Self-supervised learning graph recommendation method, system and equipment based on intention guidance and storage medium

Abstract

The invention discloses an intention-oriented self-supervised learning graph recommendation method, system and device and a storage medium, and the method comprises the steps: constructing a user-article interaction graph according to the interaction data of a user and an article, inputting the ID of the user-article in the user-article interaction graph into an intention-perceived graph encoder, obtaining a user-article embedded representation under each intention and aggregating the user-article embedded representation; inputting the aggregated user-article embedded representation into an online network in a trained self-supervised learning framework, and predicting a user-article recommendation score; wherein the self-supervised learning framework comprises an online network and a target network, the online network is used for predicting a user-article recommendation score according to the aggregated user-article embedded representation, and the target network is used for performing data enhancement on the aggregated user-article embedded representation in a training process. According to the method, the interest of the user can be better modeled according to the intention, so that more accurate and more efficient user recommendation is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence machine learning technology, and in particular to an intent-oriented self-supervised learning graph recommendation method, system, terminal device and computer-readable storage medium. Background Art

[0002] In the information age with the rapid development of the Internet and digital technology, people are faced with a huge amount of information and content. How to filter out information that meets the interests and preferences of users has become an urgent problem to be solved. Recommendation systems can recommend items and content that users may interact with in the future based on the items that users have interacted with before and various behavioral data, thereby improving the user experience, increasing user stickiness, and promoting the transaction of various commodities.

[0003] At present, most traditional recommendation algorithms only use collaborative filtering algorithms for recommendation. The disadvantage of this type of recommendation algorithm is that when training the recommendation algorithm, it will face the problems of data sparsity and false negative samples in the data. These problems will reduce the accuracy of the recommendation algorithm and thus affect the performance of the recommendation system. Summary of the invention

[0004] In order to solve at least one of the problems in the above-mentioned prior art, the present invention provides an intent-guided self-supervised learning graph recommendation method, system, terminal device and computer-readable storage medium, which can better model user interests and thus achieve more accurate and efficient user recommendations.

[0005] The first object of the present invention is to provide an intention-guided self-supervised learning graph recommendation method.

[0006] The second object of the present invention is to provide an intention-guided self-supervised learning graph recommendation system.

[0007] The third object of the present invention is to provide a terminal device.

[0008] A fourth object of the present invention is to provide a computer-readable storage medium.

[0009] The first object of the present invention can be achieved by adopting the following technical solutions:

[0010] A method for graph recommendation based on intention-oriented self-supervised learning, the method comprising:

[0011] Build a user-item interaction graph based on the interaction data between users and items, input the user-item IDs in the user-item interaction graph into the intent-aware graph encoder, obtain the user-item embedding representation under each intent, and aggregate the user-item embedding representation under each intent;

[0012] The aggregated user-item embedding representation is input into an online network in a trained self-supervised learning framework to predict the user-item recommendation score; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the user-item recommendation score according to the aggregated user-item embedding representation, and the target network is used to perform data enhancement on the aggregated user-item embedding representation during training.

[0013] Furthermore, the step of inputting the IDs of the user-item in the user-item interaction graph into the intent-aware graph encoder to obtain the embedded representation of the user-item under each intent includes:

[0014] Convert the IDs of users and items into learnable initial embedding representations, and then decompose the initial embedding representations into K different embedding representations according to K intents; K is a positive integer greater than 1;

[0015] Based on the embedding representation of each intent and the probability of the user interacting with the item under that intent, a user-item interaction graph is constructed for each intent.

[0016] Perform graph convolution operations on the user-item interaction graph under each intent to obtain the embedded representation of the user-item under each intent.

[0017] Furthermore, the data enhancement adopts a node specialization approach.

[0018] Furthermore, let the output of the online network be in, and are the output of the user and the output of the item respectively;

[0019] Assume the perturbation output of the target network is in, and are the output of the user and the output of the item respectively;

[0020] During the training process, back propagation loss is only performed on the online network, and the loss function is the negative cosine loss function C·; gradient propagation is stopped on the target network, and sg(·) is used as the stop gradient operator; the loss function of the self-supervised learning framework is:

[0021]

[0022] Furthermore, the cross-prediction method is used to predict the interaction score between users and items, which is the recommendation score:

[0023] s(E u ,E i )=h(E u )·(Ei ) T +E u ·h(E i ) T

[0024] Among them, s(E u ,E i ) is the recommendation score of user u for item i, (E u ,E i ) is the embedding representation of user-item, E u and E i are the embedding representations of users and items respectively; (·) T represents the transposition operation, and h(·) is the online network.

[0025] Furthermore, a denoising module is added to the self-supervised learning framework to achieve data denoising; the output of the denoising module is compared with the output of the target network to reduce the impact of noise during the training process.

[0026] Furthermore, the data denoising includes:

[0027] The user-item interaction graph is input into the denoising module, and the reliability of the user-item interaction is calculated based on the structural similarity between the node and its adjacent nodes;

[0028] The adjacency matrix is ​​reconstructed based on the reliability of the interaction between users and items, and the interactions whose reliability is lower than the set threshold are discarded to achieve denoising;

[0029] Perform graph convolution on the denoised adjacency matrix to obtain the denoised embedding representation of users and items, which is the output of the denoising module.

[0030] Furthermore, the online network is a single-layer perceptron.

[0031] The second object of the present invention can be achieved by adopting the following technical solutions:

[0032] An intention-guided self-supervised learning graph recommendation system, the system comprising:

[0033] The acquisition module is used to construct a user-item interaction graph based on the interaction data between users and items, input the IDs of users and items in the user-item interaction graph into the intent-aware graph encoder, obtain the embedded representation of users and items under each intent, and aggregate the embedded representation of users and items under each intent;

[0034] A prediction module is used to input the embedded representation of aggregated users and items into an online network in a trained self-supervised learning framework to predict the recommendation score of the user-item; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the recommendation score of the user-item based on the embedded representation of aggregated users and items, and the target network is used to perform data enhancement on the embedded representation of aggregated users and items during training.

[0035] The third object of the present invention can be achieved by adopting the following technical solutions:

[0036] A terminal device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned intention-oriented self-supervised learning graph recommendation method is implemented.

[0037] The fourth object of the present invention can be achieved by adopting the following technical solutions:

[0038] A computer-readable storage medium stores a program, which, when executed by a processor, implements the above-mentioned intention-oriented self-supervised learning graph recommendation method.

[0039] The present invention has the following beneficial effects compared with the prior art:

[0040] The present invention adopts self-supervised learning to effectively solve the problem of data sparsity; by designing corresponding auxiliary tasks, it can better utilize the information in the existing user-item interaction data; by adopting the intent-aware graph encoder, it can better model the various interests of users, thereby improving the performance of recommendations; compared with the BPR loss function adopted in many recommendation system methods, the present invention can only use the observed user-item interaction information when calculating the loss function in the self-supervised learning framework of the twin network, thereby avoiding the problem of false negative samples in the data and improving the accuracy of recommendations; the denoising module is combined with the recommendation task in the form of contrastive learning, thereby reducing the impact of data noise on the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0042] Figure 1 This is a flowchart of the intention-guided self-supervised learning graph recommendation method according to Embodiment 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of the intention-guided self-supervised learning graph recommendation method according to Embodiment 1 of the present invention;

[0044] Figure 3 This is a schematic diagram of a flow chart of an intent-aware image encoder according to Embodiment 1 of the present invention;

[0045] Figure 4 This is a schematic diagram of the data enhancement process of Example 1 of the present invention;

[0046] Figure 5 This is a schematic diagram of the process of data denoising in Example 1 of the present invention;

[0047] Figure 6 This is a structural block diagram of an intention-guided self-supervised learning graph recommendation system according to Embodiment 2 of the present invention;

[0048] Figure 7 This is a structural block diagram of the terminal device of embodiment 3 of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. It should be understood that the specific embodiments described are only used to explain the present application and are not used to limit the present application.

[0050] Embodiment 1:

[0051] like Figure 1 , 2 As shown, this embodiment provides an intention-oriented self-supervised learning graph recommendation method, comprising the following steps:

[0052] S101. Build a user-item interaction graph based on the interaction data between users and items, input the IDs of users and items in the user-item interaction graph into the intent-aware graph encoder, obtain the embedded representation of users and items under each intent, and aggregate them.

[0053] like Figure 3 As shown in FIG. 1 , the process of inputting the user-item ID into the intent-aware graph encoder includes:

[0054] First, the IDs of users and items are converted into learnable initial embedding representations, and then the initial embedding representations are decomposed into K different embedding representations according to K intents; K is a hyperparameter of the algorithm;

[0055] Secondly, the interaction between users and items is remodeled so that each different intention corresponds to an interaction probability, which indicates the probability of the user interacting with the item under the intention. This interaction information is used to create a user-item interaction graph under the intention.

[0056] Thirdly, after obtaining the intent-aware embedding representation and the intent-aware interaction probability, we can use the graph convolution network to perform graph convolution operations on the user-item interaction graph for each intent, thereby obtaining the user-item embedding representation (E uk ,E ik ), where k = 1…K; E uk represents the embedding representation of the user under the kth intent, E ik Represents the embedding representation of the item under the k-th intent;

[0057] Finally, after obtaining the embedding representations of users and items under each intent, these embedding representations need to be aggregated:

[0058] E u =CONCAT(E uk ,=1…K), i =CONCAT(E ik ,=1…K)

[0059] Among them, CONCAT(· represents the embedded connection operation, E u , E i They are the aggregated embedding representations of users and items, respectively.

[0060] The aggregated embedding representation E u , E i is the output of the intent-aware graph encoder.

[0061] In this embodiment, the value of K is 3.

[0062] This step mainly divides the user-item embedding representation into multiple different embedding representations, which can well model the intention of user-item interaction, thereby effectively decoupling the intention and extracting information related to each intention.

[0063] S102: Input the aggregated user-item embedding representation into a self-supervised learning framework to predict the user-item recommendation score.

[0064] The self-supervised learning framework mainly consists of two different neural networks, and the outputs of these two neural networks are used for recommendation tasks.

[0065] The two neural networks are the online network and the target network. The main structure of the online network is a predictor h, which can be different neural networks and is used to recommend the main task. The target network is an embedding enhancer that can perturb or modify the embedded representation of the input to enhance the representation of the positive sample. The target network enhances the output of the intent-aware graph encoder to improve the generalization and robustness of the model, enrich the training data of the model, and reduce the risk of overfitting.

[0066] Specifically, this embodiment uses a single-layer perceptron as the predictor h.

[0067] Specifically, in this embodiment, the online network implements a specific perturbation on the representation of the output of the intent-aware graph encoder to enhance the data, which can avoid destroying the original structure of the graph. This embodiment does not use random noise to perform perturbations, but instead uses a node-specific data enhancement method. By using a perturbation method with a globally aware representation, the representation of each node has a certain global perspective.

[0068] like Figure 4 As shown, taking user node u as an example, all items are uniformly randomly sampled with probability ρ to obtain a candidate item set is the set of all items; then the cosine similarity between user u and each item in the candidate item set is calculated, and the M item sets with the closest similarity to user u are obtained for the subsequent perturbation of the node embedding of user u. M is a hyperparameter, and then the perturbation vector e of the corresponding node is obtained by taking the average of the item embeddings in the item set. p , e p The calculation method is:

[0069]

[0070] In this way, the perturbation vector of node u is obtained, and then the hyperparameter ε is used to modify the size of the perturbation vector. The final perturbation of user node u is expressed as follows:

[0071]

[0072] The calculation method of the enhanced representation of items is similar to that of users. This method allows each node representation to obtain a global view and to some extent utilize the information of negative samples in the data. Because items with high similarity to the aggregated user representation are likely to be unmined false negative samples, this enhancement method is also conducive to improving the representation ability of user nodes.

[0073] It can be understood that the embedded representations of users and items output by the intent-aware graph encoder are input into the online network and the target network respectively, and the output of the online network is in, For user output, is the output of the item; the perturbation output of the target network is in, For user output, Output of items.

[0074] The loss function of the online network in this embodiment is the negative cosine loss function C(·), the main purpose of which is to maximize the prediction of a user on the perturbed item and the prediction of an item on the perturbed user. Because only positive samples are used in training, the user-item pair can still have a high interaction score after the training is completed. For this reason, this embodiment only performs back propagation loss on the online network during training, stops gradient propagation on the target network, and uses sg(·) as the stop gradient operator. The final loss function of the self-supervised learning framework is as follows:

[0075]

[0076] The prediction of the traditional collaborative filtering model is defined as the inner product of the embedding representation of the user and the item. This embodiment uses the cross-prediction method to predict the interaction score between the user and the item. Let (E u ,E i ) is the output of the intent-aware graph encoder, the final recommendation score is calculated as:

[0077] s(E u ,E i )=h(E u )·(E i ) T +E u ·h(E i ) T

[0078] Among them, s(E u ,E i ) represents the recommendation score of end user u for item i, and h(·) represents the online network mentioned above.

[0079] This step only uses the observed user-item interaction data when performing the recommendation task, so it can avoid the problem of false negative samples in the dataset.

[0080] In order to reduce the impact of data noise during the training process, this embodiment also adds a denoising module to the self-supervised learning framework.

[0081] The denoising module is used to achieve data denoising. The output of the denoising module is compared with the output of the target network to reduce the impact of noise during training.

[0082] There are some noisy interactions in the training data of existing recommendation systems, and graph neural networks use a message passing scheme that aggregates neighborhood information, which may aggregate a lot of noisy information together, thereby amplifying the impact of noise. In order to denoise the data, this embodiment adds a denoising module to the original recommendation framework, and uses the output of the denoising module to perform comparative learning with the output of the target network to assist the main task of recommendation, thereby reducing the impact of noise in the training data.

[0083] The denoising module implements data denoising mainly in two steps: interactive reliability modeling and denoising based on interactive reliability.

[0084] (1) Interaction reliability modeling.

[0085] Mainly calculates the reliability of the interaction between users and items.

[0086] Nodes with similar features or structures are more likely to interact with each other than nodes with different features or structures, so the reliability of the interaction is modeled based on the similarity of features or structures. However, in the recommendation task, the embedding initialization of users and items is randomly initialized based on their IDs, so the features of the nodes are unavailable. Therefore, the reliability of the interaction data is estimated based on the structural similarity between the node and its neighboring nodes in the interaction graph. The greater the structural similarity between nodes, the lower the probability that there is interaction noise between the target node and its neighboring nodes that interact with it. For example, an outlier user with many erroneous interactions with items will result in low structural consistency with other users in the collaborative relationship of the interaction data. In each layer of the graph neural network, the structural information of the user and item nodes is first obtained:

[0087]

[0088] in, They represent the structural information of users and items respectively, l is the number of layers of the current neural network, and A is the adjacency matrix of the user-item interaction graph.

[0089] In order to estimate the reliability of the interaction between user u and item i, the structural similarity between them in the interaction graph is used to measure it. The structural similarity is calculated using cosine similarity. After calculating the cosine similarity, it is also necessary to normalize the similarity so that it is between [0,1]. The structural similarity obtained after normalization is the reliability of the interaction. Specifically, the reliability of the interaction between user u and item i is The calculation method is:

[0090]

[0091] in, and They are The embedding representation of user u in The embedding representation of item i in represents the structural characteristics of user u and item i.

[0092] (2) Denoising based on interactive reliability.

[0093] Denoising is mainly performed based on the reliability of the interaction between users and items.

[0094] The adjacency matrix is ​​reconstructed using the reliability of the interactions between users and items, and the interactions whose reliability is lower than a set threshold are discarded to achieve denoising.

[0095] Specifically, first, the interaction reliability between user u and item i is used to replace the value of the adjacency matrix in the user-item interaction graph; then, the interaction reliability The weights of interactions below the threshold β are set to 0, where β is a hyperparameter of the algorithm. This way, the denoised adjacency matrix of the user-item interaction graph is obtained. Finally, graph convolution is performed on the denoised adjacency matrix to obtain the denoised embedding representation of users and items, which is the output of the denoising module.

[0096] After obtaining the denoised embedding representations of users and items, the denoising module is combined with the target network in the form of contrastive learning, which is achieved by maximizing the mutual information of different views of the same node and minimizing the mutual information of different views of different nodes. Formally, the InfoNCE loss function commonly used in contrastive learning is used as the loss function of contrastive learning, specifically:

[0097]

[0098] in, is the output representation of the target network of user u in the self-supervised learning module, h u is the output representation of user u in the denoising step; τ is a hyperparameter, called the temperature coefficient, which is used to control the scale of similarity.

[0099] Similarly, the contrast loss of the item can be obtained Combining these two loss functions, we can get the target loss function that uses contrastive learning between the denoising module and the target network:

[0100] The data denoising process can be referred to Figure 5 .

[0101] Since the main goal of the recommendation task is to simulate the interaction between users and items, the loss function of the contrastive learning is regarded as a supplement. The recommendation task and the self-supervised learning task are jointly optimized through a multi-task learning strategy. The complete loss function of the final overall model is:

[0102]

[0103] Among them, Θ is the set of parameters in the complete model, λ1 and λ2 are hyperparameters that control the strength of self-supervised learning and L2 regularization, respectively.

[0104] It is understandable that the model in this embodiment specifically refers to a self-supervised learning framework. Before adding the denoising module, the self-supervised learning framework includes a prediction network and an online network; after adding the denoising module, it also includes the denoising module.

[0105] Those skilled in the art will appreciate that all or part of the steps in the method for implementing the above embodiments may be completed by instructing related hardware through a program, and the corresponding program may be stored in a computer-readable storage medium.

[0106] It should be noted that although the method operations of the above embodiments are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. On the contrary, the steps depicted can be performed in a different order. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple steps.

[0107] Embodiment 2:

[0108] like Figure 6 As shown, this embodiment provides an intention-oriented self-supervised learning graph recommendation system, which includes an acquisition module 601 and a prediction module 602, wherein:

[0109] An acquisition module 601 is used to construct a user-item interaction graph based on the interaction data between users and items, input the IDs of users and items in the user-item interaction graph into an intent-aware graph encoder, obtain the embedded representation of users and items under each intent, and aggregate the embedded representation of users and items under each intent;

[0110] The prediction module 602 is used to input the embedded representation of aggregated users and items into the online network in the trained self-supervised learning framework to predict the recommendation score of the user-item; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the recommendation score of the user-item according to the embedded representation of aggregated users and items, and the target network is used to perform data enhancement on the embedded representation of aggregated users and items during the training process.

[0111] The specific implementation of each module in this embodiment can refer to the above-mentioned embodiment 1, which will not be described one by one here; it should be noted that the system provided in this embodiment is only illustrated by the division of the above-mentioned functional modules. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0112] Embodiment 3:

[0113] This embodiment provides a terminal device, which can be a computer. Figure 7 As shown, a processor 702, a memory, an input device 703, a display 704 and a network interface 705 connected via a system bus 701 are provided. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 706 and an internal memory 707. The non-volatile storage medium 706 stores an operating system, a computer program and a database. The internal memory 707 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 702 executes the computer program stored in the memory, the intention-oriented self-supervised learning graph recommendation method of the above-mentioned embodiment 1 is implemented as follows:

[0114] Build a user-item interaction graph based on the interaction data between users and items, input the user-item IDs in the user-item interaction graph into the intent-aware graph encoder, obtain the user-item embedding representation under each intent, and aggregate the user-item embedding representation under each intent;

[0115] The aggregated user-item embedding representation is input into an online network in a trained self-supervised learning framework to predict the user-item recommendation score; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the user-item recommendation score according to the aggregated user-item embedding representation, and the target network is used to perform data enhancement on the aggregated user-item embedding representation during training.

[0116] Embodiment 4:

[0117] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the method for recommending graphs based on intention-oriented self-supervised learning in the above embodiment 1 is implemented as follows:

[0118] Build a user-item interaction graph based on the interaction data between users and items, input the user-item IDs in the user-item interaction graph into the intent-aware graph encoder, obtain the user-item embedding representation under each intent, and aggregate the user-item embedding representation under each intent;

[0119] The aggregated user-item embedding representation is input into an online network in a trained self-supervised learning framework to predict the user-item recommendation score; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the user-item recommendation score according to the aggregated user-item embedding representation, and the target network is used to perform data enhancement on the aggregated user-item embedding representation during training.

[0120] It should be noted that the computer-readable storage medium of the present embodiment may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0121] The above is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and inventive concept of the present invention within the scope disclosed by the present invention, which shall fall within the protection scope of the present invention.

Claims

1. An intention-guided self-supervised learning graph recommendation method, characterized in that: The method comprises: Build a user-item interaction graph based on the interaction data between users and items, input the user-item IDs in the user-item interaction graph into the intent-aware graph encoder, obtain the user-item embedding representation under each intent, and aggregate the user-item embedding representation under each intent; The aggregated user-item embedding representation is input into an online network in a trained self-supervised learning framework to predict the user-item recommendation score; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the user-item recommendation score according to the aggregated user-item embedding representation, and the target network is used to perform data enhancement on the aggregated user-item embedding representation during training.

2. The self-supervised learning graph recommendation method according to claim 1, characterized in that: The IDs of the user-items in the user-item interaction graph are input into the intent-aware graph encoder to obtain the embedded representation of the user-items under each intent, including: Convert the IDs of users and items into learnable initial embedding representations, and then decompose the initial embedding representations into K different embedding representations according to K intents; K is a positive integer greater than 1; Based on the embedding representation of each intent and the probability of the user interacting with the item under that intent, a user-item interaction graph is constructed for each intent. Perform graph convolution operations on the user-item interaction graph under each intent to obtain the embedded representation of the user-item under each intent.

3. The self-supervised learning graph recommendation method according to claim 1, characterized in that: The data enhancement adopts a node specialization method.

4. The self-supervised learning graph recommendation method according to claim 1, characterized in that: Assume that the output of the online network is in, and are the output of the user and the output of the item respectively; Assume the perturbation output of the target network is in, and are the output of the user and the output of the item respectively; During the training process, only the back propagation loss is performed on the online network, and the loss function is the negative cosine loss function C(·); the gradient propagation is stopped on the target network, and sg(·) is used as the stop gradient operator; then the loss function of the self-supervised learning framework is:

5. The self-supervised learning graph recommendation method according to claim 1, characterized in that: The cross-prediction method is used to predict the interaction score between users and items, which is the recommendation score: if u ,AND i )=h(E u )·(AND i ) T +E u ·h(E i ) T Among them, s(E u ,E i ) is the recommendation score of user u for item i, (E u ,E i ) is the embedding representation of user-item, E u and E i are the embedding representations of users and items respectively; (·) T represents the transposition operation, and h(·) is the online network.

6. The self-supervised learning graph recommendation method according to any one of claims 1 to 5, characterized in that: A denoising module is added to the self-supervised learning framework to achieve data denoising; the output of the denoising module is compared with the output of the target network to reduce the impact of noise during training.

7. The self-supervised learning graph recommendation method according to claim 6, characterized in that: The data denoising comprises: The user-item interaction graph is input into the denoising module, and the reliability of the user-item interaction is calculated based on the structural similarity between the node and its adjacent nodes; The adjacency matrix is ​​reconstructed based on the reliability of the interaction between users and items, and the interactions whose reliability is lower than the set threshold are discarded to achieve denoising; Perform graph convolution on the denoised adjacency matrix to obtain the denoised embedding representation of users and items, which is the output of the denoising module.

8. The self-supervised learning graph recommendation method according to any one of claims 1 to 5, characterized in that: The online network is a single-layer perceptron.

9. An intention-guided self-supervised learning graph recommendation system, characterized in that: The system comprises: The acquisition module is used to construct a user-item interaction graph based on the interaction data between users and items, input the IDs of users and items in the user-item interaction graph into the intent-aware graph encoder, obtain the embedded representation of users and items under each intent, and aggregate the embedded representation of users and items under each intent; A prediction module is used to input the embedded representation of aggregated users and items into an online network in a trained self-supervised learning framework to predict the recommendation score of the user-item; wherein the self-supervised learning framework includes an online network and a target network, the online network is used to predict the recommendation score of the user-item based on the embedded representation of aggregated users and items, and the target network is used to perform data enhancement on the embedded representation of aggregated users and items during training.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the self-supervised learning graph recommendation method according to any one of claims 1 to 8 is implemented.

Images