CN115222044A

CN115222044A - Model training method, graph data processing method, device, equipment and storage medium

Info

Publication number: CN115222044A
Application number: CN202210827890.8A
Authority: CN
Inventors: 孙嘉齐; 赵胜林; 白星宇; 石书玮
Original assignee: Shenzhen Tencent Information Technology Co Ltd
Current assignee: Shenzhen Tencent Information Technology Co Ltd
Priority date: 2022-07-13
Filing date: 2022-07-13
Publication date: 2022-10-21

Abstract

The application provides a model training method, a graph data processing method, a device, equipment and a storage medium; the method comprises the following steps: acquiring a preset graph neural network model and training data; determining a training expression vector of each training node by using a graph neural network model; determining a complementary graph structure of the training graph structure and a first adjacent matrix of the complementary graph structure, and acquiring a second adjacent matrix of the training graph structure; determining a first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors of the respective training nodes; determining a joint loss function based on a second loss function component preset by the graph neural network model and the first loss function component; and training the graph neural network model by using the joint loss function, the training expression vectors of the training nodes and the training labels of the training nodes to obtain the trained graph neural network model. By the method and the device, comprehensiveness of the graph neural network model to signal description can be improved.

Description

Model training method, graph data processing method, device, equipment and storage medium

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a method, an apparatus, a device, and a computer-readable storage medium for processing graph data.

Background

In recent years, graph-structured data has become a research hotspot in the field of data mining, such as social networks, citation networks, partner networks, knowledge graphs, recommendation systems, and the like. Graph data is made up of a number of nodes and a number of edges connecting between the nodes, and the use of a full understanding of such data does not depart from generating a representation vector for each of the nodes therein. In the related technology, when the expression vector of the node is determined by using a graph convolution network and a graph attention network, high-frequency signal components can not be captured, and the classification information of different nodes tends to be consistent with the increase of the layer number, so that the problem of over-smoothness occurs, and the accuracy of downstream task processing is greatly reduced; for example, when a task is recommended, the same object is recommended to users with large differences, and accurate recommendation cannot be achieved.

Disclosure of Invention

The embodiment of the application provides a model training method, a graph data processing device and a storage medium, and can improve comprehensiveness of a neural network model on signal description.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides a model training method, which comprises the following steps:

acquiring a preset graph neural network model and training data, wherein the training data comprises a training graph structure and label information of each training node in the training graph structure;

determining a training representation vector of each training node by using the graph neural network model;

determining a complementary graph structure of the training graph structure and a first adjacency matrix of the complementary graph structure, and acquiring a second adjacency matrix of the training graph structure;

determining a first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors for the respective training nodes;

determining a joint loss function based on a second loss function component preset by the graph neural network model and the first loss function component;

and training the graph neural network model by using the joint loss function, the training expression vectors of the training nodes and the training labels of the training nodes to obtain the trained graph neural network model.

An embodiment of the present application provides a graph data processing method, including:

responding to a received task processing request, acquiring to-be-processed graph structure data and a trained graph neural network model corresponding to the task processing request, wherein the trained graph neural network model is obtained by utilizing a model training method provided by the embodiment of the application;

predicting the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in the graph structure, wherein the representation vector can represent the global characteristics and the local characteristics of the node;

performing task processing based on the expression vectors of the nodes to obtain a processing result;

and outputting the processing result.

The embodiment of the application provides a model training device, includes:

the device comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring a preset graph neural network model and training data, and the training data comprises a training graph structure and label information of each training node in the training graph structure;

a first determining module, configured to determine a training representation vector of each training node by using the graph neural network model;

a second determining module, configured to determine a complementary graph structure of the training graph structure and a first adjacent matrix of the complementary graph structure, and obtain a second adjacent matrix of the training graph structure;

a third determining module for determining a first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors of the respective training nodes;

a fourth determining module, configured to determine a joint loss function based on a second loss function component preset by the graph neural network model and the first loss function component;

and the first training module is used for training the graph neural network model by using the joint loss function, the training expression vectors of the training nodes and the training labels of the training nodes to obtain the trained graph neural network model.

An embodiment of the present application provides a graph data processing apparatus, including:

the second obtaining module is configured to obtain, in response to a received task processing request, to-be-processed graph structure data and a trained graph neural network model corresponding to the task processing request, where the trained graph neural network model is obtained by training using the model training method provided in the embodiment of the present application;

the first prediction module is used for performing prediction processing on the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in a graph structure, and the representation vector can represent the global characteristics and the local characteristics of the node;

the first processing module is used for carrying out task processing based on the expression vectors of all the nodes to obtain a processing result;

and the result output module is used for outputting the processing result.

An embodiment of the present application provides a computer device, including:

a memory for storing executable instructions;

and the processor is used for realizing the model training method or the graph data processing method provided by the embodiment of the application when executing the executable instructions stored in the memory.

The embodiment of the present application provides a computer-readable storage medium, which stores executable instructions for causing a processor to implement the model training method or the graph data processing method provided by the embodiment of the present application when executed.

The embodiments of the present application provide a computer program product, which includes a computer program or instructions, and when the computer program or instructions are executed by a processor, the computer program or instructions implement the model training method or the graph data processing method provided by the embodiments of the present application.

The model training method provided by the embodiment of the application has the following beneficial effects:

in the training process of the graph neural network model, a high-pass filter is not required to be added in each layer, the trained graph neural network model does not need to have the capacity of expressing high-frequency signals through calculation of a high-order polynomial of a Laplace matrix, the loss function used for training the graph neural network model is improved on the basis of not changing the graph neural network model, a first loss function component is added on the basis of a preset second loss function component, the first loss function component comprises a first regular sub-item determined through a first adjacent matrix of a complementary graph structure corresponding to the training graph structure and a second regular sub-item determined through a second adjacent matrix corresponding to the training graph structure, and the first regular sub-item can represent one high-pass filter, so that the trained graph neural network model can have the capacity of expressing high-frequency signals through the first loss function component, and therefore the low-frequency neural network model capable of providing the high-frequency signals and the high-frequency signals can be efficiently trained on the premise of not changing the number of model parameters and the complexity of calculation.

Drawings

FIG. 1 is a block diagram illustrating a network architecture of a data processing system 100, according to an embodiment of the present application;

fig. 2 is a schematic structural diagram of a server 400 provided in an embodiment of the present application;

FIG. 3 is a schematic diagram of an implementation flow of model training provided in an embodiment of the present application;

FIG. 4 is a schematic flow chart of an implementation of a graph data processing method provided in an embodiment of the present application;

fig. 5 is a schematic flowchart of another implementation of the graph data processing method according to the embodiment of the present application;

fig. 6 is a schematic flowchart of another implementation of the graph data processing method according to the embodiment of the present application;

fig. 7 is a schematic view of a virtual object recommendation interface obtained by using the graph data processing method according to the embodiment of the present application;

fig. 8 is a schematic flow chart illustrating an implementation process of the image data processing method applied to a game item recommendation scene according to the embodiment of the present application.

Detailed Description

In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the attached drawings, the described embodiments should not be considered as limiting the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order or importance, but rather "first \ second \ third" may, where permissible, be interchanged in a particular order or sequence so that embodiments of the present application described herein can be practiced in other than the order shown or described herein.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the application.

Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.

1) The following drawings: a data structure is composed of vertexes and edges, wherein one edge can only be connected with two vertexes;

2) The Graph Neural Network (GNN) is an algorithm which uses a Neural Network to learn Graph structure data, extracts and explores features and modes in the Graph structure data, and meets the requirements of Graph learning tasks such as clustering, classification, prediction, segmentation and generation;

3) The Multi-Layer Perceptron (MLP) is characterized in that one to a plurality of hidden layers (hidden layers) are introduced on the basis of a single-Layer neural network, and the hidden layers are positioned between an input Layer and an output Layer;

4) Graph Convolution Networks (GCNs), generalizing the Convolution operation from traditional data (images or meshes) to Graph data. The key point is to learn a function and generate a representation vector of the node by aggregating the characteristics of the function and the characteristics of the neighbors;

5) Graph Attention Network (GAT), a new neural Network architecture based on Graph structure data, utilizes a hidden self-Attention layer to solve the shortcomings of the previous methods based on Graph convolution or its approximation;

6) The high-frequency signal and the low-frequency signal, from the viewpoint of signal processing, the component with high frequency of the signal changing with time is called the high-frequency signal, the component with low frequency of the signal changing with time is called the low-frequency signal, the low-frequency signal is generally considered to reflect the global characteristics of the signal, the high-frequency signal reflects the local characteristics of the signal, and the whole signal can be more completely described by considering the two signals.

In order to better understand the graph data processing method provided in the embodiment of the present application, a graph data processing method in the related art is first described.

The mathematical representation of the graph data is as follows: graph data

Wherein

Indicating sectionSet of points, ∈ = { e = { _i }, | ε | = M denotes a set of edges, A ∈ R ^N×N ,A _ij E 0,1 represents the adjacency matrix,

representing a node degree matrix, and L = D-a representing a laplacian matrix of the graph structure, typically in a normalized form

Corresponding to the eigenvector Λ = { λ) of the laplace matrix with normalized frequency components of the resulting graph structure _i }∈[0,2],

X＝R ^N×d Initial representation of a representation node, Y _train Representing a known sample label.

The graph convolution network defines that each layer of convolution calculation is to replace the representation of the central node by the average representation of all the neighbor nodes of the central node. The graph attention network introduces an attention mechanism, and replaces direct average calculation in the graph convolution network with weighted summation of neighbor nodes by attention weights among nodes. GraphSAGE proposes a fixed number of samples of the neighbors of each node, thereby reducing the computational complexity of the model. However, these mainstream graph neural network models generally only retain the low-frequency components of the graph signals at each layer, and can be regarded as a low-pass filter, and the superposition of such low-pass filters may cause the representation vectors of each learned node to tend to be consistent, which seriously affects the results of downstream tasks, which is generally referred to as an over-smoothing problem. On the other hand, these low-pass filters are more suitable for homographs, i.e., neighboring nodes tend to have the same label, while for graph data in which neighboring nodes tend to have inconsistent labels, i.e., heterograph data, these methods perform less optimally.

Based on this, the embodiment of the present application provides a graph data processing method, which can provide high-frequency signal components for a graph neural network regarded as a low-pass filter, and can not introduce additional training parameters on the basis of an original model. For all graph neural networks, the graph data processing method provided by the embodiment of the application can be a universal plug-in, can be efficiently deployed in the existing network architecture, and can be applied to large-scale data scenes.

The following describes an exemplary application of the electronic device provided in the embodiments of the present application, and the electronic device provided in the embodiments of the present application may be implemented as various types of user terminals such as a notebook computer, a tablet computer, a desktop computer, a set-top box, a mobile device (e.g., a mobile phone, a portable music player, a personal digital assistant, a dedicated messaging device, and a portable game device), and may also be implemented as a server. In the following, an exemplary application will be explained when the device is implemented as a server.

Referring to fig. 1, fig. 1 is a schematic diagram of a network architecture of a data processing system 100 according to an embodiment of the present application, and as shown in fig. 1, the system includes: the terminal 200 is connected with the server 400 through the network 300, and the network 300 can be a wide area network or a local area network, or a combination of the two.

The graph data processing method provided by the embodiment of the application can be applied to various scenes, such as recommendation of purchased articles, video recommendation, music recommendation, item recommendation in games, matching among game players and the like. In the embodiment of the present application, a scenario in which an item is recommended to a player is described as an example.

The terminal 400 may have various applications installed therein, such as a game application, a shopping application, a video viewing application, and the like. The application may be an application that needs to be downloaded and installed, or may be an application (applet) that is to be used on demand, which is not limited in this embodiment of the application. In the embodiment of the present application, the game application may be any application capable of providing a virtual environment in which virtual objects substituted and operated by a user are active. For example, a Massively Multiplayer Online Role Playing (MMORP) Game, a Third-person Shooting Game (TPS), a First-person Shooting Game (FPS), and the like may be used. The server stores game match data of a plurality of players, such as virtual objects selected and used in the match process, selected props, match data in the game match process, match results and the like, the server 400 acquires prop information selected or used by the players from the database 500, establishes a graph structure based on the prop information selected or used by the players, determines the graph structure as graph structure data to be processed, and then acquires a trained graph neural network model, wherein a loss function of the trained graph neural network model in the training process at least comprises a first loss function component capable of enhancing high-frequency signals; and then, carrying out prediction processing on the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in the graph structure, wherein the representation vector can represent the global features and the local features of the nodes. After a game application program client is started and before game play is started, a player can select a virtual object and props in the game play process, after an operation instruction for selecting game props is triggered, a recommendation request is sent to a server, the server determines the recommendation request as a task processing request, then responds to the task processing request, corresponding processing is carried out on each node based on a representation vector of each node to obtain a processing result, in a prop recommendation scene, the processing result comprises first N prop information matched with the player, then the processing result (namely the first N prop information) is sent to a terminal 200, the terminal 200 displays the processing result in a display interface of the terminal 200, namely the first N prop information is displayed, and then the terminal 200 determines one or more props as props selected by the game play according to the fact that the terminal receives selection operation of one or more props, and starts the game play.

In some embodiments, the server 400 may be an independent physical server, may also be a server cluster or a distributed system formed by a plurality of physical servers, and may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web services, cloud communication, middleware services, domain name services, security services, CDNs, and big data and artificial intelligence platforms. The terminal 200 may be, but is not limited to, a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a vehicle-mounted smart terminal, and the like. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the embodiment of the present application is not limited.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a server 400 according to an embodiment of the present application, where the server 400 shown in fig. 2 includes: at least one processor 410, at least one network interface 420, a bus system 430, and a memory 440. The various components in server 400 are coupled together by a bus system 430. It is understood that the bus system 430 is used to enable connected communication between these components. The bus system 430 includes a power bus, a control bus, and a status signal bus in addition to the data bus. For clarity of illustration, however, the various buses are designated as bus system 430 in FIG. 2.

The Processor 410 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

The memory 440 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 440 optionally includes one or more storage devices physically located remote from processor 410.

Memory 440 includes volatile memory or nonvolatile memory, and can include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 440 described in embodiments herein is intended to comprise any suitable type of memory.

In some embodiments, memory 440 is capable of storing data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below.

An operating system 441 including system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and handling hardware-based tasks;

a network communication module 442 for communicating to other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), etc.;

in some embodiments, the apparatus provided by the embodiments of the present application may be implemented in software, and fig. 2 illustrates a model training apparatus 443 stored in the memory 440, which may be software in the form of programs and plug-ins, and includes the following software modules: the first obtaining module 4431, the first determining module 4432, the second determining module 4433, the third determining module 4434, the fourth determining module 4435 and the first training module 4436 are logical and thus may be arbitrarily combined or further divided according to the functions implemented. The functions of the respective modules will be explained below.

In other embodiments, the apparatus provided in the embodiments of the present Application may be implemented in hardware, and for example, the apparatus provided in the embodiments of the present Application may be a processor in the form of a hardware decoding processor, which is programmed to perform the model training method provided in the embodiments of the present Application, for example, the processor in the form of the hardware decoding processor may be one or more Application Specific Integrated Circuits (ASICs), DSPs, programmable Logic Devices (PLDs), complex Programmable Logic Devices (CPLDs), field Programmable Gate Arrays (FPGAs), or other electronic components.

The graph data processing method provided by the embodiment of the present application will be described with reference to exemplary applications and implementations of the server provided by the embodiment of the present application.

Before explaining the graph data processing method provided in the embodiment of the present application, artificial intelligence and several branches of artificial intelligence, and the branches related to the artificial intelligence-based molecular classification method provided in the embodiment of the present application will be explained first.

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the implementation method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject, and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence base technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like. The method provided by the embodiment of the application mainly relates to the research direction of machine learning.

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach to make computers have intelligence, and is applied in various fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning.

The embodiment of the present application provides a model training method, which is applied to an electronic device, where the electronic device may be a terminal or a server, and the description is given by taking the electronic device as the server in the embodiment of the present application. Referring to fig. 3, fig. 3 is a schematic flow chart of an implementation of the model training method provided in the embodiment of the present application, and each step of the model training method provided in the embodiment of the present application will be described below with reference to fig. 3.

And S101, acquiring a preset graph neural network model and training data.

Here, the training data includes a training graph structure and label information of each training node in the training graph structure. The label information of the training node may be class information of the training node, for example, when the training node is multimedia information, the label information of the training node may be a class of the multimedia information. If the multimedia information is music, the label information can be rock, light music, popular music, etc.; if the multimedia information is a video, the label information can be suspicion, speech, science and education and the like. When the training nodes are virtual objects, the label information of the training nodes can be warriors, combatants, shooters, assistance, jurisprudence, tanks, and the like.

In the embodiment of the present application, the training graph structure types of the graph neural network models corresponding to different task processing requests are different. For example, when the task processing request is a multimedia recommendation request, then the training graph structure is a user-multimedia data interaction graph; when the task processing request is a player matching request, the training graph structure is a player-player interaction graph; when the task processing request is a prop recommendation request, the training graph structure is a player-prop interaction graph.

The preset graph neural network model can be a graph convolution neural network model or a graph attention neural network model, and the model parameters are default parameter values.

And S102, determining the training expression vector of each training node by using a graph neural network model.

In the practical application process, when the step is realized, firstly, the attribute information of each training node is obtained, and the feature vector of each training node is determined based on the attribute information of each training node; and then, determining a second adjacent matrix of the training graph structure, and encoding the feature vector of each training node and the second adjacent matrix of the training graph structure by using a graph neural network model to obtain a training expression vector of each training node.

The attribute information of the training nodes at least comprises identification of the training nodes, and when the training nodes are users, the attribute information of the training nodes can also comprise information such as gender, age, region, user grade and the like. When the training node is a virtual article (e.g., a game item), the attribute information of the training node may further include an effect of a high virtual article, a player level at which the virtual article may be selected, an attack power of the virtual article, and the like; when the training nodes are multimedia data, attribute information of the training nodes includes the type, language and deduction information of the multimedia data; when the training node is an actual article, the attribute information of the training node may include information such as material and price of the article.

When determining that the feature vector of each training node is implemented based on the attribute information of each training node, feature extraction may be performed on the attribute information of each training node by using a feature extraction layer in a graph neural network model to obtain the feature vector of each training node.

And coding the characteristic vector of each training node and the second adjacency matrix by using a graph neural network model to obtain a training expression vector of each training node. When the method is realized, the characteristic vector of each training node and the second adjacency matrix are coded by using a coder of the graph neural network model, and the training expression vector of each training node is obtained.

Step S103, acquiring a second adjacent matrix of the training graph structure, and determining a complementary graph structure of the training graph structure and a first adjacent matrix of the complementary graph structure.

The supplementary graph structure of the training graph structure comprises all training nodes in the training graph structure, but connecting edges in the supplementary graph structure do not exist in the training graph structure. That is, the complementary graph structure and the training graph structure do not have the same connecting edges. Determining a complementary graph structure of the training graph structure, wherein when the complementary graph structure is implemented, a non-neighbor node set of each training node can be determined based on a second adjacency matrix of the training graph structure, then randomly selecting a preset number of target nodes from the non-neighbor node set of each training node, and establishing connecting edges of each training node and the corresponding target nodes, so as to form the complementary graph structure.

Since the graph structure is represented by two arrays, one being a one-dimensional array, for storing the vertex information in the graph structure, in the embodiment of the present application, the vertex information may be the node identifier of the training node. And the other is a two-dimensional array, namely an adjacent matrix, which is used for storing information of edges in the graph structure, for the undirected graph structure, if a connecting edge is arranged between an ith vertex and a jth vertex, elements in an ith row and a jth column and elements in an ith row and an ith column in the adjacent matrix are 1, and if the ith vertex and the jth vertex do not have a connecting edge, elements in the ith row and the jth column and elements in the jth row and the ith column in the adjacent matrix are 0. In the embodiment of the present application, since the training graph structure is known, the first adjacency matrix of the training graph structure can be directly obtained.

For the complementary graph structure, a set of non-neighboring nodes for each training node may be determined based on the second adjacency matrix of the training graph structure. And then, determining a target node corresponding to each training node from the non-neighbor node set of each training node, and determining a node having a connecting edge with each training node in the complementary graph structure at the moment, so that a second adjacency matrix of the complementary graph structure can be determined.

And determining a non-neighbor node set of each training node based on a second adjacent matrix of the training graph structure, wherein when the non-neighbor node set is realized, a node corresponding to 0 in each i row in the second adjacent matrix is the non-neighbor node set of the ith training node.

For example, there are ten training nodes, V1, V2, V3, V4, V5, V6, V7, V8, V9, V10 respectively, then the first behavior of the second adjacency matrix [1, 0,1,0] then the set of non-neighboring nodes for the first training node is { V4, V5, V7, V9, V10}. <xnotran> 3, , V5, V7, V9, [0,0,0,0,1,0,1,0,1,0]. </xnotran>

Step S104, determining a first loss function component based on the first adjacency matrix, the second adjacency matrix and the training expression vectors of the training nodes.

In an embodiment of the present application, the first loss function component includes a first regular subentry determined using the first adjacency matrix and the training representation vector of each training node, and a second regular subentry determined using the second adjacency matrix and the training representation vector of each training node. After the first regular subitem and the second regular subitem are determined, the first regular subitem and the second regular subitem are weighted and summed by using the weight coefficient (first superparameter) of the first regular subitem and the weight coefficient (second superparameter) of the second regular subitem to obtain a first loss function component.

The first loss function component comprises a first regular subentry determined by a first adjacency matrix of a complementary graph structure and a training representation vector of each training node, while a second regular subentry determined by a second adjacency matrix of an original training graph structure can shorten the distance between two training nodes with a connection relation in the training graph structure, and the graph signal passing through the graph neural network model can be smoother through the constraint of the second regular subentry, which is equivalent to the low-pass filtering processing of the graph signal in a frequency domain.

Step S105, determining a joint loss function based on a second loss function component preset by the graph neural network model and the first loss function component.

In the embodiment of the present application, the second loss function component preset by the graph neural network model may be a cross entropy function, or may be another type of loss function, for example, an L1 loss function, a KL distance loss function, or the like. In the embodiment of the present application, a preset second loss function component is taken as an example to illustrate, and since the convergence rate of the cross entropy loss function is fast, the update rate of the model parameter can be increased, and the training duration of the pre-training model can be reduced.

When the step is realized, a preset regular term coefficient can be obtained, the regular term coefficient is a weight coefficient of the first loss function component, and then the joint loss function is determined based on the following formula (1-2)

Wherein, the first and the second end of the pipe are connected with each other,

a second loss function component preset for the graph neural network model, gamma being a regularization term coefficient,

is the first loss function component.

And step S106, training the graph neural network model by using the joint loss function, the training expression vectors of the training nodes and the training labels of the training nodes to obtain the trained graph neural network model.

Through the above steps S101 to S106, in the training process of the neural network model, it is not necessary to add a high pass filter to each layer as in the related art, nor to make the trained neural network model have the capability of expressing high frequency signals through the calculation of the high order polynomial of the laplace matrix, but instead, on the basis of not changing the neural network model, the loss function used for training the neural network model is improved, and a first loss function component is added on the basis of a preset second loss function component, and the first loss function component includes a first regularization sub-item determined by a first adjacent matrix of a complementary graph structure corresponding to the training graph structure and a second regularization sub-item determined by a second adjacent matrix corresponding to the training graph structure.

In some embodiments, the "determining the complementary graph structure of the training graph structure" in the above step S103 may be implemented by:

and step S1031, determining a non-neighbor node set of each training node in the training graph structure based on the second adjacency matrix of the training graph structure.

The second adjacency matrix of the training graph structure can represent the connection relation between the training nodes, and if a connection edge exists between the ith training node and the jth training node, the element of the ith row and the jth column of the second adjacency matrix and the element of the jth row and the ith column are 1. That is, the ith row of the second adjacency matrix of the training graph structure represents the connection relationship between the ith training node and each training node, and when the non-neighbor node set of the ith training node in the training graph structure is determined, the node corresponding to the value 0 in the ith row of the second adjacency matrix is placed in the non-neighbor node set.

Step S1032, a preset number of target non-neighbor nodes are determined from the non-neighbor node set of each training node.

When the step is implemented, a preset number of target non-neighbor nodes can be randomly selected from the non-neighbor node set of each training node. The preset number N is determined based on the number of elements in the non-neighbor node set of each training node, and when the method is implemented, the preset number N may be a random integer value smaller than or equal to the minimum value of each element number, and the preset number N may be obtained by subtracting a preset integer value from the minimum value of each element number, for example, the first preset integer value may be 1, and the preset number N is a value obtained by-1 from the minimum value; the preset number N may be the minimum value of the number of each element.

Assuming that the preset number N is 2, this step is to randomly select two target non-neighbor nodes from the non-neighbor node set of each training node.

And step S1033, connecting each training node with the corresponding target non-neighbor node to form a complementary graph structure of the training graph structure.

Through the steps S1031 to S1033, a non-neighbor node set of each training node is determined, then a preset number of target non-neighbor nodes are randomly determined from the non-neighbor node set of each training node, and each training node is connected with the corresponding non-neighbor node, so that a complementary graph structure of the training graph structure is formed, and a data basis is provided for subsequently determining the first loss function component capable of enhancing the high-frequency signal.

In some embodiments, the first loss function component includes a first regular sub-term and a second regular sub-term, and correspondingly, the step S104 "determining the first loss function component based on the first adjacency matrix, the second adjacency matrix and the training representation vectors of the training nodes" can be implemented by:

step S1041, acquiring a first hyper-parameter and a second hyper-parameter.

Here, the first superparameter is a weight coefficient of the first canonical subentry, and the second superparameter is a weight coefficient of the second canonical subentry. The first hyper-parameter and the second hyper-parameter may be understood as model parameters that need to be adjusted when performing model training.

Step S1042, a product of a first adjacency coefficient corresponding to the ith row and the jth column of the first adjacency matrix, a training vector of the ith training node, and a transposed product of the training vector of the jth training node is summed to obtain a first regular subentry capable of representing a high-pass filter.

Wherein i =1,2, \8230, M, j =1,2, \8230, and M, M is the total number of nodes in the training graph structure.

When implemented, can be implemented by

A first regularization sub-term capable of representing a high pass filter is determined, wherein,

a first adjacency coefficient, h, corresponding to the ith row and jth column of the first adjacency matrix _i Is the training vector for the ith training node,

is the transpose of the training vector for the jth training node.

Step S1043, summing products of a second adjacency coefficient corresponding to the ith row and the jth column of the second adjacency matrix, the training vector of the ith training node, and the transpose of the training vector of the jth training node, to obtain a second regular subentry capable of representing a low-pass filter.

In implementation, can be realized by

A second regularization sub-term is derived that can represent a low pass filter, where A _ij A second adjacency coefficient, h, corresponding to the ith row and the jth column of the second adjacency matrix _i Is the training vector for the ith training node,

is the transpose of the training vector for the jth training node.

Step S1044 is to perform weighted summation on the first regular sub-term and the second regular sub-term by using the first hyper-parameter and the second hyper-parameter, so as to obtain a first loss function component.

In an implementation, the first loss function component may be determined by equation (1-1):

wherein alpha is a first hyperparameter and beta is a second hyperparameter.

Through the steps S1041 to S1044, the first regular subentry determined by the first adjacency matrix and the training representation vector of each training node may be utilized, the second regular subentry determined by the second adjacency matrix and the training representation vector of each training node may be utilized, and then the first loss function component may be determined based on the first regular subentry and the second regular subentry. Furthermore, the low-frequency signals reflect the global characteristics of the training nodes, and the high-frequency signals reflect the local characteristics of the training nodes, so that the difference between the predicted node characteristics of the graph structure can be ensured through the trained graph neural network model, the problem of over-smoothness caused by the fact that node vectors of all nodes tend to be consistent is avoided, and the accuracy of downstream task processing is improved.

In some embodiments, the step S106 "training the graph neural network model by using the joint loss function, the training expression vectors of the training nodes, and the training labels of the training nodes to obtain the trained graph neural network model" may be implemented by:

step S1061, performing prediction processing on the training expression vectors of the training nodes by using the graph neural network model to obtain prediction information of the training nodes.

In the embodiment of the application, after the training expression vector of each training node is obtained, the training expression vector of each training node is subjected to prediction processing by using a classification module of a graph neural network model, the classification module may be a full connection layer or any other suitable classifier (e.g., a decision tree, a bayesian classifier, a support vector machine, etc.), and prediction information of each training node can be obtained by using the classification module.

Step S1062, determining a loss value based on the prediction information of each training node, the label information of each training node, and the joint loss function.

In the embodiment of the application, the loss value can represent the difference degree between the prediction information of each training node and the corresponding label information, and when the step is implemented, the prediction information of each training node and the label information of each training node are substituted into the joint loss function, that is, the loss value is determined.

And S1063, when the joint loss function is determined to be not converged according to the loss value, adjusting parameters of the graph neural network model based on the joint loss function until the trained graph neural network model is obtained.

When the loss value continuously tends to a constant, determining the convergence of the loss function, namely when the difference value between the loss value and the constant is smaller than a preset difference threshold value, determining the convergence of the combined loss function; and when the difference value of the loss value and the constant is larger than or equal to the difference threshold value, determining that the joint loss function is not converged, and continuing to train the graph neural network model. When determining that the joint loss function is not converged, adjusting parameters of the graph neural network model by adopting an Adam optimization algorithm or an exponential descent random gradient descent algorithm based on the joint loss function, then calculating a loss value again, and obtaining the trained graph neural network model when determining that the joint loss function is converged based on the loss value.

In some embodiments, after the trained neural network model is obtained through the above steps S101 to S106, it may be further determined whether the trained neural network model reaches the evaluation index through the following steps, and the training of the neural network model is continued when the evaluation index is not reached:

step S107, test data is acquired.

The test data comprises a test pattern structure and label information of each test node in the test pattern structure. The node types of the test graph structure and the training graph structure are the same.

And S108, predicting the test chart structure by using the trained chart neural network model to obtain the prediction information of each test node in the test chart structure.

And step S110, determining an evaluation index value of the trained graph neural network model based on the prediction information of each test node and the corresponding label information.

In the embodiment of the present application, the evaluation index may be an accuracy rate, a recall rate, or the like.

And step S111, determining that the evaluation index value is smaller than a preset index threshold value, and acquiring the training data again.

The training data comprises a training graph structure and label information of each training node in the training graph structure.

In some embodiments, when the evaluation index value is greater than or equal to the index threshold value, the trained graph neural network model is determined to reach the evaluation passing standard, and the training of the trained graph neural network model does not need to be continued.

And step S112, continuously training the trained graph neural network model by using the training data until the evaluation index value of the retrained graph neural network model reaches the index threshold value.

When the method is realized, a trained graph neural network model is used for conducting prediction processing on a training graph structure to obtain prediction information of each training node, then loss values are determined based on the prediction information of each training node, label information of each training node and a joint loss function, and when the joint loss function is determined to be not converged based on the loss values, parameters of the graph neural network model are adjusted by adopting an Adam optimization algorithm or a gradient descent method until evaluation index values of the retrained graph neural network model reach an index threshold value.

Through the steps S107 to S112, after the trained graph neural network model is preliminarily obtained, the performance index of the preliminarily trained graph neural network model is evaluated to obtain an evaluation index value, and when it is determined that the evaluation index value does not reach the preset index threshold value, it indicates that the performance index of the preliminarily trained graph neural network model does not reach the standard, at this time, new training data needs to be obtained again, and the graph neural network model continues to be trained until the evaluation index of the trained graph neural network model reaches the index threshold value, so that the accuracy of graph data processing by using the trained graph neural network model can be ensured.

Based on the foregoing embodiments, an embodiment of the present application provides a graph data processing method, which is applied to an electronic device, where the electronic device may be a server, fig. 4 is a schematic diagram of an implementation flow of the graph data processing method provided in the embodiment of the present application, and each step of the graph data processing method provided in the embodiment of the present application is described below with reference to fig. 4.

Step S201, in response to the received task processing request, obtaining the graph structure data to be processed and the trained graph neural network model corresponding to the task processing request.

In the embodiment of the present application, before executing the graph data processing method provided in the embodiment of the present application, the server may construct a graph structure according to the historical data stored in the database, and further acquire graph structure data. The graph structure data may include information of each node in the graph structure and information of connection edges between nodes, where the connection edge information may include vertex node identifiers of connection edges and may further include weights of the connection edges. Taking the construction of the graph structure between the player and the prop as an example, in the implementation, the record of the player purchasing the prop can be inquired and obtained from the purchasing database, and the record of the player using the prop can be inquired and obtained from the game data. The player and the prop are regarded as nodes, and the purchasing/using relation is used as an edge between the nodes to construct a player-prop interaction graph. And then taking the construction of a graph structure between players as an example, the player information in each game can be inquired and obtained from the game data in the realization, the players are taken as nodes, and the game relationship is taken as an edge between the nodes to construct a player-player interaction graph. When constructing the graph structure between the user and the multimedia data, the watching or listening record, the downloading record, the like, of the user may be obtained first, the user and the multimedia data are regarded as nodes, the watching/downloading/favoring/collecting relationship is regarded as an edge between the nodes, and a user-multimedia data interaction graph is constructed, different connecting edges may have different weights based on different connecting relationships, for example, the connecting relationship between the user and the multimedia data is watching, downloading or collecting, the weight may be 2, and if the connecting relationship between the user and the multimedia data is favoring, the weight may be 1.

For example, when the mission processing request is a player matching request, the graph structure corresponding to the mission processing request is a player-player graph structure.

In the embodiment of the present application, the graph neural network model may be a graph convolution neural network model or a graph attention neural network model, the trained graph neural network model is obtained by training using the model training method, and the trained graph neural network model not only has the capability of representing low-frequency signals but also has the capability of representing high-frequency signals by learning the high-frequency signal representation through the first loss function component included in the joint loss function used in the training process.

And S202, performing prediction processing on the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in the graph structure.

In the embodiment of the application, the trained graph neural network model includes a first loss function component capable of enhancing a high-frequency signal in a joint loss function in a training process, a distance between non-existent edges on graph data can be shortened through the first loss function component, the first loss function component implies the representation of a high-pass filter, a constraint for expressing the high-frequency signal can be applied to the representation learned by the graph neural network model, the high-frequency signal can reflect local characteristics of the training nodes, and the low-frequency signal can reflect global characteristics of the training nodes, so that after the graph structure data is subjected to prediction processing through the trained graph neural network model, an obtained expression vector of the node can represent the global characteristics and the local characteristics of the node, and the data comprehensiveness of the expression vector of the node is improved.

Step S203, processing the task based on the expression vector of each node to obtain a processing result.

The task processing request may be a classification request or a recommendation request. In the actual application process, the step can be implemented in different ways based on different types of task processing requests. When the task processing request is a recommendation request, firstly determining a target node corresponding to the task processing request, then determining a matching degree score between the target node and each other node except the target node in the graph structure based on the representation vector of each node, and determining the top N nodes with the highest matching degree as processing results. When the task processing request is a classification request, the classifier is required to perform prediction processing on the expression vector of each node to obtain a class probability vector of each node, and the class information of each node, that is, the processing result, is determined based on the class probability vector of each node.

And step S204, outputting the processing result.

In this embodiment of the application, the server outputs the processing result, and may send the processing result to a terminal corresponding to the task processing request. And after receiving the processing result, the terminal displays the processing result on a display device of the terminal and carries out subsequent operation based on the processing.

In the graph data processing method provided by the embodiment of the application, after graph structure data to be processed are obtained, a trained graph neural network model is obtained first, a loss function of the trained graph neural network model in a training process at least comprises a first loss function component capable of enhancing a high-frequency signal, then the trained graph neural network model is used for carrying out prediction processing on the graph structure data to obtain a representation vector of each node in a graph structure, and the representation vector can represent the global characteristic and the local characteristic of the node; in response to the received task processing request, performing task processing based on the expression vectors of the nodes to obtain and output a processing result, wherein the loss function of the trained neural network model in the training process comprises a first loss function component capable of enhancing a high-frequency signal, so that the trained neural network model not only can represent the original low-frequency signal but also has the capability of representing the high-frequency signal, that is, the expression vectors of the nodes obtained by using the trained neural network model can represent the global characteristics of the nodes and can also ensure the local characteristics, thereby improving the data richness and the comprehensiveness of the node expression vectors, and further improving the accuracy of the processing result when the expression vectors of the nodes are used for performing task processing.

Based on the foregoing embodiments, an embodiment of the present application provides a graph data processing method, which is applied to the network structure shown in fig. 1, fig. 5 is a schematic diagram of another implementation flow of the graph data processing method provided in the embodiment of the present application, and each step of the graph data processing method provided in the embodiment of the present application is described below with reference to fig. 5.

Step S301, the terminal responds to the operation instruction for starting the game client, starts the game client and presents a game interface.

In this embodiment, the game client may be a game application client, and the operation instruction for opening the client may be an instruction generated based on a user clicking or touching a game application icon in a display screen of the terminal.

Step S302, the terminal receives an operation instruction for selecting the prop triggered through the game interface.

Before game play, a player can select one or more virtual items according to the game level of the player or the skill which is desired to be released in the game, or can select the virtual items according to the virtual objects selected by the player.

In step S303, the terminal sends a task processing request to the server in response to the operation instruction.

The task processing request is used for requesting the server to provide selectable props for the terminal, and therefore in the embodiment of the application, the task processing request is a recommendation request. The server may be a server corresponding to the game client.

And step S304, the server responds to the task processing request and acquires the graph structure data to be processed and the trained graph neural network model.

Since in a game scenario, a player may have various requirements, for example, matching enemy players belonging to different camps, or matching friend players of the same camps, and then matching suitable items, or requesting a recommended virtual object, etc., different graph structures are constructed in the server according to different requirements, for example, a player-item graph structure, a player-player graph structure, a player-virtual object graph structure, etc. may be constructed. In the embodiment of the application, after the server receives the task processing request, the graph structure corresponding to the type of the task processing request is determined as the graph structure to be processed based on the type of the task processing request.

The loss function of the trained graph neural network model in the training process at least comprises a first loss function component capable of enhancing high-frequency signals.

And S305, the server performs prediction processing on the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in the graph structure.

Because the loss function used in the training process of the graph neural network model comprises the first loss function component capable of enhancing the high-frequency signal, the trained graph neural network model can represent the high-frequency signal, so that the node representation vector obtained by the trained graph neural network model can represent the global feature and the local feature of the node, and the data comprehensiveness of the node representation vector is improved.

In step S306, the server determines a target node based on the task processing request.

In the embodiment of the application, the task processing request carries identification information of a player corresponding to the terminal, and the target node in the graph structure can be determined by using the identification information.

Step S307, the server determines matching degree scores between the target node and other nodes in the graph structure data based on the expression vectors of the nodes.

During implementation, the inner product of the expression vector of the target node and the expression vectors of the other nodes in the graph structure data may be calculated, and then normalization processing is performed to obtain a matching degree score between the target node and the other nodes.

In some embodiments, respective distances between the representation vector of the target node and the representation vectors of the respective other nodes may also be determined, and then the respective distances may be used to determine a matching degree score between the target node and the other nodes.

In step S308, the server determines other nodes whose matching degree score with the target node is greater than a preset matching threshold as matching objects of the target node.

And step S309, the server sorts the matching objects according to the matching degree scores to obtain a sorting result, and determines the sorting result as a processing result.

In the embodiment of the application, the matching objects are sorted from high to low according to the matching degree score to obtain a sorting result, and the sorting result is determined as a final processing result.

In step S310, the server sends the processing result to the terminal.

In step S311, the terminal displays the processing result on its own display interface.

Because the processing results are sorted from high to low according to the matching degree scores, when the terminal displays the processing results, the props corresponding to the matching objects are also displayed according to the sequence from high to low of the matching degree scores, and therefore the player can preferentially view the props with high matching degree.

In step S312, the terminal receives the selection operation for the processing result, and determines the target prop to be used.

In step S313, the terminal receives an operation command for starting game play and starts game play.

In step S314, the terminal acquires the game play data from the server, and transmits the game play data generated based on the player operation to the server.

After the terminal acquires game match data from the server, a virtual scene comprising a virtual object and a graphic control for displaying a graphic visual area of the virtual object in the virtual scene are loaded and displayed based on the match data. Here, the virtual scene may refer to an image frame including a game scene, and the virtual object may include an object controlled by a user and may also include an object controlled by a machine. Then, the player controls the movement of the virtual object corresponding to the player or releases the skill and the like through the terminal, and sends the corresponding game match data to the server.

In the graph data processing method provided by the embodiment of the application, when a game client is started by a terminal and a prop selection request is triggered, a task processing request is sent to a server, the server responds to the task processing request, firstly, a graph structure to be processed and a trained graph neural network model are obtained, then, the trained graph neural network model is used for carrying out prediction processing on graph structure data to obtain a representation vector of each node in the graph structure, a target node corresponding to the task processing request is determined, then, matching degree scores of the target node and other nodes in the graph structure are determined, finally, a processing result is obtained based on the matching degree scores, and the processing result is sent to the terminal, so that the terminal selects a target prop based on the processing result and starts a game-to-game process.

In the above embodiments, the description is given by taking a task processing request as a recommendation request as an example. In some embodiments, when the task processing request is a classification request, the step S103 may be implemented by:

and step S1031, performing prediction processing on the expression vector of each node by using the trained classifier, and obtaining a category probability vector of each node.

The trained classifier can be a full connection layer in a graph neural network structure, or can be an independent classifier, and the class probability vector can represent the probability or confidence that the node possibly belongs to each class.

Step S1032, determining a classification result of each node based on the class probability vector of each node.

In implementation, the class corresponding to the highest probability in the class probability vector of the node may be determined as the classification result of the node.

Step S1033, determining the classification result of each node as the processing result.

In the embodiment of steps S1031 to S1033, the trained classifier is used to perform classification processing based on the expression vectors of the nodes to obtain the classification result of each node, and the expression vectors of the nodes can not only represent all the features but also represent local features, so that the accuracy of the obtained classification result can be ensured.

In the following, an exemplary application of the embodiments of the present application in a practical application scenario will be described.

The embodiment of the application provides a graph data processing method capable of efficiently providing high-frequency signal components. In the implementation, firstly, a hidden layer representation vector of each node is obtained through an original graph neural network, then, the structure of the original graph is subjected to high random sampling, and then, laplace regularization (namely, the representation distance between adjacent nodes is reduced) is adopted on the graph structure obtained through sampling, so that the constraint of expressing high-frequency signals is applied to the learned representation vector. And finally, calculating a cross entropy loss function of the training set by using the obtained expression vector through a node classification task to update parameters of the original graph neural network so as to train the graph neural network, so that the graph neural network obtained by training has the capacity of expressing high-frequency signals besides the original low-frequency signals.

The graph data processing method provided by the embodiment of the application can be widely applied to all graph neural networks, and high-frequency components of graph signals can be efficiently modeled. Compared with the related art, the graph data processing method provided by the embodiment of the application does not increase extra computational complexity and is a flexible plug-in.

The following describes the graph data processing method provided in the embodiments of the present application and the implementation of the graph data processing method as a plug-in to an existing graph neural network. Fig. 6 is a schematic diagram of a further implementation flow of the graph data processing method according to the embodiment of the present application, and as shown in fig. 6, the flow includes:

step S601, original map data is acquired.

Step S602, an adjacency matrix and a feature vector of each node are determined based on the original graph data.

Step S603, the feature vectors of the adjacent matrix and each node are encoded by using the neural network encoder to obtain a representation vector of each node.

In step S604, the patch data is determined based on the original map data.

Wherein, the complement data includes edges that do not exist in the original graph data.

In step S605, the complement data is sampled.

And step S606, carrying out supervised and unsupervised joint training on the graph neural network model.

The method comprises the steps of updating parameters of a main graph neural network in a supervised (unsupervised constraint) mode through loss functions such as cross entropy and the like based on expression vectors of all nodes, so that the graph neural network is trained, and enhancing high-frequency signals of the existing graph neural network by constructing a regular unsupervised constraint (unsupervised CLAR).

Inputting: graph data

Wherein

Represents a set of nodes, ∈ = { e = { _i Is |, | ε | = M represents a set of edges, A ∈ R ^N×N ,A _ij E 0,1 represents an adjacency matrix,

a node degree matrix is represented, L = D-a represents a laplacian matrix of a graph structure, a graph neural network is used as a main skeleton of the graph data processing method provided by the embodiment of the application, and it is assumed that a graph convolution network f is selected _GCN (X, A); negative sampled side multiple s =2; coefficient γ of the regularization term.

Obtaining a characterization vector H = f of a node through a graph neural network _GCN (X, a) X = RN × dytranin, and then random sampling is performed on the graph structure ∈ in order to guarantee the graphStructural balance, sampling s non-existent edges for every existent edge to obtain graph structure epsilon _s ：

Then, the sampled graph structure epsilon is obtained according to the formula (1-1) _s And carrying out regularization constraint:

wherein alpha and beta are hyper-parameters. A. The ^s Representing the sampled graph structure epsilon _s The corresponding adjacency matrix means to draw the distance between the node representations of the edges which do not exist on the original image, the regular term (corresponding to the first loss function component in other embodiments) implies the representation of the high-pass filter, since the regular term on the original image represents a low-pass filter, the edges which do not exist are the complement of the original image, and the regular term on the complement image represents a high-pass filter.

Thus, the training target of the whole task is converted into the classification task of the original graph neural network plus a regular term as shown in formula (1-2):

the training target of the original graph neural network model can be expressed by the formula (1-3):

g (-) is generally cross entropy.

The training process may be trained using an Adam optimizer to update f _GCN The main network can adopt a two-layer network, the hidden characteristic dimension is uniformly set to be 32, and gamma adopts a parameter from [0,1,2 ]]The best result searched out.

The graph data processing method provided by the embodiment of the application can be applied to a general model architecture and is not limited to a specific application scene. In any graph structure data, the representation vector learned by the embodiment of the application has richer representation capability, and can provide a better user representation vector for downstream tasks. As long as the data of the graph structure can be constructed, the graph neural network can help to obtain valuable node vector representations to help the completion of downstream tasks. For example, in a game player recommendation scene, knowing game-play history information between players, a user-user game-play map can be constructed; or under the game item recommending field, the purchase or use record between the known player and the item forms a player-item graph; or in a virtual object recommendation scene, a user-virtual object graph is constructed through game history information of a player, then a player/article/virtual object representation vector obtained through the graph data method provided by the embodiment of the application can obtain a score value matched between a user and the user, or between the user and an article, or between the user and a virtual object through a scoring function, and recommendation results of friends/props/virtual objects can be provided for the user through sequencing the score values.

Fig. 7 is a schematic view of a virtual object recommendation interface obtained by using the graph data processing method according to the embodiment of the present application, where on the interface shown in fig. 7, in order of high and low matching degree with a player, each virtual object (hero) suitable for the player to use is shown, so that the player can quickly select the virtual object used in the game play of this time.

The application process of the graph data processing method provided by the embodiment of the present application is described below by taking a game player item recommendation scenario as an example. Under the condition that the item information purchased by the player and the item information used historically are known, the representation vectors of the player and the user can be obtained through the graph data processing method provided by the embodiment of the application, so that the matching degree score between the player and the user is determined, and the item is recommended to the game player.

Fig. 8 is a schematic view of an implementation flow of the graph data processing method applied to a game item recommendation scene, and the implementation flow is described below with reference to fig. 8.

Step S801, inquiring and acquiring records of items purchased by the player from the purchase database, and inquiring and acquiring records of items used by the player from the game-play data.

Step S802, the player and the prop are taken as nodes, the purchasing/using relation is taken as the edge between the nodes, and a player-prop interaction graph is constructed.

The edges in the graph may be weighted by other information such as the number of uses or time of purchase, defaulting to 1. The structure of the figure is as follows

Wherein all nodes include user nodes and item nodes:

step S803, the obtained graph data is input to a graph neural network, and a representation vector H of each node is obtained.

The graph data processing method provided by the embodiment of the application mainly enhances the structure of the graph neural network, and improves the representation capability of the graph neural network by improving the expression capability of high-frequency information.

And step S804, calculating the matching degree between the player and all global articles based on the obtained expression vector of each node.

This step is accomplished, when implemented, by a scoring function, most directly by vector inner product and Softmax normalization as shown in equations (1-4):

step S805, based on the matching scores between the props and the players, all the props are sorted, all the props owned by the players are removed, and the top K props with the largest number are selected for recommendation.

In practical application, the top K props with matching scores can be arranged and displayed from high to low according to the matching scores.

In the present embodiment, the GCN described above is compared with the graph convolution network, GAT (graph attention network), and SAGE (GraphSAGE). + represents that the network is used as a backbone network in the graph data method provided by the embodiment of the application. In addition, two-layer polynomial frequency domain diagram neural methods ChebNet and GPRGNN under the unified computational complexity are compared. The data set selects citation network (concordant data set) Cora, citeSeer and PubMed; wikipedia webpage hyperlink (heterotranscription dataset) Squirrel, chameleon and Actor. For all data sets, the data sets were randomly divided into 0.6,0.2,0.6 training, validation, test scale, and the average of 50 random divisions was published. And setting the maximum training iteration number to be 1000, and stopping training when the loss function values of the 50 verification sets are not reduced any more. For all experimental results, accuracy is selected as an index, and can be determined by using the formula (1-5):

TABLE 1 comparison of node Classification task Effect

As can be seen from table 1, in a node classification task, the graph data processing method provided in the embodiment of the present application has a very significant improvement on the same-configuration data set, particularly, the different-configuration data set, which indicates that a high-pass filter provided by the graph data processing method provided in the embodiment of the present application can help to obtain a better node representation.

In addition, table 2 also provides the effect of the graph data processing method provided in the embodiment of the present application on the over-smoothing problem. On a Cora data set, a graph convolution neural network is used as a main model, a more difficult data partitioning mode that 20 training samples are selected from each category is adopted, and classification results of 2,4,6 and 8 layers are compared. The result shows that when the neural network of the main graph (taking the convolutional neural network as an example) is stacked in multiple layers, the performance degradation of the model due to the over-smoothing problem can be avoided to a great extent by using the graph data processing method provided by the embodiment of the application.

TABLE 2 comparison of layer number for neural network

	2	4	6	8
					GCN	82.25	77.14	58.10	37.26
GCN+	82.45	77.23	63.57	50.75

It is understood that, in the embodiments of the present application, the content related to the user information, for example, the data related to the attribute information of the node, etc., when the embodiments of the present application are applied to a specific product or technology, user permission or consent needs to be obtained, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related country and region.

Continuing with the exemplary structure of the model training device 443 provided by the embodiments of the present application as a software module, in some embodiments, as shown in fig. 2, the software module stored in the model training device 443 of the memory 440 may include:

a first obtaining module 4431, configured to obtain a preset graph neural network model and training data, where the training data includes a training graph structure and label information of each training node in the training graph structure;

a first determining module 4432, configured to determine a training representation vector of each of the training nodes using the graph neural network model;

a second determining module 4433, configured to determine a complementary graph structure of the training graph structure and a first adjacency matrix of the complementary graph structure, and obtain a second adjacency matrix of the training graph structure;

a third determination module 4434 for determining a first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors for the respective training nodes;

a fourth determining module 4435, configured to determine a joint loss function based on a second loss function component preset by the neural network model and the first loss function component;

the first training module 4436 is configured to train the graph neural network model by using the joint loss function, the training expression vectors of the training nodes, and the training labels of the training nodes, so as to obtain a trained graph neural network model.

In some embodiments, the first determining module 4432 is further configured to:

acquiring attribute information of each training node, and determining a feature vector of each training node based on the attribute information of each training node;

and coding the characteristic vector of each training node and the second adjacency matrix by using the graph neural network model to obtain a training expression vector of each training node.

In some embodiments, the second determining module 4433 is further configured to:

determining a non-neighbor node set of each training node in the training graph structure based on a second adjacency matrix of the training graph structure;

determining a preset number of target non-neighbor nodes from the non-neighbor node set of each training node;

and connecting each training node with the corresponding target non-neighbor node to form a complementary graph structure of the training graph structure.

In some embodiments, the first loss function component includes a first regular sub-term and a second regular sub-term, and the third determining module 4434 is further configured to:

acquiring a first hyper-parameter and a second hyper-parameter;

summing products of a first adjacent coefficient corresponding to the ith row and the jth column of the first adjacent matrix, a training vector of an ith training node and a transpose of the training vector of the jth training node to obtain a first regular subentry capable of representing a high-pass filter;

summing products of a second adjacent coefficient corresponding to the ith row and the jth column of the second adjacent matrix, a training vector of an ith training node and a transpose of the training vector of the jth training node to obtain a second regular subentry capable of representing a low-pass filter;

and performing weighted summation on the first regular sub-term and the second regular sub-term by using the first hyper-parameter and the second hyper-parameter to obtain the first loss function component.

In some embodiments, the first training module 4436 is further configured to:

predicting the training expression vectors of the training nodes by using the graph neural network model to obtain prediction information of the training nodes;

determining a loss value based on the prediction information of each training node, the training labels of each training node, and the joint loss function;

and when the joint loss function is determined to be not converged according to the loss value, adjusting parameters of the graph neural network model based on the joint loss function until the trained graph neural network model is obtained.

In some embodiments, the apparatus further comprises:

the third acquisition module is used for acquiring test data, wherein the test data comprises a test chart structure and label information of each test node in the test chart structure;

the second prediction module is used for performing prediction processing on the test chart structure by using the trained chart neural network model to obtain prediction information of each test node in the test chart structure;

a fifth determining module, configured to determine an evaluation index value of the trained neural network model based on the prediction information of each test node and the corresponding label information;

the fourth acquisition module is used for determining that the evaluation index value is smaller than a preset index threshold value and acquiring the training data again;

and the second training module is used for continuously training the trained graph neural network model by using the training data until the evaluation index value of the retrained graph neural network model reaches the index threshold value.

It should be noted that, the embodiments of the present application are described with respect to the model training apparatus, and similar to the description of the method embodiments above, and have similar beneficial effects to the method embodiments. For technical details not disclosed in the embodiments of the apparatus, reference is made to the description of the embodiments of the method of the present application for understanding.

In some embodiments, the graph data processing apparatus provided in the embodiments of the present application may be implemented in a software manner, and may be software in the form of programs, plug-ins, and the like, and include the following software modules: the second acquisition module, the first prediction module, the first processing module and the result output module are logical, so that the modules can be arbitrarily combined or further split according to the realized functions. The functions of the respective modules are explained below.

the first prediction module is used for performing prediction processing on the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in a graph structure, and the representation vector can represent global features and local features of the node;

and the result output module is used for outputting the processing result.

In some embodiments, when the task processing request is a recommendation request, the first processing module is further configured to:

determining a target node based on the task processing request;

determining matching degree scores between a target node and other nodes in the graph structure data based on the representation vectors of the nodes;

determining other nodes with the matching degree score larger than a preset matching threshold value with the target node as matching objects of the target node;

and determining the matching object as a processing result.

In some embodiments, when the task processing request is a classification request, the first processing module is further configured to:

predicting the expression vector of each node by using a trained classifier to obtain a class probability vector of each node;

determining a classification result of each node based on the class probability vector of each node;

and determining the classification result of each node as the processing result.

It should be noted that, the embodiments of the present application are described with respect to the data processing apparatus, and similar to the description of the method embodiments described above, and have similar beneficial effects to the method embodiments. For technical details not disclosed in the embodiments of the apparatus, reference is made to the description of the embodiments of the method of the present application for understanding.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, so that the computer device executes the model training method or the graph data processing method described in the embodiment of the present application.

Embodiments of the present application provide a computer-readable storage medium storing executable instructions, which when executed by a processor, will cause the processor to perform a model training method or a graph data processing method provided by embodiments of the present application, for example, a model training method as shown in fig. 3, and a graph data processing method as shown in fig. 4 and 5.

In some embodiments, the computer-readable storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

By way of example, executable instructions may correspond, but do not necessarily have to correspond, to files in a file system, and may be stored in a portion of a file that holds other programs or data, such as in one or more scripts in a hypertext Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

By way of example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or distributed across multiple sites and interconnected by a communication network.

The above description is only an example of the present application, and is not intended to limit the scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method of model training, the method further comprising:

determining training representation vectors of the training nodes by using the graph neural network model;

determining a first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors of the respective training nodes;

2. The method of claim 1, wherein determining the training representation vector for each of the training nodes using the graph neural network model comprises:

and coding the feature vector of each training node and the second adjacency matrix by using the graph neural network model to obtain a training expression vector of each training node.

3. The method according to claim 1, wherein the determining a complementary graph structure of the training graph structure comprises:

4. The method of claim 1, wherein the first loss function component comprises a first regular subentry and a second regular subentry, and wherein determining the first loss function component based on the first adjacency matrix, the second adjacency matrix, and the training representation vectors for the respective training nodes comprises:

acquiring a first hyper-parameter and a second hyper-parameter;

summing products of a first adjacency coefficient corresponding to the ith row and the jth column of the first adjacency matrix, a training vector of an ith training node and a transpose of the training vector of the jth training node to obtain a first regular subentry capable of representing a high-pass filter;

summing products of a second adjacent coefficient corresponding to the ith row and the jth column of the second adjacent matrix, a training vector of the ith training node and a transpose of the training vector of the jth training node to obtain a second regular subentry capable of representing a low-pass filter;

5. The method of claim 1, wherein the training the graph neural network model using the joint loss function, the training representation vector of each training node, and the training label of each training node to obtain the trained graph neural network model comprises:

6. The method according to any one of claims 1 to 5, further comprising:

acquiring test data, wherein the test data comprises a test chart structure and label information of each test node in the test chart structure;

predicting the test chart structure by using the trained chart neural network model to obtain prediction information of each test node in the test chart structure;

determining an evaluation index value of the trained graph neural network model based on the prediction information of each test node and the corresponding label information;

determining that the evaluation index value is smaller than a preset index threshold value, and acquiring training data again;

and continuously training the trained graph neural network model by using the training data until the evaluation index value of the retrained graph neural network model reaches the index threshold value.

7. A graph data processing method, comprising:

responding to a received task processing request, acquiring to-be-processed graph structure data and a trained graph neural network model corresponding to the task processing request, wherein the trained graph neural network model is obtained by training by using the model training method of any one of claims 1 to 6;

predicting the graph structure data by using the trained graph neural network model to obtain a representation vector of each node in a graph structure, wherein the representation vector can represent global features and local features of the node;

and outputting the processing result.

8. The method according to claim 7, wherein when the task processing request is a recommendation request, the performing task processing based on the representation vector of each node to obtain a processing result includes:

determining a target node based on the task processing request;

and determining the matching object as a processing result.

9. The method according to claim 7, wherein when the task processing request is a classification request, performing task processing based on the representation vectors of the respective nodes to obtain a processing result includes:

10. An apparatus for training a neural network model, the apparatus comprising:

a second determining module, configured to determine a complementary graph structure of the training graph structure and a first adjacency matrix of the complementary graph structure, and obtain a second adjacency matrix of the training graph structure;

11. A graph data processing apparatus, characterized in that the apparatus comprises:

a second obtaining module, configured to obtain, in response to a received task processing request, to-be-processed graph structure data and a trained graph neural network model corresponding to the task processing request, where the trained graph neural network model is obtained by training using the model training method according to any one of claims 1 to 6;

and the result output module is used for outputting the processing result.

12. A computer device, characterized in that the computer device comprises:

a memory for storing executable instructions;

a processor for implementing the method of training a graph neural network model according to any one of claims 1 to 6, or the method of processing graph data according to any one of claims 7 to 9, when executing executable instructions stored in the memory.

13. A computer-readable storage medium storing executable instructions which, when executed by a processor, implement the method of training a graph neural network model according to any one of claims 1 to 6, or the method of processing graph data according to any one of claims 7 to 9.

14. A computer program product comprising a computer program or instructions, characterized in that the computer program or instructions, when executed by a processor, implement the method of training a graph neural network model according to any one of claims 1 to 6, or the method of graph data processing according to any one of claims 7 to 9.