on the initial graph structure data, is used to:

[0162] Based on the execution logic of the multiple sampling process and the multiple feature extraction process, a second graph structure data is generated to characterize the dependencies between operators;

[0163] Based on the second graph structure data, determine the topological order of each operator;

[0164] Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling and feature extraction processes.

[0165] In one possible implementation, the loop module 402, when determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the historical execution graph structure data processing stage and the graph neural network training stage, is used to:

[0166] Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment.

[0167] When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

[0168] In one possible implementation, the loop module 402, when training the target neural network based on the target graph structure data, is used to:

[0169] The execution code corresponding to the training phase of the graph neural network is input into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code;

[0170] The target graph structure data is processed according to the third graph structure data to train the target neural network.

[0171] The graph neural network training apparatus provided in this embodiment uses a distributed training terminal to train the graph neural network. During the training process, based on historical execution data of the previous graph structure data processing stage and the graph neural network training stage, the processing timing of the current graph structure data processing stage is determined. The graph structure data processing stage and the graph neural network training stage are then performed according to the determined processing timing. By controlling the processing timing of the graph structure data processing stage, the allocation of computing resources for the graph structure data processing stage and the graph neural network training stage can be adjusted in real time to achieve a reasonable allocation of computing resources. Furthermore, by reasonably allocating computing resources for the graph structure data processing stage and the graph neural network training stage, the computation time of the graph structure data processing stage can be the same as or similar to the computation time of the graph neural network training stage, avoiding a waiting state in a certain training stage and thus improving the utilization rate of computing resources.

[0172] Reference Figure 5 The diagram shown is an architectural schematic of an abnormal account identification device provided in an embodiment of this disclosure. The device includes: a second acquisition module 501 and a determination module 502; wherein,

[0173] The second acquisition module 501 is used to acquire the historical browsing records corresponding to the account to be verified.

[0174] The determining module 502 is used to determine a target identification result that characterizes whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network; wherein the target neural network is trained based on any of the graph neural network training methods described in the first aspect.

[0175] In one possible implementation, the determining module 502, when determining the target identification result for characterizing whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network, is used to:

[0176] Based on the historical browsing records, the browsing feature information corresponding to the account to be verified is determined;

[0177] Based on the browsing feature information, construct the unverified graph structure data corresponding to the unverified account;

[0178] The graph structure data to be verified is input into the target neural network to obtain the target recognition result output by the target neural network.

[0179] In one possible implementation, after determining the target identification result used to characterize whether the account to be verified is an abnormal account, the determining module 502 is further configured to:

[0180] If the account to be verified is an abnormal account, adjust the permission information corresponding to the account to be verified in order to control the permissions of the account.

[0181] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0182] Based on the same technical concept, this disclosure also provides a computer device. (See also...) Figure 6 The diagram shows the structure of a computer device 600 provided in this embodiment of the present disclosure, including a processor 601, a memory 602, and a bus 603. The memory 602 stores execution instructions and includes main memory 6021 and external memory 6022. The main memory 6021, also called internal memory, is used to temporarily store computational data in the processor 601 and data exchanged with external memory 6022 such as a hard disk. The processor 601 exchanges data with the external memory 6022 through the main memory 6021. When the computer device 600 is running, the processor 601 and the memory 602 communicate through the bus 603, causing the processor 601 to execute the following instructions:

[0183] The initial graph structure data corresponding to the terminal device is obtained; the initial graph structure data obtained by the multiple distributed training terminals are derived from the same sample graph structure data.

[0184] The following graph structure data processing stage and graph neural network training stage are executed iteratively until the target neural network that meets the training requirements is obtained:

[0185] Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage;

[0186] According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing.

[0187] The target neural network is trained based on the target graph structure data during the graph neural network training phase.

[0188] In one possible implementation, the initial graph structure data in the instructions of the processor 601 is allocated based on the following method:

[0189] The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data.

[0190] Based on the greedy algorithm, the number of distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

[0191] In one possible implementation, the instructions of the processor 601, wherein the graph structure data processing of the initial graph structure data includes:

[0192] Based on the execution logic of multiple sampling and multiple feature extraction processes, a second graph structure data is generated to characterize the dependencies between operators.

[0193] Based on the second graph structure data, determine the topological order of each operator;

[0194] Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling and feature extraction processes.

[0195] In one possible implementation, the instructions of the processor 601 determine the processing timing of the current execution of the graph structure data processing stage based on historical execution data from the previous execution of the graph structure data processing stage and the graph neural network training stage, including:

[0196] Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment.

[0197] When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

[0198] In one possible implementation, the instructions of the processor 601, wherein training the target neural network based on the target graph structure data, includes:

[0199] The execution code corresponding to the training phase of the graph neural network is input into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code;

[0200] The target graph structure data is processed according to the third graph structure data to train the target neural network; or,

[0201] This causes processor 601 to execute the following instructions:

[0202] Retrieve the browsing history of the account to be verified;

[0203] Based on the historical browsing records and the trained target neural network, a target identification result is determined to characterize whether the account to be verified is an abnormal account; wherein, the target neural network is trained based on any of the graph neural network training methods described in the first aspect.

[0204] In one possible implementation, the instructions of the processor 601, wherein determining the target identification result for characterizing whether the account to be verified is an abnormal account based on the historical browsing records and the trained target neural network, includes:

[0205] Based on the historical browsing records, the browsing feature information corresponding to the account to be verified is determined;

[0206] Based on the browsing feature information, construct the unverified graph structure data corresponding to the unverified account;

[0207] The graph structure data to be verified is input into the target neural network to obtain the target recognition result output by the target neural network.

[0208] In one possible implementation, after determining the target identification result used to characterize whether the account to be verified is an abnormal account, the method further includes the following instructions in the processor 601:

[0209] If the account to be verified is an abnormal account, adjust the permission information corresponding to the account to be verified in order to control the permissions of the account.

[0210] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the graph neural network training method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0211] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the graph neural network training method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0212] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0213] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0214] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0215] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0216] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0217] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A training method for a graph neural network, characterized in that, The method is applied to any terminal device among multiple distributed training terminals, wherein the multiple distributed training terminals are used to train the same target neural network, and the method includes: The initial graph structure data corresponding to the terminal device is obtained; the initial graph structure data obtained by the multiple distributed training terminals are derived from the same sample graph structure data. The following graph structure data processing stage and graph neural network training stage are executed iteratively until the target neural network that meets the training requirements is obtained: Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage; According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing. During the training phase of the graph neural network, the target neural network is trained based on the target graph structure data. The step of determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the historical execution graph structure data processing stage and the graph neural network training stage includes: Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment. When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

2. The method according to claim 1, characterized in that, The initial graph structure data was allocated based on the following method: The sample graph structure data is segmented based on the breadth-first search algorithm to obtain multiple first graph structure data. Based on the greedy algorithm, the number of distributed training terminals, and the number of parts of the first graph structure data obtained by segmentation, the initial graph structure data is allocated to the distributed training terminals from the plurality of first graph structure data.

3. The method according to claim 1, characterized in that, The graph structure data processing of the initial graph structure data includes: Based on the execution logic of multiple sampling and multiple feature extraction processes, a second graph structure data is generated to characterize the dependencies between operators. Based on the second graph structure data, determine the topological order of each operator; Based on the topological order of each operator, the initial graph structure data is subjected to multiple sampling and feature extraction processes.

4. The method according to claim 1, characterized in that, The training of the target neural network based on the target graph structure data includes: The execution code corresponding to the training phase of the graph neural network is input into the target code compiler to obtain the third graph structure data generated by the target code compiler after compiling the execution code; The target graph structure data is processed according to the third graph structure data to train the target neural network.

5. A training system for a graph neural network, characterized in that, It includes multiple distributed training terminals and databases, among which: The distributed training terminal is used to determine the processing timing of the current graph structure data processing stage based on historical execution data from previous graph structure data processing stages and graph neural network training stages; based on the processing timing, graph structure data processing is performed on the initial graph structure data in the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing; and the target neural network is trained based on the target graph structure data in the graph neural network training stage. The database is used to store the initial graph data structure corresponding to each distributed training terminal, and to send the initial graph data structure to each of the distributed training terminals. The step of determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the historical execution graph structure data processing stage and the graph neural network training stage includes: Based on the performance parameters of the terminal device, the first historical average duration of the previous executions of the graph structure data processing stage, and the second historical average duration of the previous executions of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment. When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

6. A method for identifying abnormal accounts, characterized in that, include: Retrieve the browsing history of the account to be verified; Based on the historical browsing records and the trained target neural network, a target identification result is determined to characterize whether the account to be verified is an abnormal account; wherein, the target neural network is trained based on the training method of the graph neural network according to any one of claims 1 to 4.

7. The method according to claim 6, characterized in that, The determination of the target identification result, based on the historical browsing records and the trained target neural network, to characterize whether the account to be verified is an abnormal account includes: Based on the historical browsing records, the browsing feature information corresponding to the account to be verified is determined; Based on the browsing feature information, construct the unverified graph structure data corresponding to the unverified account; The graph structure data to be verified is input into the target neural network to obtain the target recognition result output by the target neural network.

8. The method according to claim 6, characterized in that, After determining the target identification result used to characterize whether the account to be verified is an abnormal account, the method further includes: If the account to be verified is an abnormal account, adjust the permission information corresponding to the account to be verified in order to control the permissions of the account.

9. A training device for a graph neural network, characterized in that, An apparatus applicable to any one of multiple distributed training terminals, wherein the multiple distributed training terminals are used to train the same target neural network, the apparatus comprising: The first acquisition module is used to acquire the initial graph structure data corresponding to the terminal device; the initial graph structure data acquired by the multiple distributed training terminals respectively comes from the same sample graph structure data; The loop module is used to repeatedly execute the following graph structure data processing stage and graph neural network training stage until the target neural network that meets the training requirements is obtained: Based on the historical execution data of the previous execution graph structure data processing stage and the graph neural network training stage, determine the processing timing of the current execution graph structure data processing stage; According to the aforementioned processing timing, the initial graph structure data is processed during the graph structure data processing stage to generate target graph structure data; the graph structure data processing includes data sampling processing and feature extraction processing. During the training phase of the graph neural network, the target neural network is trained based on the target graph structure data. The step of determining the processing timing of the current execution graph structure data processing stage based on historical execution data from the historical execution graph structure data processing stage and the graph neural network training stage includes: Based on the performance parameters of the terminal device, the first historical average duration of the previous execution of the graph structure data processing stage, and the second historical average duration of the previous execution of the graph neural network training stage, it is determined whether to execute the current graph structure data processing at the current moment. When determining whether to perform the current subgraph structure data processing at the current time, the processing timing is determined to be the current time; otherwise, after a preset time interval, it is re-determined whether to perform the current subgraph result data processing.

10. An abnormal account identification device, characterized in that, include: The second acquisition module is used to acquire the historical browsing records corresponding to the account to be verified. The determination module is used to determine, based on the historical browsing records and the trained target neural network, a target identification result that characterizes whether the account to be verified is an abnormal account; wherein the target neural network is trained based on the training method of the graph neural network according to any one of claims 1 to 4.

11. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the training method for a graph neural network as described in any one of claims 1 to 4 are performed. Alternatively, perform the steps of the abnormal account identification method as described in any one of claims 6 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the training method for a graph neural network as described in any one of claims 1 to 4. Alternatively, perform the steps of the abnormal account identification method as described in any one of claims 6 to 8.

Images