Method for predicting a collaborative filtering recommendation model

Abstract

The application provides a prediction method of a graph collaborative filtering recommendation model, filters noise by screening interaction features, improves the accuracy of a recommendation result, and introduces a node-level prototype contrast learning task to mine potential features between users or items, and alleviates the data sparsity problem faced by collaborative filtering. In addition, by rewriting the hypergraph convolution formula, the convolution process only needs to be performed once, avoiding the feature smoothing problem caused by multiple convolutions of the traditional graph collaborative filtering model, so that the efficiency and performance of the recommendation system model are improved.

Description

Technical Field

[0001] This invention relates to the fields of recommender systems and graph convolutional neural networks, and more specifically, to a prediction method for a graph collaborative filtering recommender model based on prototype comparison and feature selection. Background Technology

[0002] With the development of the information age into the big data era, the scale of data on the internet is exploding, leading to information overload becoming an increasingly serious problem in our lives. Recommendation systems are an effective solution to alleviate information overload, making it easier for users to find information, and can also increase traffic and revenue for service providers. Therefore, efficient and accurate personalized recommendation systems have become a hot topic in academia and industry. At the same time, research on graph neural networks (GNNs) is gradually emerging. Since most information in recommendation systems has a graph structure, GNNs have advantages in graph representation learning and have therefore been widely applied in recommendation systems.

[0003] Traditional graph convolution-based recommendation models fail to analyze graph convolution in the frequency domain, treating features of different frequencies in the user-item interaction graph equally. This inevitably introduces noise. In fact, only a small portion of the smoothed or coarse features of neighbors in the user-item interaction information actually contribute to the recommendation; most of the graph information can be considered noise. Furthermore, repeated graph convolution operations only smooth neighbor features, failing to effectively filter noise and making graph convolution inefficient. Additionally, to mitigate the impact of data sparsity, traditional graph convolution-based recommendation models employ contrastive learning methods to improve performance. However, these methods typically use random sampling to construct contrast pairs, ignoring the potential relationships between users (or items) and failing to fully utilize the potential of contrastive learning in recommendation systems.

[0004] In view of this, it is indeed necessary to propose a prediction method for a graph collaborative filtering recommendation model based on prototype comparison and feature selection in order to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a prediction method for a graph collaborative filtering recommendation model, which can effectively improve the efficiency of graph convolution and alleviate the problem of data sparsity.

[0006] To achieve the above objectives, this invention provides a prediction method for a graph collaborative filtering recommendation model, which mainly includes the following steps:

[0007] Step 1, Define the problem: Based on the user-project interaction matrix A∈R m×n Establish a bipartite graph to clarify the input and output of the model;

[0008] Step 2, Initial Embedding Layer: Using embedding vector u k ∈Rd and i k ∈R d To describe a user and an item, where d is the embedding dimension;

[0009] Step 3, Feature Filtering: Use the model's graph convolutional filter to filter interactive features and perform only one convolution layer, which will roll up to nodes that are arbitrarily far away.

[0010] Step 4, Prototype contrastive learning task: Use the embedding of semantically similar cluster centers in the space to build contrast pairs with the target node. The node of the cluster center is the prototype. Use the expectation-maximization algorithm to infer the prototype. At this time, the contrastive learning task of the target node can be completed by comparing with the prototype.

[0011] Step 5, Embedding Propagation Layer: The embedding propagation layer in the model is built on the message passing architecture of GCN and is used to capture collaborative filtering signals and the graph structure of user-item interaction; the model uses hypergraph convolution and rewrites the convolution formula with feature-selected filters.

[0012] Step 6, Result Prediction: After learning through convolution in the embedding propagation layer and the contrast task, the final representation of user u and item i is obtained through the pooling function. A formula for the probability of target user u interacting with item i is established. The final representation of user u interacting with item i is combined as the final interaction prediction. The model is trained using the loss function.

[0013] As a further improvement of the present invention, step 1 specifically includes: the input user-item interaction data is represented by an undirected graph G = (V, E), where the node set V consists of user nodes u ∈ U and item nodes i ∈ I, and the edge set E consists of the interaction relationships between users and items. If there is an interaction between a user and an item, there is an edge between them; otherwise, there is no edge. All the interaction data between users and items is the input of the model, and the prediction of whether the target user will interact with the target item is the output of the model.

[0014] As a further improvement of the present invention, step 2 specifically includes: the user embedding vector and the item embedding vector are represented by embedding matrices U and I, respectively:

[0015] U = {u1, u2, ... u} m}

[0016] I = {i1, i2, ..., i n}

[0017] The model uses one-hot encoding as input as the initial embedding of users and items, so matrices U and I can be regarded as the initial features of users and items, i.e., the input features of the model.

[0018] As a further improvement of the present invention, step 3 specifically includes: dividing the interaction graph G into smooth G... S Figure, Roughness G R Image and noise G N The graphs, composed of smooth, coarse, and noise features respectively, are used in G... S and G R The embeddings on the surface generate the final representation, while those from G... N The embedding is filtered out, which is called a bandpass filter:

[0019]

[0020] Where λ t The feature value represents the corresponding interaction feature. Represents user / item;

[0021] Using hypergraph convolution, the interaction between the user and the project can be represented as follows:

[0022]

[0023]

[0024] Where D u, D i, This is the user-project interaction matrix, where R represents the user-project interaction matrix. T Let represent the transpose of this matrix. Using a filter to replace the diagonal matrix composed of the original eigenvalues ​​in the hypergraph convolution, the resulting embedding expression on the hypergraph is as follows:

[0025]

[0026] Where E represents the embedding matrix, {P,π} represents the first m smoothest or coarsest {eigenvectors, eigenvalues}, and P... T Let P be the transpose of P.

[0027] As a further improvement of the present invention, step 4 specifically includes: applying the K-Means algorithm to cluster the embedded users and projects into K clusters, making the nodes within a cluster as close together as possible while maximizing the distance between clusters; then using the EM algorithm to learn the prototype of each cluster, i.e., maximizing the log-likelihood function.

[0028]

[0029] Among them, e u c represents the embedding of user u. i Let R be the latent prototype of user u, R be the interaction matrix, and Φ be a set of model parameters. First, use Jensen's inequality to obtain its lower bound.

[0030]

[0031] Where Q(c) i |e u To observe e u Latent variable c i The distribution of Q(c) is used to estimate Q(c). i |e u When redirecting the target, e u Maximize the function on the above, and in the maximization step, utilize Rewrite the objective function

[0032]

[0033] Assume that users follow an isotropic Gaussian distribution across all clusters, then

[0034]

[0035] Among them, c j The target node represents the prototype of other clusters besides the target node, and δ represents the variance of the Gaussian distribution. Minimizing this objective function will complete the user's contrastive learning task.

[0036] As a further improvement of the present invention, step 5 specifically includes: representing the embedded representation generated on the smoothed hypergraph as follows:

[0037]

[0038]

[0039] in These are the m1 and n1 smoothest {eigenvectors, eigenvalues} in the user adjacency matrix and item adjacency matrix, respectively. γ(·) outputs the importance weights of different features, E U and E I Embedding matrices representing users and items, respectively. Similarly, embeddings are performed on coarse hypergraphs with the coarsest m² and n² {eigenvectors, eigenvalues} to learn their heterogeneity:

[0040]

[0041]

[0042] in, These are the coarsest m² and n² eigenvectors and eigenvalues ​​in the user adjacency matrix and the item adjacency matrix, respectively.

[0043] As a further improvement of the present invention, step 6 specifically includes: setting m = m1 + m2 and n = n1 + n2 in step 5, and generating the final representation using a pooling function:

[0044]

[0045]

[0046] The pooling function here uses a summation method; where β is the coefficient of the exponential function, and k is the order of the Maclaurin expansion. This represents an adjacency matrix with noise removed. The subscripts U / I represent users / items, and O represents... U and O I As the final representation of the user and the project, the inner product of the two is used. To represent the final predicted probability of user and project interaction, i.e.

[0047]

[0048] We use the Bayesian Personalized Ranking (BPR) loss as the loss function for graph convolution. Formally, this loss function is as follows:

[0049]

[0050] Among them, C={(u,i,j)|(u,i)∈R + ,(u,j)∈R - Let} represent paired interaction data, σ() represent the sigmoid activation function, and j represent items that the user did not interact with. Then, the overall loss function of the model is:

[0051] L = L BPR +L P ,

[0052] Where L P The loss function is used for the contrast learning task.

[0053] The beneficial effects of this invention are as follows: By filtering the co-occurrence matrix features of interaction information, this invention filters noise in explicit interaction features and uses prototypes to construct node-level contrastive learning objectives to enrich the final representations of users and items. Therefore, it can effectively improve the efficiency of graph convolution and alleviate the problem of data sparsity. In addition, by rewriting the hypergraph convolution formula, the convolution process only needs to be done once, avoiding the feature smoothing problem caused by multiple convolutions in traditional graph collaborative filtering models, thus improving the efficiency and performance of the recommendation system model. Attached Figure Description

[0054] Figure 1 A diagram illustrating the user-project interaction.

[0055] Figure 2 A two-part diagram of user-project interaction.

[0056] Figure 3 This is a diagram of the feature filtering module.

[0057] Figure 4 This is a schematic diagram of the prototype.

[0058] Figure 5 This is a diagram of the overall model. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] It should be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0061] Additionally, it should be noted that the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] like Figure 1 The user-item interaction relationship shown can be used to build a recommendation system in real e-commerce scenarios. In traditional collaborative filtering recommendation systems, graph convolution and contrastive learning can be used to mine the potential features of users and products, thereby building a more granular recommendation system.

[0063] like Figure 5 As shown, this invention provides a prediction method for a graph collaborative filtering recommendation model, comprising the following steps:

[0064] Step 1: Define the problem. For example... Figure 2 As shown, the user-project interaction matrix A∈R m×nThe model is constructed as a bipartite graph to clarify the problem under study. The input user and project interaction data is represented by an undirected graph G = (V, E), where the node set V consists of user nodes u ∈ U and project nodes i ∈ I, and the edge set E consists of the interaction relationships between users and projects. If a user interacts with a project, an edge is established between them; otherwise, no edge is established. All user and project interaction data constitute the input of the model, and the prediction of whether the target user will interact with the target project is the output of the model.

[0065] Step 2: Initialize the embedding layer. Use the embedding vector u k ∈R d and i k ∈R d Let's describe a user and an item, where d is the embedding dimension. The user embedding vector and the item embedding vector can be represented by embedding matrices U and I, respectively:

[0066] U = {u1, u2, ... u} m}

[0067] I = {i1, i2, ..., i n}

[0068] The model uses one-hot encoding as input as the initial embeddings of users and items. The embeddings in matrices U and I can serve as the initial features of users and items, i.e., the input features of the model.

[0069] In this embodiment, the node is optimized in subsequent convolution and contrastive learning processes to provide a more effective representation for the recommendation task, because the latent features of the node are injected into the node representation during the optimization process.

[0070] Step 3: Feature Filtering. For example... Figure 3 As shown, the accuracy of graph collaborative filtering recommendations is mainly contributed by smooth and coarse features. However, most features are neither too smooth nor too coarse, making them unsuitable for capturing these two types of features and can be considered noise. In this step, it is necessary to filter the features in the user-item interaction matrix, dividing the interaction graph G into smooth and coarse features. S Roughness G R and noise G N The graphs, composed of smooth, coarse, and noise features respectively, are used in G... S and G R The embeddings on the surface generate the final representation, while those from G... N The embeddings need to be filtered out by a bandpass filter, leaving only the coarsest and smoothest features, i.e.

[0071]

[0072] Where λt The feature value represents the corresponding interaction feature. The importance of an interaction feature depends not only on the feature value but also on the user / item. Therefore, γ(u / i,λ) t The importance weight of the output features to user u.

[0073] In this embodiment, to obtain stronger and more informative embeddings, hypergraph convolution is used, and the interaction between the user and the item can be represented as:

[0074]

[0075]

[0076] Where D u, D i, It is a user-project interaction matrix, where R represents the user-project interaction matrix. T This represents the transpose of the matrix. Since stacking multiple GCNs inevitably leads to oversmoothing, the graph convolutional filter in this model filters interactive features and performs only one convolution layer. However, even a single convolution layer can reach arbitrarily distant nodes. Therefore, a filter is used to replace the diagonal matrix composed of the original feature values ​​in the hypergraph convolution. Furthermore, the hypergraph convolution formula is rewritten to ensure that the convolution process is performed only once, avoiding the feature smoothing problem caused by multiple convolutions in traditional graph collaborative filtering models. This improves both the efficiency and performance of the recommendation system model. The embedding expression generated on the hypergraph is:

[0077]

[0078] Where E represents the embedding matrix, {P,π} represents the first m smoothest or coarsest {eigenvectors, eigenvalues}, and P... T Let P be the transpose matrix. Treating the interaction feature importance function γ(·) as a continuous function of eigenvalues ​​(ignoring the influence of user / project parameters for now), the terms in the parentheses above can be rewritten using a Taylor series:

[0079]

[0080] When m→M in It is the adjacency matrix after removing noise, and K is the highest-order matrix with non-zero derivatives. These are the coefficients of the k-th order Maclaurin expansion. If the model is to capture as many neighbor signals as possible that are effective in generating the final representation, then γ(·) should be infinitely differentiable, meaning that its derivative of any order is positive. Therefore, we choose γ(π) = e^(-π / 2). βπ As the exponential kernel, the convolution formula is rewritten using the graph convolution paradigm of collaborative filtering:

[0081]

[0082] It can then be convolved to its infinitely distant neighbors.

[0083] Step 4: Prototype Comparative Learning Task. Besides direct user and item interactions, there are various potential relationships useful for recommendation tasks. For example, users with similar representations form semantic neighbors. In this step, we consider constructing a node-level comparative learning task. Node-level comparison targets typically require pairwise learning for each node pair, which is time-consuming for a large number of neighbors. Therefore, we consider using the embeddings of semantically similar cluster centers in the space to construct comparison pairs with the target nodes. These cluster center nodes are the prototypes. The node comparative learning task can then be completed by comparing with the prototypes. The prototype represents the center of a cluster composed of nodes with similar semantics; it is a potential node. Therefore, we consider using the Expectation-Maximization (EM) algorithm to infer the prototype.

[0084] Specifically, the K-Means algorithm is first applied to the embedded user and project data to cluster them into K clusters, ensuring that nodes within a cluster are as close together as possible while maximizing the distance between clusters. Then, the EM algorithm is used to learn the prototype of each cluster, which is to maximize the log-likelihood function.

[0085]

[0086] Among them, e u c represents the embedding of user u. i Let R be the latent prototype of user u, R be the interaction matrix, and Φ be a set of model parameters. First, use Jensen's inequality to obtain its lower bound.

[0087]

[0088] Where Q(c) i |e u To observe e u Latent variable c i The distribution of Q(c) is used to estimate Q(c). i |e u When redirecting the target, e u Maximizing the function on, that is, in the step of finding the expectation, since e u If it is fixed, then Q(c) i |e u The embeddings of all users can be estimated using the K-means algorithm. If a user belongs to cluster i, then the cluster center c... i For this user's prototype, that is, only for the prototype c of this cluster i ,have Non-cluster prototypes are 0. In the maximization step, utilizing... Rewrite the objective function

[0089]

[0090] Assume that users follow an isotropic Gaussian distribution across all clusters, then

[0091]

[0092] Among them, c j The target node represents the prototype of other clusters besides the target node, and δ represents the variance of the Gaussian distribution. Minimizing this objective function will complete the user's contrastive learning task.

[0093] like Figure 4 As shown, by finding the prototype using the K-means algorithm and the expectation-maximization algorithm and comparing it with the target node, the loss function for the contrastive learning task can be obtained. Taking the user's prototype contrastive loss function as an example:

[0094]

[0095] In the above formula, e u and c i If we normalize beforehand, then we have (e u -c i ) 2 =2-2e u c i If each Gaussian distribution has the same derivation and is expressed as a temperature hyperparameter τ, then it can be rewritten as...

[0096]

[0097] Where c i Let u be the prototype of user u. Minimizing this loss function minimizes the distance between the target user and its own cluster prototype and maximizes its distance from other cluster prototypes. Similarly, the loss function for the item can be expressed as:

[0098]

[0099] c k For the prototype of project i, the loss function for the complete contrastive learning task is the weighted sum of the above two equations.

[0100] L P =L U +μL I ,

[0101] Where μ is the weight parameter.

[0102] Step 5: Embedding Propagation Layer. The embedding propagation layer in the model is built on the messaging architecture of GCN and is used to capture collaborative filtering signals and the graph structure of user item interactions. The basic idea of ​​GCN is to learn node representations through features on a smooth graph. To propagate embeddings on both smooth and coarse hypergraphs, the embedding representations generated on the smooth hypergraph are as follows:

[0103]

[0104]

[0105] in These are the m1 and n1 smoothest {eigenvectors, eigenvalues} in the user adjacency matrix and item adjacency matrix, respectively. γ(·) outputs the importance weights of different features, E U and E I Embedding matrices representing users and items, respectively. Similarly, embeddings are performed on coarse hypergraphs with the coarsest m² and n² {eigenvectors, eigenvalues} to learn their heterogeneity:

[0106]

[0107]

[0108] in, These are the coarsest m² and n² eigenvectors and eigenvalues ​​in the user adjacency matrix and the item adjacency matrix, respectively.

[0109] In this embodiment, in order to obtain stronger and more informative embeddings, hypergraph convolution is used, and the convolution formula is rewritten with a feature-selected filter. Only one convolution layer is needed to learn the representation of nodes from the selected features, thereby improving the efficiency and performance of graph convolution.

[0110] Step 6: Result Prediction. After learning through convolutional embeddings and the comparison task, the overall framework of the graph collaborative filtering recommendation model based on prototype comparison and feature selection is as follows: Figure 5 As shown, in step 5, let m = m1 + m2 and n = n1 + n2. The final representation is generated through the pooling function. After optimization by hypergraph convolution and contrastive learning tasks, the final representations of user and item nodes are obtained.

[0111]

[0112]

[0113] To avoid introducing additional complexity that could compromise the computational simplicity of the model, the pooling function here uses a summation approach, where β is the coefficient of the exponential function and k is the order of the Maclaurin expansion. This represents the adjacency matrix after noise removal, with the subscripts U / I indicating user / item. Since single-layer feature filtering can reasonably weight the important graph features for recommendation, there is no need to stack more layers. U and O I As the final representation of the user and the project, the inner product of the two is used. The formula used to represent the final predicted probability of user and project interaction is as follows:

[0114]

[0115] The final score not only extracts the collaborative filtering signal between users and items, but also takes into account behavioral semantics.

[0116] Bayesian Personalized Ranking (BPR) loss is used as the loss function for graph convolution. Formally, this loss function is as follows:

[0117]

[0118] Among them, C={(u,i,j)|(u,i)∈R + ,(u,j)∈R - Let} represent paired interaction data, σ() represent the sigmoid activation function, and j represent items that the user did not interact with. Then the overall loss function of the model is...

[0119] L = L BPR +L P ,

[0120] Where L P This is the loss function for the comparative learning task. By optimizing this loss function, the interaction between the user and the item can be modeled.

[0121] In summary, this invention filters noise by selecting interaction features, thereby improving the accuracy of recommendation results. It also introduces a node-level prototype contrastive learning task to mine potential features between users or items, alleviating the data sparsity problem faced by collaborative filtering. Furthermore, by rewriting the hypergraph convolution formula, the convolution process only needs to be performed once, avoiding the feature smoothing problem caused by multiple convolutions in traditional graph collaborative filtering models. This improves both the efficiency and performance of the recommendation system model.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A prediction method for a graph collaborative filtering recommendation model, characterized in that, The main steps include: Step 1, Define the problem: Based on the user-project interaction matrix A∈R m×n Establish a bipartite graph to clarify the input and output of the model; Step 2, Initial Embedding Layer: Using embedding vector u k ∈R d and i k ∈R d To describe a user and an item, where d is the embedding dimension; Step 3, Feature Filtering: Use the model's graph convolutional filter to filter interactive features and perform only one convolution layer, which will roll up to nodes that are arbitrarily far away. Step 4, Prototype contrastive learning task: Use the embedding of semantically similar cluster centers in the space to build contrast pairs with the target node. The node of the cluster center is the prototype. Use the expectation-maximization algorithm to infer the prototype. At this time, the contrastive learning task of the target node can be completed by comparing with the prototype. Step 5, Embedding Propagation Layer: The embedding propagation layer in the model is built on the message passing architecture of GCN and is used to capture collaborative filtering signals and the graph structure of user-item interaction; the model uses hypergraph convolution and rewrites the convolution formula with feature-selected filters. Step 6, Result Prediction: After learning through convolution in the embedding propagation layer and the contrast task, the final representation of user u and item i is obtained through the pooling function. A formula for the probability of target user u interacting with item i is established. The final representation of user u interacting with item i is combined as the final interaction prediction. The model is trained using the loss function. Step 3 specifically includes: Divide the interaction graph G into smooth G. S Figure, roughness G R Image and noise G N The graphs, composed of smooth, coarse, and noise features respectively, are used in G... S and G R The embeddings on the surface generate the final representation, while those from G... N The embedding is filtered out, which is called a bandpass filter: Where λ t The feature value represents the corresponding interaction feature. Represents user / item; Using hypergraph convolution, the interaction between the user and the project can be represented as follows: Among them, D u, D i This is the user-project interaction matrix, where R represents the user-project interaction matrix. T Let represent the transpose of this matrix. Using a filter to replace the diagonal matrix composed of the original eigenvalues ​​in the hypergraph convolution, the resulting embedding expression on the hypergraph is as follows: Where E represents the embedding matrix, {P,π} represents the first m smoothest or coarsest {eigenvectors, eigenvalues}, and P... T Let P be the transpose matrix; Step 5 specifically includes: The embeddings generated on the smoothed hypergraph are represented as follows: in, These are the m1 and n1 smoothest {eigenvectors, eigenvalues} in the user adjacency matrix and item adjacency matrix, respectively. γ(·) outputs the importance weights of different features, E U and E I Embedding matrices representing users and items, respectively. Embedding is performed on a coarse hypergraph with m² and n² {eigenvectors, eigenvalues} at the coarsest level to learn its heterogeneity: in, These are the coarsest m² and n² {eigenvectors, eigenvalues} in the user adjacency matrix and the item adjacency matrix, respectively.

2. The prediction method of the graph collaborative filtering recommendation model according to claim 1, characterized in that, Step 1 specifically includes: The input user and project interaction data is represented by an undirected graph G = (V, E), where the node set V consists of user nodes u ∈ U and project nodes i ∈ I, and the edge set E consists of the interaction relationships between users and projects. If there is an interaction between a user and a project, there is an edge between them; otherwise, there is no edge. All user and project interaction data is the input of the model, and the prediction of whether the target user will interact with the target project is the output of the model.

3. The prediction method of the graph collaborative filtering recommendation model according to claim 2, characterized in that, Step 2 specifically includes: User embedding vectors and item embedding vectors are represented by embedding matrices U and I, respectively: U={u1,u2,…u m } I={i1,i2,…,i n } The model uses one-hot encoding as input as the initial embedding of users and items, so matrices U and I can be regarded as the initial features of users and items, i.e., the input features of the model.

4. The prediction method of the graph collaborative filtering recommendation model according to claim 1, characterized in that, Step 4 specifically includes: The K-Means algorithm is used to cluster users and projects into K clusters, making the nodes within each cluster as close together as possible while maximizing the distance between clusters. Then, the EM algorithm is used to learn the prototype of each cluster, which is to maximize the log-likelihood function. Among them, e u c represents the embedding of user u. i Let R be the latent prototype of user u, R be the interaction matrix, and Φ be a set of model parameters. First, use Jensen's inequality to obtain its lower bound. Where Q(c) i |e u To observe e u Latent variable c i The distribution of Q(c) is used to estimate Q(c). i |e u When redirecting the target, e u Maximize the function on the above, and in the maximization step, utilize Rewrite the objective function Assume that users follow an isotropic Gaussian distribution across all clusters, then Among them, c j The target node represents the prototype of other clusters besides the target node, and δ represents the variance of the Gaussian distribution. Minimizing this objective function will complete the user's contrastive learning task.

5. The prediction method of the graph collaborative filtering recommendation model according to claim 1, characterized in that, Step 6 specifically includes: Let m = m1 + m2 and n = n1 + n2 in step 5, and generate the final representation using the pooling function: The pooling function here uses a summation method, where β is the coefficient of the exponential function and k is the order of the Maclaurin expansion. This represents an adjacency matrix with noise removed. The subscripts U / I represent users / items, and O represents... U and O I As the final representation of the user and the project, the inner product of the two is used. To represent the final predicted probability of user and project interaction, i.e. We use the Bayesian Personalized Ranking (BPR) loss as the loss function for graph convolution. Formally, this loss function is as follows: Among them, C={(u,i,j)|(u,i)∈R + ,(u,j)∈R - Let} represent paired interaction data, σ() represent the sigmoid activation function, and j represent items that the user did not interact with. Then, the overall loss function of the model is: L=L BPR +L P , Where L P The loss function is used for the contrast learning task.

Images