Recommendation method, system, electronic device and computer-readable medium based on graph neural network

Through the single-layer neural network fitting the convergence state of the graph neural network and explicit learning of product co-occurrence graph information, combined with negative sampling technology, the problem of the reduction in the recommended effect when the number of layers of the graph convolution neural network increases, and the training of the recommendation system is improved Efficiency and accuracy make models easier to deploy.

CN114925268BInactive Publication Date: 2024-12-03SOUTH CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210468095.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-29
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing recommended method based on graph neural networks is used to reduce the number of layers of graph convolutional neural networks, and it is impossible to effectively use node embedding vectors to express the user's preference for products.

Method used

The convergence state of the neural network is adopted for single-layer neural network fitting graph, and by constructing the user-product binary graph and product co-occurrence matrix, the loss function is optimized, negative sampling data enhancement is introduced, and pseudo-negative samples are constructed to improve the difficulty and accuracy of model training data. .

Benefits of technology

It improves the training efficiency and accuracy of the recommendation system, makes the model easier to deploy, solves the problem that the recommendation effect decreases when the number of layers increases, and further improves the recommendation accuracy by explicitly learning product co-occurrence graph information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114925268B_ABST
    Figure CN114925268B_ABST
Patent Text Reader

Abstract

The present invention discloses a recommendation method based on a graph neural network, including: constructing a user-item bipartite graph based on the interaction data between users and items; initializing all user embedding vectors and item embedding vectors in the user-item bipartite graph; respectively calculating a fitted user embedding vector and a fitted item embedding vector based on the relationship between the item embedding vector and the user embedding vector in the converged state of the graph neural network; obtaining a loss function based on the fitted user embedding vector and the fitted item embedding vector; and back-propagating based on the loss function to update the user embedding vector and the item embedding vector. The present invention also discloses a recommendation system, an electronic device, and a computer-readable medium based on a graph neural network. Compared with existing recommendation systems based on graph neural networks, the present invention has higher training efficiency and is easier to deploy in actual production scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet recommendation systems, and specifically relates to a recommendation method, system, electronic device and computer-readable medium based on a graph neural network. Background Art

[0002] With the rapid development of the Internet industry, a huge amount of information is generated every day, forcing people to face a problem - information overload. Recommendation System is an advanced business intelligence platform based on massive data mining, providing customers with personalized information services and decision support. Figure 1 The logical framework of the personalized recommendation system shown in the figure outputs a list of recommended items that the user may be interested in and may interact with from a candidate item library for a specific triple of user information, item information, and scene information.

[0003] Graph Neural Network (GNN) is a neural network based on graph data structure, which contains the following basic elements: node, edge, and node embedding. Each time, the adjacent neighbor node information is aggregated, and the embedding representation of the current central node is updated. After the same operation is performed on all nodes in the graph, multiple layers are iterated, and finally the node can aggregate the information of multi-hop neighbor nodes. In recent years, graph neural networks have received more and more attention, and have been widely used in text classification, neural machine translation, image classification, recommendation systems, and other fields.

[0004] During the training of graph convolutional neural networks, as the network depth increases and the number of iterations increases, all node embedding vectors of the same connected component in the graph tend to converge to the same value. In existing collaborative filtering recommendation solutions based on graph neural networks, there is only one connected component in the entire graph structure. This results in all node embedding vectors in the graph tending to converge to the same value when there are many layers. Naturally, the inner product of the node embedding vector cannot be used to represent the user's preference for the product, and no useful information can be learned. This results in poor final recommendation results of the current method. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a recommendation method, system, electronic device and computer-readable medium based on a graph neural network with high training efficiency and easy deployment.

[0006] In the first aspect, a recommendation method based on a graph neural network comprises the following steps:

[0007] Based on the interaction data between users and items, a user-item bipartite graph is constructed;

[0008] Initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph;

[0009] According to the relationship between the product embedding vector and the user embedding vector in the converged state of the graph neural network, the fitted user embedding vector and the fitted product embedding vector are calculated respectively;

[0010] Obtaining a loss function according to the fitted user embedding vector and the fitted product embedding vector;

[0011] According to the loss function, back propagation is performed to update the user embedding vector and the item embedding vector.

[0012] As a preferred solution, the following steps are also included:

[0013] Constructing a product co-occurrence matrix G according to the user-product bipartite graph;

[0014] Convert the commodity co-occurrence matrix G into a sparse matrix;

[0015] According to the similarity relationship between products in the product co-occurrence matrix G, the loss function is optimized.

[0016] As a preferred solution, the following steps are also included:

[0017] Extract at least one original negative sample, and calculate the embedding vectors of all the extracted original negative samples;

[0018] An embedding vector of at least one positive sample is extracted, and is mixed with the embedding vectors of all original negative samples to synthesize an embedding vector of a pseudo negative sample, where the embedding vector of the pseudo negative sample is used to fit the calculation of the product embedding vector.

[0019] As a preferred solution, it also includes:

[0020] If there are multiple original negative samples extracted, for a specific user u, an inner product operation is performed between the embedding vector of the user u and the embedding vectors of all the pseudo negative samples, and the pseudo negative sample with the largest inner product is selected; and / or

[0021] The mixing process uses an interpolation operation.

[0022] In the second aspect, a recommendation system based on a graph neural network includes:

[0023] The bipartite graph construction module is used to construct a user-item bipartite graph based on the interaction data between users and items;

[0024] An initialization module, used to initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph;

[0025] The fitting embedding vector module is used to calculate the fitting user embedding vector and the fitting product embedding vector respectively according to the relationship between the product embedding vector and the user embedding vector in the converged state of the graph neural network;

[0026] A first loss module, used for obtaining a loss function according to the fitted user embedding vector and the fitted product embedding vector;

[0027] An updating module is used to back-propagate according to the loss function to update the user embedding vector and the item embedding vector.

[0028] As a preference, it also includes:

[0029] A co-occurrence matrix construction module, used to construct a product co-occurrence matrix G according to the user-product bipartite graph;

[0030] A sparse module, used to convert the commodity co-occurrence matrix G into a sparse matrix;

[0031] The second loss module is used to optimize the loss function according to the similarity relationship between commodities in the commodity co-occurrence matrix G.

[0032] As a preference, it also includes:

[0033] A negative sample extraction module, used to extract at least one original negative sample and calculate the embedding vectors of all the extracted original negative samples;

[0034] The pseudo negative sample synthesis module is used to extract the embedding vector of at least one positive sample, mix it with the embedding vectors of all original negative samples, and synthesize the embedding vector of the pseudo negative sample. The embedding vector of the pseudo negative sample is used to fit the calculation of the product embedding vector.

[0035] As a preference, it also includes:

[0036] The pseudo negative sample selection module is used to perform an inner product operation on the embedding vector of the user u and the embedding vectors of all the pseudo negative samples for a specific user u if there are multiple original negative samples extracted, and select the pseudo negative sample with the largest inner product.

[0037] According to a third aspect, an electronic device includes:

[0038] one or more processors;

[0039] a storage device having one or more programs stored thereon;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement a recommendation method based on a graph neural network as described in the first aspect.

[0041] In a fourth aspect, a computer-readable medium stores a computer program thereon, wherein when the program is executed by a processor, the recommendation method based on a graph neural network as described in the first aspect is implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. By using a single-layer neural network to fit the convergence state of the graph neural network, the topological information hidden in the user-item bipartite graph is obtained. This solves the technical problem that the recommendation effect of the traditional recommendation method based on the graph convolutional neural network will decrease when the number of graph convolutional neural network layers is large due to the constraints of the graph structure. This improves the training efficiency and makes the model easier to deploy in actual production scenarios.

[0044] 2. By explicitly learning the information of the product co-occurrence graph, the training efficiency of the recommendation system is further improved;

[0045] 3. By introducing negative sampling as a data enhancement method and constructing pseudo-negative samples that are more difficult to classify, the training data of the model can be expanded, which can effectively improve the accuracy of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] 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 these drawings without paying creative work.

[0047] Figure 1 It is a flowchart diagram of steps S101 to S105 of a recommendation method based on a graph neural network of the present invention;

[0048] Figure 2 (a) is a schematic diagram of the structure of a neural network in the prior art;

[0049] Figure 2 (b) is a schematic diagram of a graph neural network structure of a recommendation method based on a graph neural network of the present invention;

[0050] Figure 3 It is a flowchart diagram of steps S201 to S203 of a recommendation method based on a graph neural network of the present invention;

[0051] Figure 4 A schematic diagram of a commodity co-occurrence matrix G of a recommendation method based on a graph neural network of the present invention;

[0052] Figure 5It is a flowchart diagram of steps S301 to S302 of a recommendation method based on a graph neural network of the present invention;

[0053] Figure 6 It is a flowchart diagram of steps S301 to S303 of a recommendation method based on a graph neural network of the present invention;

[0054] Figure 7 This is a structural schematic diagram of a recommendation system based on graph neural network in the present invention.

[0055] Among them, 1 single-layer fitting convergence state module; 11 bipartite graph construction module; 12 initialization module; 13 fitting embedding vector module; 14 first loss module; 15 update module;

[0056] 2 commodity co-occurrence module; 21 co-occurrence matrix construction module; 22 sparsification module; 23 second loss module;

[0057] 3 Negative sampling enhancement module; 31 Negative sample extraction module; 32 Pseudo negative sample synthesis module; 33 Pseudo negative sample selection module. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of 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.

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] First, a recommendation method based on graph neural network, such as Figure 1 As shown, the following steps are included:

[0061] S101: Construct a user-item bipartite graph based on the interaction data between users and items.

[0062] In this step, the user-item bipartite graph is a bipartite graph used to reflect the interactive relationship between users and items. Users and various items are regarded as nodes in the user-item bipartite graph; users' browsing, clicking, subscribing, purchasing, liking, coining, rewarding and other operations can all be regarded as edges in the bipartite graph.

[0063] S102: Initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph.

[0064] In this step, the initialization may be performed in a random manner, or certain fixed features may be added when initializing the embedding vector according to the characteristics of the user.

[0065] S103: According to the relationship between the product embedding vector and the user embedding vector in the converged state of the graph neural network, a fitted user embedding vector and a fitted product embedding vector are calculated respectively.

[0066] In this step, the convergence state of the graph neural network refers to the state in which the embedding vectors of all nodes of the graph neural network converge to a value when the number of layers of the graph neural network increases infinitely; the fitted user / product embedding vector refers to the user / product embedding vector fitted under the convergence state. The inner product of the fitted user embedding vector and the fitted product embedding vector can be used to obtain personalized product recommendation results for users.

[0067] S104: Obtain a loss function according to the fitted user embedding vector and the fitted product embedding vector.

[0068] In this step, the loss function can be formed by adding the loss term obtained by the conventional constraint and the loss term obtained by the constraint required to approximate the convergence state of the graph neural network. The conventional constraint can be a constraint that ensures that the user likes positive samples and dislikes negative samples.

[0069] S105: According to the loss function, back-propagate to update the user embedding vector and the item embedding vector.

[0070] In this embodiment, the recommendation method based on graph neural network shown in steps S101-S105 is based on the principle of modifying the graph neural network structure, no longer using the multi-layer stacking aggregation transmission method, but using a new method of mining implicit information in the graph, that is, using a single-layer neural network to fit the convergence state of the graph neural network to obtain the topological information hidden in the user-product bipartite graph. It can solve the technical problem that the traditional recommendation method based on graph convolutional neural network is restricted by the graph structure, and when the number of graph convolutional neural network layers is greater than 3, the recommendation effect will decrease.

[0071] In this embodiment, the specific principles of steps S103 to S104 are described as follows:

[0072] In the traditional collaborative filtering recommendation system based on graph neural network, the most common update aggregation formula in graph neural network is shown in formula (1):

[0073]

[0074] Among them, e mrepresents the embedding vector of node m (in this embodiment, it can represent the user embedding vector e u and product embedding vector e i ), the superscripts (l+1) and (l) in the upper right corner represent the current layer number; represents the set of all neighbor nodes of node m; d m represents the degree of node m (degree refers to the number of neighbors of a node in the graph; a node with only one neighbor has a degree of 1, and a node with three neighbors has a degree of 3); k is the number of neighbors in the graph. The node selected from (in this embodiment, if node m is a user embedding vector, then k is a product embedding vector; if node m is a product embedding vector, then k is a user embedding vector); d k represents the degree of node k.

[0075] If we increase the number of graph neural network layers infinitely, then the embedding vectors of all nodes will eventually converge to the same value, as shown below:

[0076]

[0077] In the converged state, the node embedding representations of the last two layers should be consistent. From formula (2), we can see that in the converged state:

[0078]

[0079] At this time, although the nodes in the graph contain the topological information of the graph, they cannot describe the unique information of each user / product node. Substituting formula (1) into the right side of formula (3), we can get the following relationship between the central node and the neighboring nodes in the graph neural network after the final training:

[0080]

[0081] Formula (4) describes the relationship that should be satisfied between the embedding vector of the central node and the embedding vector of the neighboring nodes when the graph neural network is in a converged state. The applicant found that if the opposite is considered: for all nodes in the graph, when the embedding vectors of the central node and its neighboring nodes satisfy the above relationship, the embedding vectors of the nodes in the graph also contain part of the information when convergence is reached. This can also achieve an effect similar to traditional message passing.

[0082] Therefore, in this embodiment, the number of layers of the graph neural network is no longer stacked, but the state of the node at the final convergence is directly approximated. Generally speaking, the inner product can be used to roughly measure the similarity between two vectors, so this embodiment adopts the most intuitive approximation method: when the graph neural network reaches the convergence state, the current center node and its neighbor nodes satisfy formula (4). Using formula (4), the fitted user embedding vector e is calculated respectively uand the fitted product embedding vector e i , and then perform an inner product operation on the two. The larger the result, the closer the embedding vectors on both sides of the equal sign are. The embedding vector of the node is also closer to the state when the graph neural network converges, and naturally contains the structural information of the entire graph to a certain extent.

[0083] In summary, the constraints that need to be satisfied to achieve the convergence state of the direct approximation graph neural network are as follows:

[0084]

[0085] In this formula, u represents the user, U represents the set of all users, and e u represents the fitted user embedding vector, e i represents the fitted product embedding vector, (d u represents the degree of user node u, d i represents the degree of the product node). Formula (5) is equivalent to fitting the user embedding vector e u and the fitted product embedding vector e i Maximize the cosine similarity between .

[0086] To facilitate optimization, this embodiment further adds a sigmoid activation function and a negative log-likelihood loss to equation (5), and obtains the following loss term L: C :

[0087]

[0088] Remove the new loss term L due to the change in network structure C In addition, the recommendation model should also have conventional constraints. In this embodiment, the constraint of ensuring that users like positive samples and dislike negative samples is selected. This constraint can be in the following form:

[0089]

[0090] Wherein, σ represents the sigmoid function. The loss function L = L O +L C After obtaining the loss function, we can use positive and negative samples, and use back propagation and gradient descent to learn the appropriate user / item embedding vector. Finally, we can use the vector inner product to reflect the user's preference for the item, and finally generate a personalized recommendation list for the specified user.

[0091] like Figure 2 As shown in (a), the traditional graph neural network model needs to go through multiple layers of transmission, and then learn the embedding vector of the user / product, and finally use the user embedding vector e uand product embedding vector e i The inner product between is used to measure the user's preference for the product. However, this embodiment directly skips the multi-layer graph neural network aggregation and transmission process, such as Figure 2 As shown in (b), by selecting the network state when direct fitting converges, the embedding vector of the product / user can be directly learned, and the user's preference for the product can be obtained by fitting the inner product of the user embedding vector and the product embedding vector, and a personalized recommendation list recommended to the specified user is generated. This change improves the training efficiency and makes the model in this embodiment easier to deploy in actual production scenarios.

[0092] As a preferred embodiment, Figure 3 As shown, the aforementioned recommendation method based on graph neural network may also include the following steps:

[0093] S201: Figure 4 As shown, a product co-occurrence matrix G is constructed based on the user-product bipartite graph.

[0094] In this step, the commodity co-occurrence matrix G can be obtained by the following formula:

[0095] G=A T A (8);

[0096] Where A is the adjacency matrix (interaction matrix) of the user-item bipartite graph. In the item co-occurrence matrix G, the diagonal elements G i,i Indicates the number of times product i appears (interacts) among all users. The off-diagonal element G i,j Represents the number of times items i and j co-occur (interact) among all users.

[0097] S202: Convert the product co-occurrence matrix G into a sparse matrix.

[0098] In this step, since the product co-occurrence matrix G is dense, directly using the product co-occurrence matrix G to assist in training the graph neural network will lead to the introduction of too much noise and unsatisfactory learning effect. Therefore, it is necessary to convert the product co-occurrence matrix G into a sparse matrix first.

[0099] S203: Optimize the loss function according to the similarity relationship between the products in the product co-occurrence matrix G.

[0100] In this step, the similarity relationship between product i and product j in the product co-occurrence matrix G can be expressed as follows:

[0101]

[0102] Among them, ω i,j It is used to measure the similarity between product i and product j. i,jrepresents the value of the (i, j) position in the commodity co-occurrence matrix G, g i Represents the degree of product i in the product co-occurrence graph.

[0103] After obtaining the sparse product co-occurrence matrix, the user-product interaction matrix can be used to assist in mining the implicit information in the product co-occurrence matrix. This can make the loss function have a consistent form, making the model easier to train. In summary, the information contained in the product co-occurrence matrix can be obtained by adding the following loss term L I get:

[0104]

[0105] In this formula, for a specified product i, the first k products that are most similar to product i are selected, and the set of these products is recorded as S(i).

[0106] The loss function L = L used to train the neural network in this embodiment O +L C +L I .

[0107] The applicant found in the research that explicitly learning the information contained in the product co-occurrence graph has a significant positive impact on improving the accuracy of the recommendation system. This embodiment further improves the recommendation accuracy of the personalized recommendation system by further adding information from the product co-occurrence graph to the recommendation system based on the graph neural network for training.

[0108] The recommendation method proposed in the above embodiment uses a bipartite graph data structure to store basic user-item interaction information, which changes the structure of the traditional graph neural network and makes it possible to flexibly learn the information contained in the graph data structure (such as user-item interaction graph and item co-occurrence graph, etc.).

[0109] As a preferred embodiment, Figure 5 As shown, the aforementioned recommendation method based on graph neural network may also include the following steps:

[0110] S301: extract at least one original negative sample, and calculate the embedding vectors of all the extracted original negative samples;

[0111] S302: extracting an embedding vector of at least one positive sample, mixing it with the embedding vectors of all original negative samples, and synthesizing an embedding vector of a pseudo negative sample, wherein the embedding vector of the pseudo negative sample is used for fitting the calculation of the product embedding vector.

[0112] Among them, the original negative sample refers to the commodity node that does not interact with the user node in the user-commodity interaction graph. This embodiment introduces negative sampling as a data enhancement method to expand the training data of the model, which can effectively improve the accuracy of the recommendation method. Specifically, samples that are difficult for a classifier to classify correctly are called "difficult samples" in this field. Difficult samples contain more valuable information, and training with difficult samples can obtain better classification boundaries. The applicant has found that the "pseudo-negative samples" artificially constructed in this embodiment are "difficult samples" that are more difficult to classify correctly than the ordinary and real negative samples used in the existing negative sampling mode; therefore, training with the "pseudo-negative samples" to form "difficult samples" can obtain better classification boundaries.

[0113] Furthermore, the mixing process may be an interpolation operation, specifically as shown in the following formula (11):

[0114]

[0115] in, represents the embedding vector of the pseudo negative sample, represents the embedding vector of the original negative sample, and m represents the number of negative samples extracted (m should be much smaller than the total number of negative samples). Represents the embedding vector of the positive sample (i.e., the product node that has interacted with the user). α represents the coefficient for mixing the existing negative samples with the positive samples, which is a random number between 0 and 1.

[0116] Further, after embedding the vector of the synthetic pseudo negative sample, as Figure 6 As shown, the following steps are also included:

[0117] S303: If there are multiple extracted original negative samples, for a specific user u, an inner product operation is performed between the embedding vector of the user u and the embedding vectors of all the pseudo negative samples, and a pseudo negative sample with the largest inner product is selected.

[0118] Among them, using the pseudo-negative samples with the largest inner product selected for training can further obtain better classification boundaries.

[0119] On the second aspect, a recommendation system based on graph neural network, such as Figure 7 As shown, it includes a single-layer fitting convergence state module 1, specifically:

[0120] A bipartite graph construction module 11, used to construct a user-item bipartite graph based on the interaction data between users and items;

[0121] An initialization module 12, used to initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph;

[0122] A fitting embedding vector module 13, used to calculate a fitting user embedding vector and a fitting product embedding vector respectively according to the relationship between the product embedding vector and the user embedding vector in a converged state of the graph neural network;

[0123] A first loss module 14, configured to obtain a loss function according to the fitted user embedding vector and the fitted product embedding vector;

[0124] The updating module 15 is used for back-propagating according to the loss function to update the user embedding vector and the item embedding vector.

[0125] The single-layer fitting convergence state module 1 of this embodiment corresponds to steps S101 to S105 of the method recommended in the first aspect.

[0126] As a preferred embodiment, the recommendation device based on graph neural network is as follows: Figure 7 As shown, it also includes a commodity co-occurrence module 2, which specifically includes:

[0127] A co-occurrence matrix construction module 21 is used to construct a product co-occurrence matrix G according to the user-product bipartite graph;

[0128] A sparseness module 22, used to convert the commodity co-occurrence matrix G into a sparse matrix;

[0129] The second loss module 23 is used to optimize the loss function according to the similarity relationship between commodities in the commodity co-occurrence matrix G.

[0130] The product co-occurrence module 2 of this embodiment corresponds to steps S201 to S203 of the recommendation method of the first aspect.

[0131] As a preferred embodiment, the recommendation device based on graph neural network is as follows: Figure 7 As shown, it also includes a negative sampling enhancement module 3, which specifically includes:

[0132] A negative sample extraction module 31 is used to extract at least one original negative sample and calculate the embedding vectors of all the extracted original negative samples;

[0133] The pseudo negative sample synthesis module 32 is used to extract the embedding vector of at least one positive sample, mix it with the embedding vectors of all original negative samples, and synthesize the embedding vector of the pseudo negative sample. The embedding vector of the pseudo negative sample is used to fit the calculation of the product embedding vector.

[0134] The product co-occurrence module 3 of this embodiment corresponds to steps S301 - S302 of the recommendation method of the first aspect.

[0135] Furthermore, the recommendation device based on graph neural network, such as Figure 7As shown, the negative sampling enhancement module 3 further includes the following on the basis of the previous embodiment:

[0136] The pseudo negative sample selection module 33 is used to perform an inner product operation on the embedding vector of the user u and the embedding vectors of all the pseudo negative samples for a specific user u if there are multiple extracted original negative samples, and select the pseudo negative sample with the largest inner product.

[0137] The product co-occurrence module 3 of this embodiment corresponds to steps S301 to S303 of the recommendation method of the first aspect.

[0138] According to a third aspect, an electronic device includes:

[0139] one or more processors;

[0140] a storage device having one or more programs stored thereon;

[0141] When the one or more programs are executed by the one or more processors, the one or more processors implement any one of the graph neural network-based recommendation methods described in the first aspect.

[0142] In a fourth aspect, a computer-readable medium stores a computer program thereon, wherein when the program is executed by a processor, the recommendation method based on a graph neural network described in any one of the first aspects is implemented.

[0143] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A recommendation method based on graph neural network, characterized in that: The steps include: Based on the interaction data between users and items, a user-item bipartite graph is constructed; Initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph; According to the relationship between the product embedding vector and the user embedding vector in the converged state of the graph neural network, the fitted user embedding vector and the fitted product embedding vector are calculated respectively; Obtaining a loss function according to the fitted user embedding vector and the fitted product embedding vector; According to the loss function, back-propagation is performed to update the user embedding vector and the item embedding vector; The method further comprises the steps of: Constructing a product co-occurrence matrix G according to the user-product bipartite graph; Convert the commodity co-occurrence matrix G into a sparse matrix; Optimize the loss function based on the similarity between products in the product co-occurrence matrix G; The method further comprises the steps of: Extract at least one original negative sample, and calculate the embedding vectors of all the extracted original negative samples; An embedding vector of at least one positive sample is extracted, and is mixed with the embedding vectors of all original negative samples to synthesize an embedding vector of a pseudo negative sample, where the embedding vector of the pseudo negative sample is used to fit the calculation of the product embedding vector.

2. A recommendation method based on graph neural network as claimed in claim 1, characterized in that: Also includes: If there are multiple original negative samples extracted, for a specific user u, an inner product operation is performed between the embedding vector of the user u and the embedding vectors of all the pseudo negative samples, and the pseudo negative sample with the largest inner product is selected; and / or The blending process uses an interpolation operation.

3. A recommendation system based on graph neural network, characterized in that: include: The bipartite graph construction module is used to construct a user-item bipartite graph based on the interaction data between users and items; An initialization module, used to initialize all user embedding vectors and product embedding vectors in the user-product bipartite graph; The fitting embedding vector module is used to calculate the fitting user embedding vector and the fitting product embedding vector respectively according to the relationship between the product embedding vector and the user embedding vector in the converged state of the graph neural network; A first loss module, used for obtaining a loss function according to the fitted user embedding vector and the fitted product embedding vector; An updating module, configured to back-propagate according to the loss function to update the user embedding vector and the item embedding vector; The system also includes: A co-occurrence matrix construction module, used to construct a product co-occurrence matrix G according to the user-product bipartite graph; A sparse module, used to convert the commodity co-occurrence matrix G into a sparse matrix; The second loss module is used to optimize the loss function according to the similarity relationship between commodities in the commodity co-occurrence matrix G; The system also includes: A negative sample extraction module, used to extract at least one original negative sample and calculate the embedding vectors of all the extracted original negative samples; The pseudo negative sample synthesis module is used to extract the embedding vector of at least one positive sample, mix it with the embedding vectors of all original negative samples, and synthesize the embedding vector of the pseudo negative sample. The embedding vector of the pseudo negative sample is used to fit the calculation of the product embedding vector.

4. A recommendation system based on graph neural network as claimed in claim 3, characterized in that: Also includes: The pseudo negative sample selection module is used to perform an inner product operation on the embedding vector of the user u and the embedding vectors of all the pseudo negative samples for a specific user u if there are multiple original negative samples extracted, and select the pseudo negative sample with the largest inner product.

5. An electronic device comprising: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in claim 1 or 2.

6. A computer readable medium having a computer program stored thereon, wherein: When the program is executed by a processor, the method according to claim 1 or 2 is implemented.