Graph neural network training, task processing method and apparatus, electronic device, and storage medium

Abstract

The present disclosure relates to a graph neural network training method, a task processing method, an apparatus, an electronic device and a storage medium. The graph neural network training method comprises: obtaining object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of a multimedia resource in which the sample object has performed a preset operation in a plurality of preset time periods; performing denoising processing on the positive sample data to obtain target sample data; and performing reconstruction training on a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data, to obtain a target graph neural network. The embodiment of the present disclosure can greatly improve the representation accuracy of the multimedia resource, and further improve the recommendation accuracy.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and particularly relates to a graph neural network training method and device, a task processing method and device, an electronic device and a storage medium. BACKGROUND

[0002] With the development of Internet technology, a large number of information recommendation platforms are also constantly upgrading, and a large number of products with multimedia resource production + recommendation as the core business have emerged. As we all know, the accurate representation of multimedia resources is one of the most core technologies for personalized recommendation products.

[0003] In the related art, a graph is often constructed in combination with multimedia resources of platform content, nodes in the graph represent multimedia resources, and edges between nodes are constructed (such as an edge between two videos continuously watched by the same user) in combination with user behavior operations on multimedia resources to generate graph data; then, graph neural networks are trained based on the graph data to learn vector representation (embedding) of multimedia resources; but in the above related technology, due to the uncertainty of user behavior, the graph data of multimedia resources used for learning has a lot of noise, and the noise propagates with the graph structure, resulting in that the graph neural network often cannot accurately learn the feature representation of the node, causing the representation accuracy of the multimedia resource and the subsequent recommendation accuracy to be poor. SUMMARY

[0004] The present disclosure provides a graph neural network training method, a task processing method, a device, an electronic device and a storage medium to at least solve the problem that the representation accuracy of multimedia resources and the subsequent recommendation accuracy are poor in the related art. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a graph neural network training method is provided, comprising:

[0006] obtaining object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of multimedia resources on which the sample object has performed a preset operation in a plurality of preset time periods;

[0007] performing denoising processing on the positive sample data to obtain target sample data;

[0008] performing reconfiguration training on a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data to obtain a target graph neural network.

[0009] Optionally, the denoising processing on the positive sample data to obtain target sample data comprises:

[0010] The positive sample data is denoised based on a data denoising network to obtain target sample data.

[0011] Optionally, the method further comprises:

[0012] obtaining first sample resource information of a multimedia resource corresponding to the positive sample data, second sample resource information of a multimedia resource corresponding to the negative sample data, and initial graph data, the initial graph data comprising nodes connected based on an association relationship between multimedia resources corresponding to the positive sample data, the nodes being the positive sample data;

[0013] inputting the initial graph data into a preset graph neural network to extract features to obtain initial node information corresponding to the nodes in the initial graph data;

[0014] performing data denoising training on a preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain the data denoising network.

[0015] Optionally, the performing reconstruction training on the preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data to obtain a target graph neural network comprises:

[0016] generating first target graph data based on an association relationship between multimedia resources corresponding to the target sample data;

[0017] inputting the first target graph data into the preset graph neural network to extract features to obtain first target node information corresponding to the nodes in the first target graph data;

[0018] inputting object information of a sample object corresponding to the target sample data and the first target node information into a preset generation network to perform sample reconstruction processing to obtain first reconstructed sample data;

[0019] determining first loss information according to the first reconstructed sample data and the target sample data;

[0020] performing reconstruction training on the preset graph neural network based on the first loss information to obtain the target graph neural network.

[0021] Optionally, the performing reconstruction training on the preset graph neural network based on the first loss information to obtain the target graph neural network comprises:

[0022] if the first loss information does not satisfy a first preset condition, adjusting network parameters of the preset generation network and the preset graph neural network based on the first loss information;

[0023] generate a network update based on the preset generation network after adjusting network parameters;

[0024] repeating the data denoising training of the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information until the first loss information is determined based on the first reconstructed sample data and the target sample data;

[0025] In the case where the first loss information meets the first preset condition, the preset graph neural network corresponding to the case where the first preset condition is met is taken as the target graph neural network.

[0026] Optionally, the denoising processing of the positive sample data by the data denoising network to obtain the target sample data includes:

[0027] The object information of the sample object corresponding to the positive sample data, the initial node information, and the first sample resource information are input into the data denoising network for confidence identification processing to obtain target confidence data, the target confidence data representing the probability that the sample object will perform a preset operation on the multimedia resource corresponding to the positive sample data.

[0028] Based on the target confidence data, the positive sample data is resampled to obtain the target sample data.

[0029] Optionally, the resource information includes resource identification information, and the initial graph data is obtained in the following manner:

[0030] The co-occurrence relationship between two resource identification information is determined by co-occurrence analysis of the resource identification information, the co-occurrence relationship representing the proportion of the pair of resource identification information corresponding to the pair of multimedia resources that have been executed by the same object in the same preset time period in the positive sample data.

[0031] Based on the co-occurrence relationship, the association relationship between the multimedia resources corresponding to the positive sample data is determined.

[0032] The resource identification information of the multimedia resource corresponding to the positive sample data is taken as a node, and the node is connected based on the association relationship to obtain the initial graph data.

[0033] Optionally, the data denoising training of the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain the data denoising network includes:

[0034] input the initial node information, the first sample resource information, the second sample resource information, and the object information into the preset data denoising network for confidence identification processing to obtain predicted confidence data, the predicted confidence data representing a probability that the sample object will perform a preset operation on multimedia resources corresponding to sample data;

[0035] According to the predicted confidence data and the corresponding preset confidence data, determine second loss information;

[0036] Based on the second loss information, the preset data denoising network is trained to obtain the data denoising network.

[0037] Optionally, the denoising processing of the positive sample data to obtain the target sample data includes:

[0038] Obtain behavior duration information of the sample object on the multimedia resources corresponding to the positive sample data;

[0039] Based on the behavior duration information, generate a sampling probability corresponding to the positive sample data;

[0040] Based on the sampling probability, resample the positive sample data to obtain the target sample data.

[0041] Optionally, the reconstruction training of the preset graph neural network based on the target sample data and the target object information of the sample object corresponding to the target sample data to obtain the target graph neural network includes:

[0042] Based on the association relationship between the multimedia resources corresponding to the target sample data, generate second target graph data;

[0043] Input the second target graph data into the preset graph neural network for feature extraction to obtain second target node information corresponding to nodes in the second target graph data;

[0044] Input the object information of the sample object corresponding to the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data;

[0045] According to the second reconstructed sample data and the target sample data, determine third loss information;

[0046] Based on the third loss information, the preset graph neural network is reconstructed and trained to obtain the target graph neural network.

[0047] Optionally, the reconstruction training of the preset graph neural network based on the third loss information to obtain the target graph neural network includes:

[0048] In a case where the third loss information does not satisfy a third preset condition, network parameters of the preset generation network and the preset graph neural network are adjusted based on the third loss information.

[0049] The second target node information is updated based on the preset generation network after the network parameters are adjusted.

[0050] The object information of the sample object corresponding to the target sample data and the second target node information are input into the preset generation network for sample reconstruction processing repeatedly based on the updated second target node information until the second reconstruction sample data is obtained, and the third loss information is determined based on the second reconstruction sample data and the target sample data.

[0051] In a case where the third loss information satisfies the third preset condition, the preset graph neural network corresponding to the case where the third preset condition is satisfied is taken as the target graph neural network.

[0052] According to a second aspect of the embodiments of the present disclosure, a task processing method is provided, including:

[0053] The preset graph data is input into the target graph neural network obtained by the graph neural network training method to obtain resource feature information corresponding to nodes in the preset graph data; the preset graph data includes nodes connected based on an association relationship between a plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources;

[0054] The target task is executed based on the resource feature information.

[0055] Optionally, the target task includes an information recall task, and the target task is executed based on the resource feature information, including:

[0056] The similarity between the plurality of multimedia resources is determined based on the resource feature information.

[0057] The associated multimedia resource corresponding to each multimedia resource is determined based on the similarity.

[0058] In the information recall processing based on any multimedia resource in the plurality of multimedia resources, the information recall processing is performed based on the corresponding associated multimedia resource.

[0059] Optionally, the target task includes a training task of a target network, and the target task is executed based on the resource feature information, including:

[0060] The preset annotation information of the multimedia resource corresponding to the resource feature information is obtained, and the preset annotation information represents preset task result information of an associated task corresponding to the target network.

[0061] inputting the resource feature information into a preset neural network for correlation task processing to obtain predicted task result information;

[0062] determining third loss information according to the preset annotation information and the predicted task result information;

[0063] training the preset neural network based on the third loss information to obtain the target network.

[0064] According to a third aspect of the embodiments of the present disclosure, a graph neural network training apparatus is provided, comprising:

[0065] A first data acquisition module configured to perform acquisition of object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of multimedia resources on which the sample object has performed a preset operation in a plurality of preset time periods;

[0066] A denoising processing module configured to perform denoising processing on the positive sample data to obtain target sample data;

[0067] A reconstruction training module configured to perform reconstruction training of a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data to obtain a target graph neural network.

[0068] Optionally, the denoising processing module is specifically configured to perform denoising processing on the positive sample data based on a data denoising network to obtain target sample data.

[0069] Optionally, the apparatus further comprises:

[0070] A second data acquisition module configured to perform acquisition of first sample resource information of a multimedia resource corresponding to the positive sample data of the sample object, second sample resource information of a multimedia resource corresponding to negative sample data, and initial graph data, the initial graph data comprising nodes connected based on an association relationship between the multimedia resources corresponding to the positive sample data, the nodes being the positive sample data;

[0071] A first feature extraction module configured to perform inputting of the initial graph data into a preset graph neural network for feature extraction to obtain initial node information corresponding to nodes in the initial graph data;

[0072] A data denoising training module configured to perform data denoising training of a preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain the data denoising network.

[0073] Optionally, the reconstruction training module comprises:

[0074] The first target graph data generation unit is configured to generate first target graph data based on an association relationship between multimedia resources corresponding to the target sample data;

[0075] The first feature extraction unit is configured to input the first target graph data into the preset graph neural network to perform feature extraction, and obtain first target node information corresponding to nodes in the first target graph data;

[0076] The first sample reconstruction processing unit is configured to input object information of a sample object corresponding to the target sample data and the first target node information into a preset generation network to perform sample reconstruction processing, and obtain first reconstructed sample data;

[0077] The first loss information determination unit is configured to determine first loss information according to the first reconstructed sample data and the target sample data;

[0078] The first reconstruction training unit is configured to perform reconstruction training on the preset graph neural network based on the first loss information, and obtain the target graph neural network.

[0079] Optionally, the first reconstruction training unit comprises:

[0080] The first network parameter adjustment unit is configured to adjust network parameters of the preset generation network and the preset graph neural network based on the first loss information in a case where the first loss information does not satisfy a first preset condition;

[0081] The initial node information updating unit is configured to update the initial node information based on the preset generation network after the network parameters are adjusted;

[0082] The first iterative training unit is configured to repeatedly perform the data denoising training on the preset data denoising network based on the updated initial node information, from the step of performing data denoising training on the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information and the object information to the step of determining the first loss information according to the first reconstructed sample data and the target sample data.

[0083] The first target graph neural network determination unit is configured to determine, in a case where the first loss information satisfies the first preset condition, a preset graph neural network corresponding to the case where the first loss information satisfies the first preset condition as the target graph neural network.

[0084] Optionally, the denoising processing module comprises:

[0085] The first confidence identification processing unit is configured to perform confidence identification processing on the object information of the sample object corresponding to the positive sample data, the initial node information corresponding to the positive sample data, and the first sample resource information input into the data denoising network, to obtain target confidence data, the target confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to the positive sample data.

[0086] The first resampling unit is configured to perform resampling on the positive sample data based on the target confidence data, to obtain the target sample data.

[0087] Optionally, the resource information includes resource identification information, and the initial graph data is obtained by using the following unit:

[0088] The co-occurrence analysis unit is configured to perform co-occurrence analysis on the resource identification information, to determine a co-occurrence relationship between two pieces of resource identification information, the co-occurrence relationship representing a proportion of a pair of resource identification information corresponding to two multimedia resources that are performed by the same object in a same preset time period in the positive sample data.

[0089] The association relationship determination unit is configured to determine an association relationship between multimedia resources corresponding to the positive sample data based on the co-occurrence relationship.

[0090] The initial graph data construction unit is configured to perform connection between nodes based on the association relationship, to obtain the initial graph data, by taking the resource identification information of the multimedia resources corresponding to the positive sample data as the nodes.

[0091] Optionally, the data denoising training module includes:

[0092] The second confidence identification processing unit is configured to perform confidence identification processing on the initial node information, the first sample resource information, the second sample resource information, and the object information input into the preset data denoising network, to obtain predicted confidence data, the predicted confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to sample data.

[0093] The second loss information determination unit is configured to determine second loss information according to the predicted confidence data and corresponding preset confidence data.

[0094] The denoising training unit is configured to train the preset data denoising network based on the second loss information, to obtain the data denoising network.

[0095] Optionally, the denoising processing module includes:

[0096] The behavior duration information acquisition unit is configured to acquire behavior duration information of the sample object on the multimedia resource corresponding to the positive sample data.

[0097] The sampling probability generation unit is configured to generate a sampling probability corresponding to the positive sample data based on the behavior duration information.

[0098] The second resampling unit is configured to resample the positive sample data based on the sampling probability to obtain the target sample data.

[0099] Optionally, the reconstruction training module comprises:

[0100] The second target graph data generation unit is configured to generate second target graph data based on the association relationship between the multimedia resources corresponding to the target sample data.

[0101] The second feature extraction unit is configured to input the second target graph data into the preset graph neural network to extract features to obtain second target node information corresponding to nodes in the second target graph data.

[0102] The second sample reconstruction processing unit is configured to input object information of a sample object corresponding to the target sample data and the second target node information into a preset generation network to perform sample reconstruction processing to obtain second reconstructed sample data.

[0103] The third loss information determination unit is configured to determine third loss information according to the second reconstructed sample data and the target sample data.

[0104] The second reconstruction training unit is configured to perform reconstruction training on the preset graph neural network based on the third loss information to obtain the target graph neural network.

[0105] Optionally, the second reconstruction training unit comprises:

[0106] The second network parameter adjustment unit is configured to adjust network parameters of the preset generation network and the preset graph neural network based on the third loss information in a case where the third loss information does not satisfy a third preset condition.

[0107] The target node information updating unit is configured to update the second target node information based on the preset generation network after adjusting the network parameters.

[0108] a second iterative training unit configured to perform, based on the updated second target node information, the steps of repeatedly inputting object information of a corresponding sample object of the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data, and determining third loss information according to the second reconstructed sample data and the target sample data;

[0109] a second target graph neural network determination unit configured to perform, in a case where the third loss information satisfies the third preset condition, a step of determining a preset graph neural network corresponding to a time when the third preset condition is satisfied as the target graph neural network.

[0110] According to a fourth aspect of the embodiments of the present disclosure, a task processing apparatus is provided, comprising:

[0111] a second feature extraction module configured to perform a step of inputting preset graph data into a target graph neural network obtained by the graph neural network training method according to any one of the first aspect to obtain resource feature information corresponding to nodes in the preset graph data; the preset graph data comprises nodes connected based on an association relationship between a plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources;

[0112] a target task execution module configured to perform a target task based on the resource feature information.

[0113] Optionally, the target task comprises an information recall task, and the target task execution module comprises:

[0114] a similarity determination unit configured to perform a step of determining a similarity between the plurality of multimedia resources according to the resource feature information;

[0115] an associated multimedia resource determination unit configured to perform a step of determining an associated multimedia resource corresponding to each multimedia resource according to the similarity;

[0116] an information recall processing unit configured to perform a step of, in information recall processing based on any one of the plurality of multimedia resources, performing information recall processing based on the corresponding associated multimedia resource.

[0117] Optionally, the target task comprises a training task of a target network, and the target task execution module comprises:

[0118] a preset annotation information acquisition unit configured to perform a step of acquiring preset annotation information of a multimedia resource corresponding to the resource feature information, the preset annotation information representing preset task result information of an associated task corresponding to the target network;

[0119] The association task processing unit is configured to perform inputting the resource feature information into a preset neural network for association task processing to obtain predicted task result information.

[0120] The third loss information determination unit is configured to perform determining third loss information according to the preset annotation information and the predicted task result information.

[0121] The network training unit is configured to perform training the preset neural network based on the third loss information to obtain the target network.

[0122] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the method according to any one of the first aspect or the second aspect.

[0123] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the method according to any one of the first aspect or the second aspect of the present disclosure.

[0124] According to a fifth aspect of an embodiment of the present disclosure, a computer program product containing instructions is provided, when the computer program product is executed on a computer, the computer can execute the method according to any one of the first aspect or the second aspect of the present disclosure.

[0125] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0126] In the process of training the graph neural network, by performing denoising processing on the positive sample data, the effectiveness and objectivity of the sample data can be improved, and the preset graph neural network can be reconstructed and trained in combination with the target graph data of the target sample data after denoising and the object information of the corresponding sample object, which can greatly improve the representation accuracy of the graph neural network for multimedia resources, and further improve the recommendation accuracy in subsequent applications.

[0127] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0128] The accompanying drawings incorporated in the specification and forming a part thereof illustrate embodiments consistent with the present disclosure and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation on the present disclosure.

[0129] Figure 1 is a schematic diagram of an application environment according to an exemplary embodiment;

[0130] Figure 2 FIG. 1 is a flowchart of a graph neural network training method according to an example embodiment;

[0131] Figure 3 FIG. 2 is a flowchart of training a data denoising network according to an example embodiment;

[0132] Figure 4 FIG. 3 is a flowchart of obtaining initial graph data according to an example embodiment;

[0133] Figure 5 FIG. 4 is a flowchart of denoising positive sample data to obtain target sample data according to an example embodiment;

[0134] Figure 6 FIG. 5 is a flowchart of reconstructing and training a preset graph neural network based on target sample data and target object information of a sample object corresponding to the target sample data to obtain a target graph neural network according to an example embodiment;

[0135] Figure 7 FIG. 6 is a schematic diagram of a graph neural network training process according to an example embodiment;

[0136] Figure 8 FIG. 7 is a flowchart of a task processing method according to an example embodiment;

[0137] Figure 9 FIG. 8 is a flowchart of executing a target task based on resource feature information according to an example embodiment;

[0138] Figure 10 FIG. 9 is another flowchart of executing a target task based on resource feature information according to an example embodiment;

[0139] Figure 11 FIG. 10 is a block diagram of a graph neural network training device according to an example embodiment;

[0140] Figure 12 FIG. 11 is a block diagram of a task processing device according to an example embodiment;

[0141] Figure 13 FIG. 12 is a block diagram of an electronic device for graph neural network training according to an example embodiment;

[0142] Figure 14 FIG. 13 is a block diagram of an electronic device for task processing according to an example embodiment. DETAILED DESCRIPTION

[0143] In order for those skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the drawings.

[0144] It should be noted that the terms "first", "second" and the like in the specification and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or a chronological sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following exemplary embodiments does not represent all implementations consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0145] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties.

[0146] Please refer to Figure 1 , Figure 1 is a schematic diagram of an application environment according to an exemplary embodiment, as Figure 1 shown, the application environment can include a server 100 and a terminal 200.

[0147] In an optional embodiment, the server 100 can be used to train a target graph neural network. Specifically, the server 100 can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or an edge computing node, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc. Basic cloud computing services.

[0148] In an optional embodiment, the terminal 200 can perform task processing based on the target graph neural network trained by the server 100. Specifically, the terminal 200 can include, but is not limited to, an electronic device such as a smart phone, a desktop computer, a tablet computer, a notebook computer, a smart speaker, a digital assistant, an augmented reality (AR) / virtual reality (VR) device, a smart wearable device, and the like, or a software such as an application program running on the electronic device. Optionally, the operating system running on the electronic device can include, but is not limited to, an Android system, an IOS system, Linux, Windows, and the like.

[0149] In addition, it should be noted that, Figure 1 The above-mentioned terminal 200 and server 100 can be directly or indirectly connected through wired or wireless communication, and the present disclosure does not limit the same.

[0150] The above-mentioned terminal 200 and server 100 can be directly or indirectly connected through wired or wireless communication, and the present disclosure does not limit the same.

[0151] Figure 2 is a flowchart of a graph neural network training method according to an exemplary embodiment, as Figure 2 The graph neural network training method is used in a terminal or a server, and includes the following steps.

[0152] In step S201, object information of a sample object and positive sample data of the sample object are obtained.

[0153] In an optional embodiment, the sample object can be a user or a user account. The sample object can include a plurality of objects, and the number of objects can be set according to actual application requirements. Optionally, the object information of the sample object can be feature representation information determined based on object attributes of the sample object. In an embodiment, the object information of the sample object can be a feature vector representation of the object attribute information. In an embodiment, the object attribute information can include, but is not limited to, information representing personal attributes of a user, such as gender, age, education, and region. In an optional embodiment, the feature vector representation of the object attribute information can be determined based on a word vector model. Optionally, the word vector model can include, but is not limited to, a trained word2vec, fasttext, glove, and the like.

[0154] In a specific embodiment, the positive sample data of the sample object can be resource information of multimedia resources on which the sample object has performed a preset operation in a plurality of preset time periods. The preset operation can include, but is not limited to, browsing, clicking, conversion (such as purchasing a related product based on the multimedia resource, or downloading a related application based on the multimedia resource, etc.), and the like. Specifically, the preset time period can be a unit collection time length of the preset sample resource data, for example, the preset time period is 30 minutes, and correspondingly, the positive sample data can be resource information of multimedia resources on which the sample object has performed a preset operation within half an hour.

[0155] In a specific embodiment, the multimedia resource can include static resources such as text and image, and can also include dynamic resources such as short video. In a specific embodiment, the resource information can include resource identification information (identification information for distinguishing different multimedia resources). Optionally, the resource information can also include resource attribute information of the multimedia resource. Specifically, the resource attribute information can be information for describing the multimedia resource. Taking the multimedia resource as a video for example, the resource attribute information can include text attribute information such as publisher information and playing time, and can also include image attribute information such as cover.

[0156] In step S203, the positive sample data is denoised to obtain target sample data.

[0157] In an optional embodiment, the denoising of the positive sample data to obtain the target sample data can include denoising the positive sample data based on the data denoising network to obtain the target sample data.

[0158] In a specific embodiment, as shown in Figure 3 The method can further include a step of training the data denoising network. Specifically, the step of training the data denoising network includes:

[0159] In step S301, first sample resource information corresponding to the multimedia resource of the positive sample data of the sample object, second sample resource information corresponding to the multimedia resource of the negative sample data, and initial graph data are obtained. The initial graph data includes nodes connected based on the association relationship between the multimedia resources corresponding to the positive sample data, and the node is the positive sample data.

[0160] In step S303, the initial graph data is input into a preset graph neural network for feature extraction to obtain initial node information corresponding to the nodes in the initial graph data.

[0161] In step S305, the preset data denoising network is trained based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain the data denoising network.

[0162] In an optional embodiment, the negative sample data of the sample object can be resource information of multimedia resources on which the sample object has not performed the preset operation in the plurality of preset time periods.

[0163] In a specific embodiment, the first sample resource information can be feature representation information determined by the resource attribute information of the multimedia resources corresponding to the positive sample data. The second sample resource information can be feature representation information determined by the resource attribute information of the multimedia resources corresponding to the negative sample data. In an optional embodiment, the feature representation information of the resource attribute information can be generated in combination with a feature representation network such as a convolutional neural network, an N-Gram (Chinese language model), etc.

[0164] In the above embodiments, the data denoising network is trained by combining the initial graph data corresponding to the initial node information, the sample resource information corresponding to the positive and negative sample data, and the object information, and the data denoising network is generated, which can improve the data denoising effect of the trained data denoising network. In combination with the data denoising network, the positive sample data is denoised, which can improve the effectiveness and objectivity of the sample data.

[0165] In a specific embodiment, the initial graph data described above can include nodes connected based on the association relationship between the multimedia resources corresponding to the positive sample data, the nodes being the positive sample data, and the positive sample data being resource identifier information of the multimedia resources on which the sample object has performed the preset operation in the plurality of preset time periods. Optionally, as shown in Figure 4 The initial graph data can be obtained by the following steps:

[0166] In step S401, co-occurrence analysis is performed on the resource identifier information to determine the co-occurrence relationship between the two resource identifier information.

[0167] In step S403, the association relationship between the multimedia resources corresponding to the positive sample data is determined based on the co-occurrence relationship.

[0168] In step S405, the resource identifier information of the multimedia resources corresponding to the positive sample data is taken as a node, and the node is connected based on the association relationship to obtain the initial graph data.

[0169] In a specific embodiment, in the case where the positive sample data is resource identifier information of the multimedia resources on which the sample object has performed the preset operation in the plurality of preset time periods, the co-occurrence relationship can represent the proportion information of the pair of resource identifier information corresponding to the two multimedia resources on which the preset operation is performed by the same object in the same preset time period in the positive sample data.

[0170] In actual application, the positive sample data can include multiple sample data groups in different preset time periods of different objects; optionally, the co-occurrence analysis on the resource identifier information can include determining resource identifier information pairs (two resource identifier information) in the same sample data group, determining, for each resource identifier information pair, a quantity of sample data groups including the resource identifier information pair, dividing the quantity by a total quantity of sample data groups to obtain a co-occurrence relationship corresponding to each resource identifier information pair; correspondingly, the co-occurrence relationship can include a resource identifier information pair corresponding to two multimedia resources that have been executed by the same object in the same preset time period, a corresponding sample data group, and a proportion information in the positive sample data corresponding sample data group. Optionally, for resource identifier information pairs (two resource identifier information) that do not co-occur in any sample data group, 0 can be taken as the corresponding co-occurrence relationship.

[0171] In one specific embodiment, in the process of determining the association relationship between the multimedia resources corresponding to the positive sample data based on the co-occurrence relationship, it can be determined that the resource identifier information pair with the proportion information corresponding to the co-occurrence relationship greater than a preset threshold has an association relationship, and vice versa, that the resource identifier information pair with the proportion information corresponding to the co-occurrence relationship less than or equal to the preset threshold does not have an association relationship; specifically, the preset threshold can be set in combination with actual application requirements. Further, taking the resource identifier information of the multimedia resource corresponding to the positive sample data as a node and connecting the nodes with an association relationship, the initial graph data is obtained.

[0172] In the above embodiments, the initial graph data representing the association relationship between the multimedia resources is constructed in combination with the operation information of the sample object on the multimedia resources in multiple preset time periods, which can effectively ensure the authenticity and objectivity of the data.

[0173] In one optional embodiment, the initial graph data can be input into a preset graph neural network for feature extraction to obtain initial node information corresponding to the nodes in the initial graph data. In one specific embodiment, the initial node information can represent the initial node information of the multimedia resource corresponding to the node. Specifically, the preset graph neural network can be a pre-set graph neural network.

[0174] In one specific embodiment, the data denoising network is trained based on the initial node information, the first sample resource information, the second sample resource information, and the object information, which can include: inputting the initial node information, the first sample resource information, the second sample resource information, and the object information into the preset data denoising network for confidence identification processing to obtain predicted confidence data; determining second loss information according to the predicted confidence data and corresponding preset confidence data; training the preset data denoising network based on the second loss information to obtain the data denoising network.

[0175] In one specific embodiment, the prediction confidence data can be the probability that the sample object will perform the preset operation on the multimedia resource corresponding to the sample data, which is predicted by the preset data denoising network. The above-mentioned preset confidence data can represent the probability that the sample object will perform the preset operation on the multimedia resource corresponding to the sample data, which is determined in combination with the obtained sample data (positive sample data and negative sample data). Specifically, the preset confidence data corresponding to the positive sample data can be 1; and the preset confidence data corresponding to the negative sample data can be 0.

[0176] In one specific embodiment, according to the prediction confidence data and the corresponding preset confidence data, determining the second loss information can include calculating the second loss information between the prediction confidence data and the corresponding preset confidence data based on a preset loss function. Further, the network parameters of the preset data denoising network can be adjusted based on the second loss information, and the steps of confidence recognition processing and second loss information determination based on the adjusted preset data denoising network can be repeated until the above-mentioned second loss information meets the second preset condition.

[0177] In one specific embodiment, the preset loss function can include but is not limited to a cross-entropy loss function, a logistic loss function, an exponential loss function, etc. The second loss information meeting the second preset condition can be that the second loss information is less than or equal to a specified threshold, or the difference between the corresponding second loss information in the previous and subsequent training processes is less than a certain threshold. In the embodiments of the present specification, the specified threshold and the certain threshold can be set in combination with the actual training requirements.

[0178] In one specific embodiment, the above-mentioned preset data denoising network can include but is not limited to a pre-set NN (Neural Network, neural network) and CNN (Convolutional Neural Networks, convolutional neural network) convolutional neural network, etc. deep learning network.

[0179] In the above-mentioned embodiments, during the data denoising training process, the confidence recognition processing is performed in combination with the initial node information, the first sample resource information, the second sample resource information and the object information, which can improve the data denoising effect and effectiveness of the trained data denoising network.

[0180] In one specific embodiment, the above-mentioned denoising processing of the positive sample data based on the data denoising network to obtain the target sample data can include: inputting the object information of the sample object corresponding to the positive sample data, the corresponding initial node information and the first sample resource information into the data denoising network to perform confidence recognition processing to obtain target confidence data; and resampling the positive sample data based on the target confidence data to obtain the target sample data.

[0181] In a specific embodiment, the target confidence data is predicted by the data denoising network, and can represent a probability that the sample object will perform a preset operation on the multimedia resource corresponding to the positive sample data.

[0182] In a specific embodiment, resampling the positive sample data based on the target confidence data to obtain target sample data can include: performing normalization processing on the target confidence data to obtain a sampling probability of each multimedia resource corresponding to the positive sample data, and resampling the positive sample data based on the sampling probability to obtain the target sample data.

[0183] In a specific embodiment, suppose that a multimedia resource viewed by an object in the positive sample data includes multimedia resource A, multimedia resource B, and multimedia resource C. The object has viewed (browsed) 1, 2, and 3 times, respectively. Suppose that the target confidence data (predicted by the data denoising network, the probability that the object views multimedia resource A, multimedia resource B, and multimedia resource C) corresponding to multimedia resource A, multimedia resource B, and multimedia resource C is 0.5, 0.8, and 0.9, respectively. Optionally, the sampling probability corresponding to multimedia resource A can be 0.5 / (0.5+0.8+0.9)=5 / 22, the sampling probability corresponding to multimedia resource B can be 0.8 / (0.5+0.8+0.9)=8 / 22, and the sampling probability corresponding to multimedia resource C can be 0.9 / (0.5+0.8+0.9)=9 / 22. Correspondingly, based on the sampling probabilities 5 / 22, 8 / 22, and 9 / 22, the data used for resampling can include resource identifier information of 5 multimedia resource A, resource identifier information of 8 multimedia resource B, and resource identifier information of 9 multimedia resource C. Suppose that 0 resource identifier information of multimedia resource A, 2 resource identifier information of multimedia resource B, and 4 resource identifier information of multimedia resource C are collected by randomly sampling the resource identifier information of 5 multimedia resource A, 8 multimedia resource B, and 9 multimedia resource C for 6 (1+2+3) times. Correspondingly, the resource identifier information of multimedia resource A in the positive sample data can be filtered out to obtain the target sample data.

[0184] In the above embodiments, resampling in combination with the target confidence data can better include various real situations, greatly achieve the denoising effect of the sample resource data, and ensure the effectiveness and accuracy of the target sample data.

[0185] In another optional embodiment, as shown in Figure 5 the above denoising processing of the positive sample data to obtain the target sample data can include:

[0186] In step S501, the behavior duration information of the sample object on the multimedia resource corresponding to the positive sample data is obtained.

[0187] In step S503, a sampling probability corresponding to the positive sample data is generated based on the behavior duration information.

[0188] In step S505, the positive sample data is resampled based on the sampling probability to obtain the target sample data.

[0189] In one specific embodiment, the behavior duration information can be the duration of the sample object performing a preset operation on the multimedia resource corresponding to the positive sample data. In one specific embodiment, assuming that the preset operation is browsing, the corresponding behavior duration information can be the duration of the sample object browsing the multimedia resource corresponding to the positive sample data. In one specific embodiment, assuming that the multimedia resources viewed by an object in the positive sample data include multimedia resource D, multimedia resource E, and multimedia resource F, and the object has viewed them 1, 2, and 3 times, respectively; assuming that the behavior duration information corresponding to multimedia resource D, multimedia resource E, and multimedia resource F is 5 minutes, 8 minutes, and 9 minutes, respectively; optionally, the sampling probability corresponding to multimedia resource D can be 5 / (5+8+9)=5 / 22; the sampling probability corresponding to multimedia resource E can be 8 / (5+8+9)=8 / 22; and the sampling probability corresponding to multimedia resource F can be 9 / (5+8+9)=9 / 22. Correspondingly, based on the sampling probabilities 5 / 22, 8 / 22, and 9 / 22, the data used for resampling can include resource identifier information of 5 multimedia resource A, resource identifier information of 8 multimedia resource B, and resource identifier information of 9 multimedia resource C; assuming that 6(1+2+3) random samplings are performed on the resource identifier information of 5 multimedia resource A, the resource identifier information of 8 multimedia resource B, and the resource identifier information of 9 multimedia resource C, 0 resource identifier information of multimedia resource D, 2 resource identifier information of multimedia resource E, and 4 resource identifier information of multimedia resource F are collected, respectively; correspondingly, the resource identifier information of multimedia resource D in the positive sample data can be filtered out to obtain the target sample data.

[0190] In the above embodiments, resampling is performed in combination with the behavior duration information of the multimedia resource corresponding to the positive sample data, which can better include various real situations, greatly achieve the denoising effect of the sample resource data, and ensure the effectiveness and accuracy of the target sample data.

[0191] In step S205, the target graph neural network is retrained based on the target sample data and the target object information of the sample object corresponding to the target sample data.

[0192] In one specific embodiment, as Figure 6As shown, the reconstruction training of the preset graph neural network based on the target sample data and the target object information of the sample object corresponding to the target sample data can include the following steps:

[0193] In step S601, the first target graph data is generated based on the association relationship between the target sample data and the multimedia resources.

[0194] In step S603, the first target graph data is input into the preset graph neural network for feature extraction to obtain the first target node information corresponding to the nodes in the first target graph data.

[0195] In step S605, the object information of the sample object corresponding to the target sample data and the first target node information are input into the preset generation network for sample reconstruction processing to obtain the first reconstructed sample data.

[0196] In step S607, the first loss information is determined according to the first reconstructed sample data and the target sample data.

[0197] In step S609, the preset graph neural network is reconstructed and trained based on the first loss information to obtain the target graph neural network.

[0198] In one specific embodiment, the specific refinement steps of generating the target graph data based on the association relationship between the target sample data and the multimedia resources can refer to the related step refinement of generating the initial graph data, which will not be repeated here.

[0199] In one specific embodiment, the first reconstructed sample data can be the resource identifier information of the multimedia resource in which the sample object has performed the preset operation in a plurality of preset time periods predicted by the preset generation network. Specifically, the target node information corresponds to a large number of multimedia resources. Optionally, the target node information can be subjected to average pooling processing and then input into the preset generation network to ensure that the size of the information input into the preset generation network meets the size requirement of the preset generation network for the input information.

[0200] In one specific embodiment, the specific refinement of determining the first loss information according to the first reconstructed sample data and the target sample data can refer to the related step refinement of determining the second loss information according to the prediction confidence data and the corresponding preset confidence data, which will not be repeated here.

[0201] In an optional embodiment, the reconstructing training of the preset graph neural network based on the first loss information to obtain the target graph neural network can include: in a case where the first loss information does not satisfy a first preset condition, adjusting network parameters of the preset generation network and the preset graph neural network based on the first loss information; updating the initial node information based on the preset generation network after the network parameters are adjusted; repeating the data denoising training of the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information and the object information to obtain the data denoising network until the step of determining the first loss information according to the first reconstructed sample data and the target sample data; and in a case where the first loss information satisfies the first preset condition, taking the preset graph neural network corresponding to the case where the first preset condition is satisfied as the target graph neural network.

[0202] In a specific embodiment, the initial graph data can be input into the preset generation network after the network parameters are adjusted to update the initial node information.

[0203] In a specific embodiment, the preset generation network can be a pre-set neural network. Specifically, the neural network can include an encoding network, a decoding network and a softmax layer.

[0204] In a specific embodiment, as shown in FIG. 7, Figure 7 Figure 7 is a schematic diagram of a graph neural network training process according to an exemplary embodiment. Specifically, in the graph neural network training process, the initial node information, the first sample resource information, the second sample resource information and the object information are input into the preset data denoising network 701 to train the data denoising network 702, then the target sample data is obtained by resampling the positive sample data in combination with the data denoising network 702; then, the target graph data generated from the target sample data is input into the preset graph neural network 703 for feature extraction to obtain the first target node information corresponding to the nodes in the target graph data; then, the first target node information and the object information of the corresponding sample object are input into the preset generation network 704 for sample reconstruction processing to obtain the first reconstructed sample data; and the first loss information is determined based on the first reconstructed sample data and the target sample data, and the preset graph neural network is reconstructed trained based on the first loss information. In the training process, the initial node information input into the preset data denoising network is updated in combination with the continuously updated preset graph neural network, and the iteration is performed in sequence until the first loss information satisfies the first preset condition.

[0205] ​In a specific embodiment, the first loss information satisfying the first preset condition can be that the first loss information is less than or equal to a specified threshold, or the difference between the corresponding first loss information in the previous and subsequent training processes is less than a certain threshold. In the embodiments of the present specification, the specified threshold and the certain threshold can be set in combination with actual training needs.

[0206] In the above embodiments, during the training of the graph neural network, the pre-set graph neural network is retrained in combination with the target graph data of the denoised target sample data and the object information of the corresponding sample object. During the training process, the initial node information of the input data denoising network is updated in combination with the continuously updated pre-set graph neural network, and iteration is continuously performed, which can greatly improve the representation accuracy of the graph neural network for multimedia resources, and further improve the recommendation accuracy in subsequent applications.

[0207] In an optional embodiment, the pre-set graph neural network is retrained based on the target sample data and the target object information of the sample object corresponding to the target sample data to obtain the target graph neural network, which includes:

[0208] Based on the association relationship between the target sample data and the corresponding multimedia resources, the second target graph data is generated;

[0209] The second target graph data is input into the pre-set graph neural network for feature extraction to obtain the second target node information corresponding to the nodes in the second target graph data;

[0210] The object information of the sample object corresponding to the target sample data and the second target node information are input into the pre-set generation network for sample reconstruction processing to obtain the second reconstructed sample data;

[0211] The third loss information is determined according to the second reconstructed sample data and the target sample data;

[0212] The pre-set graph neural network is retrained based on the third loss information to obtain the target graph neural network.

[0213] In a specific embodiment, the specific refinement of the above steps S601 to S607 from generating the second target graph data based on the association relationship between the target sample data and the corresponding multimedia resources to determining the third loss information according to the second reconstructed sample data and the target sample data can be referred to, and will not be repeated here.

[0214] In an optional embodiment, the reconstructing and training the preset graph neural network based on the third loss information to obtain the target graph neural network can include: in a case where the third loss information does not satisfy a third preset condition, adjusting network parameters of the preset generation network and the preset graph neural network based on the third loss information; updating the second target node information based on the preset generation network after the network parameters are adjusted; repeating the steps of inputting the object information of the sample object corresponding to the target sample data and the second target node information into the preset generation network for sample reconstruction processing to obtain the second reconstructed sample data, determining the third loss information according to the second reconstructed sample data and the target sample data, in a case where the third loss information satisfies the third preset condition, taking the preset graph neural network corresponding to the case where the third preset condition is satisfied as the target graph neural network.

[0215] In the above embodiments, in the graph neural network training process, the preset graph neural network is reconstructed and trained in combination with the target graph data of the denoised target sample data and the object information of the corresponding sample object, which can greatly improve the representation accuracy of the graph neural network for multimedia resources, and further improve the recommendation accuracy in subsequent applications.

[0216] As can be seen from the technical solutions provided by the above embodiments of the present specification, in the graph neural network training process, the positive sample data is denoised, which can improve the effectiveness and objectivity of the sample data, and the preset graph neural network is reconstructed and trained in combination with the target graph data of the denoised target sample data and the object information of the corresponding sample object, which can greatly improve the representation accuracy of the graph neural network for multimedia resources, and further improve the recommendation accuracy in subsequent applications.

[0217] Based on the target graph neural network obtained by the above graph neural network training method, the present disclosure further provides a task processing method, Figure 8 is a flowchart of a task processing method according to an exemplary embodiment, which can be used in a terminal or a server, such as Figure 8 as shown, can include the following steps.

[0218] In step S801, the preset graph data is input into the target graph neural network for feature extraction to obtain resource feature information corresponding to nodes in the preset graph data.

[0219] In a specific embodiment, the preset graph data includes nodes connected based on the association relationship between the plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources. In an optional embodiment, in the case where the resource information is resource identifier information, the preset graph data can include the initial graph data. In the case where the resource information further includes resource attribute information, the multimedia resources corresponding to the preset graph data can be different from the multimedia corresponding to the initial graph data.

[0220] In step S803, the target task is performed based on the resource feature information.

[0221] In actual applications, the target task can be different according to actual application requirements. In an optional embodiment, in the case where the target task includes an information recall task, the target task can be performed based on the resource feature information as shown in FIG. 8. Figure 9 The target task can be performed based on the resource feature information as shown in FIG. 8.

[0222] In step S901, the similarity between the plurality of multimedia resources is determined according to the resource feature information.

[0223] In step S903, the associated multimedia resource corresponding to each multimedia resource is determined according to the similarity.

[0224] In step S905, when performing information recall processing based on any multimedia resource of the plurality of multimedia resources, information recall processing is performed based on the corresponding associated multimedia resource.

[0225] In an optional embodiment, the multimedia resource with a similarity greater than or equal to a preset similarity threshold value with any multimedia resource can be used as the associated multimedia resource corresponding to the multimedia resource. In another optional embodiment, the multimedia resources can be sorted in descending order according to the similarity. Correspondingly, for any multimedia resource, the multimedia resources in the top preset number of positions can be used as the associated multimedia resource corresponding to the multimedia resource.

[0226] In a specific embodiment, when a user browses multimedia resources such as videos, the server receives a recommendation request. Correspondingly, the n multimedia resources recently browsed by the user can be obtained in real time, and the n multimedia resources and the associated multimedia resources corresponding to the n multimedia resources can be used as recall results for information recommendation.

[0227] In the above embodiments, the resource information of the multimedia resource extracted by the target graph neural network can effectively improve the accuracy of the representation of the resource information of the multimedia resource, thereby ensuring that the recalled multimedia resources are more in line with the user's demand, and greatly improving the recall accuracy and recommendation accuracy.

[0228] In an optional embodiment, in the case that the target task includes a training task of the target network, accordingly, as shown in FIG. 10, the performing the target task based on the resource feature information can include the following steps: Figure 10

[0229] In step S1001, the preset annotation information of the multimedia resource corresponding to the resource feature information is acquired.

[0230] In step S1003, the resource feature information is input into the preset neural network for associated task processing to obtain predicted task result information.

[0231] In step S1005, the third loss information is determined according to the preset annotation information and the predicted task result information.

[0232] In step S1007, the preset neural network is trained based on the third loss information to obtain the target network.

[0233] In an embodiment, the preset annotation information can represent the preset task result information of the associated task corresponding to the target network. Specifically, the associated task corresponding to the target network can be a task corresponding to the use of the target network. In an embodiment, the target network is used for classifying multimedia resources, and accordingly, the associated task corresponding to the target network can be a classification task of the multimedia resources, and accordingly, the preset annotation information can be a classification label (such as sports multimedia resources, food multimedia resources, game multimedia resources, etc.). In another embodiment, the target network is used for quality analysis of multimedia resources, and accordingly, the associated task corresponding to the target network can be a quality analysis task of the multimedia resources, and accordingly, the preset annotation information can be a quality analysis result. Specifically, the quality analysis result can be a character representation capable of representing the quality information of the multimedia resources, for example, high, medium, and low; or a numerical value quantified based on a certain rule from the character representation, for example, 90, 60, and 30.

[0234] In an embodiment, the specific refinement of determining the third loss information according to the preset annotation information and the predicted task result information can refer to the specific refinement of determining the first loss information according to the reconstructed sample data and the target sample data, which will not be repeated here.

[0235] ​In a specific embodiment, the network parameters of the preset neural network can be adjusted based on the third loss information, and the steps of determining the third loss information and performing the association task processing based on the adjusted preset neural network can be repeated until the third loss information meets a third preset condition. Optionally, the third loss information meeting the third preset condition can be that the third loss information is less than or equal to a specified threshold, or the difference between the third loss information in the previous training process and the third loss information in the current training process is less than a certain threshold. In the embodiments of the present specification, the specified threshold and the certain threshold can be set according to actual training needs.

[0236] In the above embodiments, the resource information of the multimedia resource extracted by the target graph neural network can effectively improve the accuracy of the representation of the resource information to the multimedia resource, thereby ensuring that the subsequent target network can better learn the features of the multimedia resource, greatly improving the learning efficiency and learning effect of the target network.

[0237] As can be seen from the technical solutions provided by the embodiments of the present specification, the resource information of the multimedia resource extracted by the target graph neural network can effectively improve the accuracy of the representation of the resource information to the multimedia resource, thereby ensuring that the features of the multimedia resource are accurately grasped during the subsequent target task execution process, greatly improving the accuracy and efficiency of the related target task execution.

[0238] Figure 11 is a block diagram of a graph neural network training device according to an example embodiment. Referring to Figure 11 The device comprises:

[0239] The first data acquisition module 1110 is configured to perform acquisition of object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of a multimedia resource in which the sample object has performed a preset operation in a plurality of preset time periods;

[0240] The denoising processing module 1120 is configured to perform denoising processing on the positive sample data to obtain target sample data;

[0241] The retraining module 1130 is configured to perform retraining of a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data, to obtain a target graph neural network.

[0242] Optionally, the denoising processing module 1120 is specifically configured to perform denoising processing on the positive sample data based on a data denoising network to obtain the target sample data.

[0243] Optionally, the device further comprises:

[0244] The second data acquisition module is configured to acquire the first sample resource information of the positive sample data corresponding to the multimedia resource, the second sample resource information of the negative sample data corresponding to the multimedia resource, and initial graph data of the sample object, the initial graph data including nodes connected based on the association relationship between the positive sample data corresponding to the multimedia resource, and the node being the positive sample data;

[0245] The first feature extraction module is configured to input the initial graph data into a preset graph neural network for feature extraction to obtain initial node information corresponding to the nodes in the initial graph data;

[0246] The data denoising training module is configured to perform data denoising training on a preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain the data denoising network.

[0247] Optionally, the reconstruction training module 1130 includes:

[0248] The first target graph data generation unit is configured to generate first target graph data based on the association relationship between the target sample data corresponding to the multimedia resource;

[0249] The first feature extraction unit is configured to input the first target graph data into a preset graph neural network for feature extraction to obtain first target node information corresponding to the nodes in the first target graph data;

[0250] The first sample reconstruction processing unit is configured to input the object information of the sample object corresponding to the target sample data and the first target node information into a preset generation network for sample reconstruction processing to obtain first reconstructed sample data;

[0251] The first loss information determination unit is configured to determine first loss information according to the first reconstructed sample data and the target sample data;

[0252] The first reconstruction training unit is configured to perform reconstruction training on the preset graph neural network based on the first loss information to obtain a target graph neural network.

[0253] Optionally, the first reconstruction training unit includes:

[0254] The first network parameter adjustment unit is configured to adjust the network parameters of the preset generation network and the preset graph neural network based on the first loss information in a case where the first loss information does not satisfy a first preset condition;

[0255] The initial node information updating unit is configured to update the initial node information based on the preset generation network after the network parameters are adjusted;

[0256] The first iterative training unit is configured to perform data denoising training on the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information and the object information, to obtain the data denoising network, and determine the first loss information according to the first reconstructed sample data and the target sample data.

[0257] The first target graph neural network determination unit is configured to perform, in a case where the first loss information meets the first preset condition, determination of a preset graph neural network corresponding to a time when the first preset condition is met as the target graph neural network.

[0258] Optionally, the denoising processing module 1120 includes:

[0259] The first confidence identification processing unit is configured to perform confidence identification processing on the data denoising network by inputting the object information of the sample object corresponding to the positive sample data, the corresponding initial node information and the first sample resource information, to obtain target confidence data, the target confidence data representing a probability that the sample object will perform the preset operation on the multimedia resource corresponding to the positive sample data.

[0260] The first resampling unit is configured to perform resampling on the positive sample data based on the target confidence data, to obtain target sample data.

[0261] Optionally, the resource information includes resource identifier information, and the initial graph data is obtained by using the following units:

[0262] The co-occurrence analysis unit is configured to perform co-occurrence analysis on the resource identifier information, to determine a co-occurrence relationship between two resource identifier information, the co-occurrence relationship representing a proportion of a pair of resource identifier information corresponding to two multimedia resources that are performed by the same object in the same preset time period in the positive sample data.

[0263] The association relationship determination unit is configured to determine an association relationship between the multimedia resources corresponding to the positive sample data based on the co-occurrence relationship.

[0264] The initial graph data construction unit is configured to perform, by taking the resource identifier information of the multimedia resources corresponding to the positive sample data as nodes, connection of the nodes based on the association relationship, to obtain the initial graph data.

[0265] Optionally, the data denoising training module includes:

[0266] The second confidence identification processing unit is configured to perform confidence identification processing on the preset data denoising network by inputting the initial node information, the first sample resource information, the second sample resource information and the object information, to obtain prediction confidence data, the prediction confidence data representing a probability that the sample object will perform the preset operation on the multimedia resource corresponding to the sample data.

[0267] The second loss information determination unit is configured to determine the second loss information according to the prediction confidence data and the corresponding preset confidence data;

[0268] The denoising training unit is configured to train the preset data denoising network based on the second loss information to obtain the data denoising network.

[0269] Optionally, the denoising processing module 1120 includes:

[0270] The behavior duration information acquisition unit is configured to acquire the behavior duration information of the sample object to the positive sample data corresponding multimedia resource;

[0271] The sampling probability generation unit is configured to generate the sampling probability corresponding to the positive sample data based on the behavior duration information;

[0272] The second resampling unit is configured to resample the positive sample data based on the sampling probability to obtain the target sample data.

[0273] Optionally, the reconstruction training module 1130 includes:

[0274] The second target graph data generation unit is configured to generate the second target graph data based on the association relationship between the target sample data corresponding multimedia resources;

[0275] The second feature extraction unit is configured to input the second target graph data into the preset graph neural network to perform feature extraction, and obtain the second target node information corresponding to the nodes in the second target graph data;

[0276] The second sample reconstruction processing unit is configured to input the object information of the sample object corresponding to the target sample data and the second target node information into the preset generation network to perform sample reconstruction processing, and obtain the second reconstructed sample data;

[0277] The third loss information determination unit is configured to determine the third loss information according to the second reconstructed sample data and the target sample data;

[0278] The second reconstruction training unit is configured to perform reconstruction training on the preset graph neural network based on the third loss information to obtain the target graph neural network.

[0279] Optionally, the second reconstruction training unit includes:

[0280] The second network parameter adjustment unit is configured to adjust the network parameters of the preset generation network and the preset graph neural network based on the third loss information in the case that the third loss information does not satisfy the third preset condition;

[0281] The target node information updating unit is configured to perform generating network updating second target node information based on the preset generated network after adjusting the network parameters;

[0282] The second iterative training unit is configured to perform the steps of repeatedly inputting object information of a sample object corresponding to target sample data and the second target node information into the preset generated network for sample reconstruction processing based on the updated second target node information, obtaining second reconstructed sample data, and determining third loss information according to the second reconstructed sample data and the target sample data.

[0283] The second target graph neural network determination unit is configured to perform, in a case where the third loss information meets a third preset condition, determining a preset graph neural network corresponding to the case where the third preset condition is met as the target graph neural network.

[0284] As to the apparatus in the above embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0285] Figure 12 is a block diagram of a task processing apparatus according to an example embodiment. Referring to Figure 12 , the task processing apparatus can include:

[0286] The second feature extraction module 1210 is configured to perform feature extraction on preset graph data by inputting the preset graph data into the target graph neural network obtained by the graph neural network training method according to any of the above first aspect, to obtain resource feature information corresponding to nodes in the preset graph data; the preset graph data includes nodes connected based on the association relationship between a plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources;

[0287] The target task execution module 1220 is configured to perform a target task based on the resource feature information.

[0288] Optionally, the target task includes an information recall task, and the target task execution module 1020 includes:

[0289] The similarity determination unit is configured to determine the similarity between the plurality of multimedia resources according to the resource feature information.

[0290] The associated multimedia resource determination unit is configured to determine the associated multimedia resource corresponding to each multimedia resource according to the similarity.

[0291] The information recall processing unit is configured to perform information recall processing based on the corresponding associated multimedia resource when performing information recall processing based on any of the plurality of multimedia resources.

[0292] Optionally, the target task includes a training task of the target network, and the target task execution module 1020 includes:

[0293] The preset annotation information acquisition unit is configured to perform acquisition of preset annotation information of the multimedia resource corresponding to the resource feature information, the preset annotation information representing preset task result information of an associated task corresponding to the target network.

[0294] The associated task processing unit is configured to perform input of the resource feature information into the preset neural network for associated task processing, to obtain predicted task result information.

[0295] The third loss information determination unit is configured to perform determination of third loss information according to the preset annotation information and the predicted task result information.

[0296] The network training unit is configured to perform training of the preset neural network based on the third loss information, to obtain the target network.

[0297] As to the apparatus in the above-described embodiments, the specific manners in which various modules perform operations have been described in details in the embodiments about the method, and thus will not be described in details here.

[0298] Figure 13 is a block diagram of an electronic device for graph neural network training according to an example embodiment. The electronic device can be a server, and its internal structure diagram can be as shown in Figure 13 The electronic device includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the electronic device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a graph neural network training method.

[0299] Figure 14 is a block diagram of an electronic device for task processing according to an example embodiment. The electronic device can be a terminal, and its internal structure diagram can be as shown in Figure 14As shown in the figure. The electronic device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the electronic device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a task processing method. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the electronic device, or an external keyboard, touchpad or mouse, etc.

[0300] Those skilled in the art can understand that, Figure 13 Or Figure 14 The structure shown in the figure is only a block diagram of part of the structure related to the present disclosure, and does not constitute a limitation on the electronic device to which the present disclosure is applied. The specific electronic device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0301] In an example embodiment, an electronic device is also provided, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the graph neural network training method or the task processing method as in the embodiments of the present disclosure.

[0302] In an example embodiment, a computer readable storage medium is also provided, when the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute the graph neural network training method or the task processing method in the embodiments of the present disclosure.

[0303] In an example embodiment, a computer program product containing instructions is also provided, when it is run on a computer, the computer executes the graph neural network training method or the task processing method in the embodiments of the present disclosure.

[0304] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0305] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the disclosure that are deemed to fall within the general principles of the disclosure and include commonly known or customary practice in the art. The specification and examples are to be considered exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

[0306] It should be understood that the present disclosure is not limited to the precise structures herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A method for training a graph neural network, the method comprising: The method comprises: obtaining object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of multimedia resources in which the sample object has performed a preset operation in a plurality of preset time periods; performing denoising processing on the positive sample data to obtain target sample data; reconstructing and training a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data, to obtain a target graph neural network; The reconstructing and training of the preset graph neural network based on the target sample data and the target object information of the sample object corresponding to the target sample data to obtain the target graph neural network comprises: generating first target graph data based on an association relationship between multimedia resources corresponding to the target sample data; inputting the first target graph data into the preset graph neural network to extract features, to obtain first target node information corresponding to nodes in the first target graph data; inputting object information of a sample object corresponding to the target sample data and the first target node information into a preset generation network to perform sample reconstruction processing, to obtain first reconstructed sample data; determining first loss information according to the first reconstructed sample data and the target sample data; and reconstructing and training the preset graph neural network based on the first loss information, to obtain the target graph neural network.

2. The graph neural network training method of claim 1, wherein, The denoising processing of the positive sample data to obtain the target sample data comprises: performing denoising processing on the positive sample data based on a data denoising network to obtain the target sample data.

3. The graph neural network training method of claim 2, wherein, The method further comprises: obtaining first sample resource information of multimedia resources corresponding to the positive sample data of the sample object, second sample resource information of multimedia resources corresponding to negative sample data, and initial graph data, the initial graph data comprising nodes connected based on an association relationship between multimedia resources corresponding to the positive sample data, the nodes being the positive sample data; inputting the initial graph data into a preset graph neural network to extract features, to obtain initial node information corresponding to nodes in the initial graph data; performing data denoising training on a preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information, to obtain the data denoising network.

4. The graph neural network training method of claim 3, wherein, The reconstructing and training of the preset graph neural network based on the first loss information to obtain the target graph neural network comprises: in a case where the first loss information does not satisfy a first preset condition, adjusting network parameters of the preset generation network and the preset graph neural network based on the first loss information; updating the initial node information based on the preset generation network after the network parameters are adjusted; repeating the steps of performing data denoising training on the preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information, to obtain the data denoising network, and determining the first loss information according to the first reconstructed sample data and the target sample data, based on the updated initial node information. In a case where the first loss information satisfies the first preset condition, a preset graph neural network corresponding to a case where the first preset condition is satisfied is taken as the target graph neural network.

5. The graph neural network training method of claim 3, wherein, The data denoising network is used to perform denoising processing on the positive sample data to obtain target sample data. Object information of a sample object corresponding to the positive sample data, initial node information corresponding to the positive sample data, and the first sample resource information are input into the data denoising network to perform confidence identification processing to obtain target confidence data, the target confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to the positive sample data. The target sample data is obtained by resampling the positive sample data based on the target confidence data.

6. The graph neural network training method of claim 3, wherein, The resource information includes resource identifier information, and the initial graph data is obtained in the following manner: Co-occurrence analysis is performed on the resource identifier information to determine a co-occurrence relationship between two resource identifier information, the co-occurrence relationship representing a proportion of a pair of resource identifier information corresponding to two multimedia resources that are performed by a same object in a same preset time period in the positive sample data; An association relationship between the multimedia resources corresponding to the positive sample data is determined based on the co-occurrence relationship; The resource identifier information of the multimedia resources corresponding to the positive sample data is taken as a node, and the node is connected based on the association relationship to obtain the initial graph data.

7. The graph neural network training method of claim 3, wherein, The data denoising network is trained based on the initial node information, the first sample resource information, the second sample resource information, and the object information to obtain a data denoising network, including: The initial node information, the first sample resource information, the second sample resource information, and the object information are input into the preset data denoising network to perform confidence identification processing to obtain predicted confidence data, the predicted confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to the sample data; Second loss information is determined according to the predicted confidence data and corresponding preset confidence data; The preset data denoising network is trained based on the second loss information to obtain the data denoising network.

8. The graph neural network training method of claim 1, wherein, The data denoising network is used to perform denoising processing on the positive sample data to obtain target sample data, including: Behavior duration information of the sample object on the multimedia resources corresponding to the positive sample data is obtained; A sampling probability corresponding to the positive sample data is generated based on the behavior duration information; The target sample data is obtained by resampling the positive sample data based on the sampling probability.

9. The graph neural network training method of claim 1, wherein, The target graph neural network is obtained by reconstructing training a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data, including: Second target graph data is generated based on an association relationship between the multimedia resources corresponding to the target sample data; Second target node information corresponding to nodes in the second target graph data is obtained by inputting the second target graph data into the preset graph neural network to perform feature extraction. inputting the object information of a sample object corresponding to the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data; determining third loss information according to the second reconstructed sample data and the target sample data; reconstructing and training the preset graph neural network based on the third loss information to obtain the target graph neural network.

10. The graph neural network training method of claim 9, wherein, The reconstructing and training the preset graph neural network based on the third loss information to obtain the target graph neural network includes: in a case where the third loss information does not satisfy a third preset condition, adjusting network parameters of the preset generation network and the preset graph neural network based on the third loss information; updating the second target node information based on the preset generation network after the network parameters are adjusted; repeating the steps of inputting the object information of a sample object corresponding to the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data to determining third loss information according to the second reconstructed sample data and the target sample data based on the updated second target node information; in a case where the third loss information satisfies the third preset condition, taking the preset graph neural network corresponding to the case where the third loss information satisfies the third preset condition as the target graph neural network.

11. A task processing method characterized by, comprises: inputting preset graph data into the target graph neural network obtained by the graph neural network training method to obtain resource feature information corresponding to nodes in the preset graph data; the preset graph data comprises nodes connected based on an association relationship between a plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources; performing a target task based on the resource feature information.

12. The task processing method according to claim 11, wherein The target task includes an information recall task, and performing the target task based on the resource feature information includes: determining a similarity between the plurality of multimedia resources according to the resource feature information; determining an associated multimedia resource corresponding to each multimedia resource according to the similarity; when performing information recall processing based on any multimedia resource in the plurality of multimedia resources, performing information recall processing based on the corresponding associated multimedia resource.

13. The task processing method of claim 11, wherein, The target task includes a training task of a target network, and performing the target task based on the resource feature information includes: obtaining preset annotation information of a multimedia resource corresponding to the resource feature information, the preset annotation information representing preset task result information of an associated task corresponding to the target network; inputting the resource feature information into a preset neural network for associated task processing to obtain predicted task result information; determining third loss information according to the preset annotation information and the predicted task result information; training the preset neural network based on the third loss information to obtain the target network.

14. A graph neural network training device, characterized in that, comprises: The first data acquisition module is configured to acquire object information of a sample object and positive sample data of the sample object, the positive sample data being resource information of multimedia resources on which the sample object has performed a preset operation in a plurality of preset time periods; The denoising processing module is configured to perform denoising processing on the positive sample data to obtain target sample data; The reconstruction training module is configured to perform reconstruction training on a preset graph neural network based on the target sample data and target object information of a sample object corresponding to the target sample data, to obtain a target graph neural network; The reconstruction training module includes: a first target graph data generation unit configured to generate first target graph data based on an association relationship between multimedia resources corresponding to the target sample data; a first feature extraction unit configured to input the first target graph data into the preset graph neural network to perform feature extraction, to obtain first target node information corresponding to nodes in the first target graph data; a first sample reconstruction processing unit configured to input object information of a sample object corresponding to the target sample data and the first target node information into a preset generation network to perform sample reconstruction processing, to obtain first reconstructed sample data; a first loss information determination unit configured to determine first loss information according to the first reconstructed sample data and the target sample data; and a first reconstruction training unit configured to perform reconstruction training on the preset graph neural network based on the first loss information, to obtain the target graph neural network.

15. The graph neural network training apparatus of claim 14, wherein, The denoising processing module is specifically configured to perform denoising processing on the positive sample data based on a data denoising network, to obtain target sample data.

16. The graph neural network training apparatus of claim 15, wherein, The device further includes: A second data acquisition module configured to acquire first sample resource information of multimedia resources corresponding to the positive sample data of the sample object, second sample resource information of multimedia resources corresponding to negative sample data, and initial graph data, the initial graph data including nodes connected based on an association relationship between the multimedia resources corresponding to the positive sample data, the nodes being the positive sample data; A first feature extraction module configured to input the initial graph data into a preset graph neural network to perform feature extraction, to obtain initial node information corresponding to nodes in the initial graph data; A data denoising training module configured to perform data denoising training on a preset data denoising network based on the initial node information, the first sample resource information, the second sample resource information, and the object information, to obtain the data denoising network.

17. The graph neural network training apparatus of claim 16, wherein, The first reconstruction training unit includes: A first network parameter adjustment unit configured to adjust network parameters of the preset generation network and the preset graph neural network based on the first loss information, in a case where the first loss information does not satisfy a first preset condition; An initial node information update unit configured to update the initial node information based on the preset generation network after the network parameters are adjusted; The first iteration training unit is configured to perform repeated data denoising training of the preset data denoising network based on the updated initial node information, to obtain the data denoising network, and to determine the first loss information according to the first reconstructed sample data and the target sample data. The first target graph neural network determination unit is configured to perform, in a case where the first loss information meets the first preset condition, determination of a preset graph neural network corresponding to the first preset condition as the target graph neural network.

18. The graph neural network training apparatus of claim 16, wherein, The denoising processing module includes: The first confidence identification processing unit is configured to perform confidence identification processing of the data denoising network by inputting the object information of the sample object corresponding to the positive sample data, the corresponding initial node information, and the first sample resource information, to obtain target confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to the positive sample data. The first resampling unit is configured to perform resampling of the positive sample data based on the target confidence data to obtain the target sample data.

19. The graph neural network training apparatus of claim 16, wherein, The resource information includes resource identifier information, and the initial graph data is obtained by using the following unit: The co-occurrence analysis unit is configured to perform co-occurrence analysis on the resource identifier information to determine a co-occurrence relationship between two resource identifier information, and the co-occurrence relationship represents a proportion of a pair of resource identifier information corresponding to two multimedia resources that are performed by the same object in the same preset time period in the positive sample data. The association relationship determination unit is configured to determine an association relationship between the multimedia resources corresponding to the positive sample data based on the co-occurrence relationship. The initial graph data construction unit is configured to construct the initial graph data by taking the resource identifier information of the multimedia resources corresponding to the positive sample data as nodes and connecting the nodes based on the association relationship.

20. The graph neural network training apparatus of claim 16, wherein, The data denoising training module includes: The second confidence identification processing unit is configured to perform confidence identification processing of the preset data denoising network by inputting the initial node information, the first sample resource information, the second sample resource information, and the object information, to obtain predicted confidence data representing a probability that the sample object will perform a preset operation on a multimedia resource corresponding to the sample data. The second loss information determination unit is configured to determine second loss information according to the predicted confidence data and the corresponding preset confidence data. The denoising training unit is configured to perform training of the preset data denoising network based on the second loss information to obtain the data denoising network.

21. The graph neural network training apparatus of claim 14, wherein, The denoising processing module includes: The behavior duration information acquisition unit is configured to acquire behavior duration information of the sample object on the multimedia resource corresponding to the positive sample data. The sampling probability generation unit is configured to generate a sampling probability corresponding to the positive sample data based on the behavior duration information. A second resampling unit, configured to perform resampling on the positive sample data based on the sampling probability to obtain the target sample data.

22. The graph neural network training apparatus of claim 14, wherein, The reconstruction training module comprises: A second target graph data generation unit, configured to perform generation of second target graph data based on an association relationship between multimedia resources corresponding to the target sample data; A second feature extraction unit, configured to perform input of the second target graph data into the preset graph neural network for feature extraction to obtain second target node information corresponding to nodes in the second target graph data; A second sample reconstruction processing unit, configured to perform input of object information of a sample object corresponding to the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data; A third loss information determination unit, configured to perform determination of third loss information according to the second reconstructed sample data and the target sample data; A second reconstruction training unit, configured to perform reconstruction training on the preset graph neural network based on the third loss information to obtain the target graph neural network.

23. The graph neural network training apparatus of claim 22, wherein, The second reconstruction training unit comprises: A second network parameter adjustment unit, configured to perform adjustment of network parameters of the preset generation network and the preset graph neural network based on the third loss information in a case where the third loss information does not satisfy a third preset condition; A target node information updating unit, configured to perform updating of the second target node information based on the preset generation network after adjustment of network parameters; A second iterative training unit, configured to perform repetition of the steps of input of object information of a sample object corresponding to the target sample data and the second target node information into a preset generation network for sample reconstruction processing to obtain second reconstructed sample data and the determination of third loss information according to the second reconstructed sample data and the target sample data based on the updated second target node information; A second target graph neural network determination unit, configured to perform determination of a preset graph neural network corresponding to satisfaction of the third preset condition as the target graph neural network in a case where the third loss information satisfies the third preset condition.

24. A task processing apparatus characterized by comprising: Comprise: A second feature extraction module, configured to perform feature extraction of preset graph data into a target graph neural network obtained by the graph neural network training method according to any one of claims 1 to 10 to obtain resource feature information corresponding to nodes in the preset graph data; the preset graph data comprises nodes connected based on an association relationship between a plurality of multimedia resources, and the nodes are resource information corresponding to the plurality of multimedia resources; A target task execution module, configured to perform a target task based on the resource feature information.

25. The task processing apparatus according to claim 24, wherein The target task comprises an information recall task, and the target task execution module comprises: A similarity determination unit, configured to perform determination of similarity between the plurality of multimedia resources according to the resource feature information; An associated multimedia resource determination unit, configured to perform determination of associated multimedia resources corresponding to each multimedia resource according to the similarity; The information recall processing unit is configured to perform information recall processing based on the corresponding associated multimedia resource when performing information recall processing based on any multimedia resource of the plurality of multimedia resources.

26. The task processing apparatus according to claim 24, wherein The target task includes a training task of a target network, and the target task execution module includes: The preset annotation information acquisition unit is configured to perform acquisition of preset annotation information of the multimedia resource corresponding to the resource feature information, the preset annotation information representing preset task result information of an associated task corresponding to the target network; The associated task processing unit is configured to perform input of the resource feature information into a preset neural network for associated task processing to obtain predicted task result information; The third loss information determination unit is configured to perform determination of third loss information according to the preset annotation information and the predicted task result information; The network training unit is configured to perform training of the preset neural network based on the third loss information to obtain the target network.

27. An electronic device, comprising: comprise: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the graph neural network training method of any one of claims 1-10 or the task processing method of any one of claims 11-13.

28. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can perform the graph neural network training method of any one of claims 1-10 or the task processing method of any one of claims 11-13.

29. A computer program product comprising computer instructions, characterized in that, The computer instructions are executed by the processor to implement the graph neural network training method of any one of claims 1-10 or the task processing method of any one of claims 11-13.

Images