Hard negative sampling comparison recommendation method based on graph generation diffusion model

Abstract

The invention proposes a hard negative sampling comparison recommendation method based on a graph generation diffusion model, and the method comprises the following steps: coding a user-article interaction graph through a graph convolutional neural network, obtaining a node embedded vector, and carrying out the inner product of the node embedded vector, and obtaining a recommendation probability; the model parameters of the graph convolutional neural network are obtained through the following training steps: S1, generating a comparison view 1 and a comparison view 2 from the node embedding vector through an inline diffusion model view generator and a parameter topological noise reduction view generator; the inline diffusion model view generator comprises an inline diffusion model, and the parameter topological noise reduction view generator comprises a parameter topological noise reduction network; and S2, carrying out comparative learning on the comparative view 1 and the comparative view 2. According to the method, a signal alignment strategy is combined, a double-view generator comparison learning framework is built, and the problems that the model generation capacity is insufficient and the influence of data noise is large are solved, so that the performance of a recommendation system is improved.

Description

Technical Field

[0001] The present invention relates to the field of personalized recommendation technology, and in particular to a hard negative sampling comparison recommendation method based on a graph generation diffusion model. Background Art

[0002] Existing recommendation systems are mainly based on user-item historical interaction graphs, which are embedded and represented by graph neural networks (GNNs) and optimized to improve recommendation performance. With the popularity of online services, recommendation systems have become a key tool to solve the problem of information overload. By analyzing user behavior and needs, these systems are able to provide users with personalized product and content recommendations. From the evolution of collaborative filtering to graph neural networks, research on recommendation systems has made continuous progress. For example, models such as PinSage, GCMC, NGCF, and LightGCN have demonstrated excellent personalized recommendation effects by capturing local and global information of user-item interaction graphs. In addition, in order to reduce dependence on data labels, self-supervised learning has been introduced into recommendation systems, such as models such as SGL and SLRec, which use the structure and pattern of the data itself to create supervision signals, thereby improving the accuracy and relevance of recommendation systems.

[0003] Although the existing technology has achieved remarkable results in recommender systems, it still faces some challenges. First, the representation ability of the recommendation method is limited, and it is difficult to fully capture the complex relationship between users and items, which limits the performance of the recommender system. Second, the data noise problem is one of the key issues that the recommender system needs to solve. Due to the randomness of user interaction behavior and the limitation of data collection, there is a large amount of noisy data in the recommender system, which may lead to inaccurate recommendation results. At the same time, the recommender system in real scenarios suffers from data sparsity. This representation mechanism based on the user-item historical interaction graph cannot ensure the accurate prediction of user data using high-quality data. In addition, the reliance of the recommender system on high-quality data is also a limitation. In real scenarios, high-quality data is often difficult to obtain, which further limits the effectiveness of the recommender system in practical applications. Finally, the problem of model collapse during training is also a technical problem that needs to be solved in the current recommender system field. Especially in self-supervised learning, how to determine an effective comparative view generation strategy to reduce the trial and error cost and how to avoid the model collapse during training are important issues in the research of recommender systems. In addition, a single view generator is also prone to the risk of model collapse in self-supervised comparative learning, which further increases the complexity of technical challenges. Summary of the invention

[0004] The present invention aims to at least solve the technical problems that the recommendation system application in real scenarios in the prior art has data sparsity and data noise, which causes poor representation of user-item interaction graphs, and most models have instability and limited generation capabilities. In particular, a hard negative sampling comparison recommendation method based on a graph generation diffusion model is innovatively proposed.

[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides a hard negative sampling comparison recommendation method based on a graph generation diffusion model, comprising the following steps:

[0006] A user-item interaction graph is constructed, where users and items are nodes and their interactions (such as purchases, ratings, etc.) are edges. Then, the user-item interaction graph is encoded through a graph convolutional neural network to obtain user and item node embedding vectors, and then the inner product of the user and item node embedding vectors is performed to obtain the recommendation probability. Finally, the top N items with the highest recommendation probabilities are recommended to the user.

[0007] The model parameters of the graph convolutional neural network are obtained through the following training:

[0008] S1, the node embedding vector is respectively generated by the inline diffusion model view generator and the parameter topology denoising view generator to generate a contrast view Figure 1 and contrast Figure 2 ; The inline diffusion model view generator includes an inline diffusion model, and the parameter topology denoising view generator includes a parameter topology denoising network; by introducing the inline diffusion model (L-Diffusion) and the parameter topology denoising network (PTDNet) as view generators; the limited representation ability and data noise problems of the recommendation method can be improved;

[0009] S2, compare the Figure 1 and contrast Figure 2 Conduct contrastive learning; alleviate the limitation of high-quality data in recommendation systems, and the two contrasting views in contrastive learning come from different view generators, which can solve the problem of model collapse during training.

[0010] Preferably, encoding the user-item interaction graph through a graph convolutional neural network to obtain a node embedding vector includes: improving the feature representation of user and item nodes by integrating multiple representation embedding propagation layers:

[0011] will u i and v j The representation embedding vectors obtained by traversing the graph neural network at layer l are defined as and The target node information is propagated from the l-1 layer to the l layer. In order to obtain a high-quality node embedding representation, the node embeddings on each propagation layer are summed, and then the inner product between the final embedding representations of the user and the item is used to predict the user's preference for the item in personalized recommendation. The specific process of calculating the node embedding vector is shown in formula (5):

[0012]

[0013] in, and u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0014] Indicates the user's preference score for the item;

[0015] They respectively represent the representation embedding vector obtained by passing the aggregate information of the user and the target node adjacent to the sampled anchor through the l-layer graph neural network and the representation embedding vector obtained by passing the aggregate information of the item and the target node adjacent to the sampled anchor through the l-layer graph neural network;

[0016] Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l-1;

[0017] Represent the final node embedding vector of the user and the final node embedding vector of the item respectively;

[0018] Represents the preference score obtained by the inner product of the user and item embedding vectors;

[0019] Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0020] and Represents each user u i and itemsv j The information representation embedding vector of

[0021] L represents the number of layers of the graph neural network;

[0022] The inner product of the node embedding vector is performed to obtain the recommendation probability; expressed as;

[0023]

[0024] in, for The transpose of .

[0025] Preferably, The calculation formula is as follows:

[0026]

[0027] Among them, E (u) 、E (v) They are matrix embedding representations of users and items respectively;

[0028] Respectively represent the aggregate information representation of users and items and the target nodes adjacent to the sampled anchor points;

[0029] Represents the normalized adjacency matrix The i-th row in the normalized adjacency matrix The jth column in ;

[0030] It is the normalized adjacency matrix representation calculated by the adjacency matrix A of the user-item interaction graph. The calculation derivation of is as follows:

[0031]

[0032] in, and Represented as diagonal matrices about users and items respectively;

[0033] Represents the normalized adjacency matrix The element at row i and column j;

[0034] A i,j The element in the i-th row and j-th column of the adjacency matrix A representing the user-item interaction graph;

[0035] L i and L j are the sets of all neighbor nodes of users and items respectively; || is the absolute value symbol. The normalized matrix can better reflect the interaction signal between users and items, and can better transmit and iterate the information aggregated by the embedding layer.

[0036] Preferably, the inline diffusion model view generator generates contrastive views for contrastive learning. Figure 1 The following steps are involved:

[0037] S2-1, embedding representation vector e of user-item interaction l Input VGAE encoder, the process is as follows:

[0038]

[0039] Among them, q φ (x|e) represents the probability distribution obtained by graph variation encoding;

[0040] N represents Gaussian distribution;

[0041] x represents the embedding representation vector e of the user-item interaction l Low-dimensional mapping via a parameterized graph variational encoder;

[0042] μ φ (e) represents the mean of the Gaussian distribution inferred by the parameter set of the neural network;

[0043] σ φ (e) represents the variance of the Gaussian distribution inferred by the parameter set of the neural network;

[0044] e is obtained from formula (5): or

[0045] μ represents the mean of the Gaussian distribution. The input data is embedded in the most likely position in the vector space. σ represents the variance of the Gaussian distribution, and the variances between dimensions are independent, reflecting the discrete situation around the mean data. φ is the parameter set of the VGAE encoder.

[0046] S2-2, low-dimensionally map the output of the VGAE encoder to x and input it into the diffusion model;

[0047] S2-2-1, Forward process: In the forward propagation process, the inline diffusion model iteratively applies K steps of adding noise, where K is a hyperparameter and the embedding vector is represented by x after adding noise for the kth time. k , each data calculation goes through an iterative process as shown in formula (7):

[0048]

[0049] Among them, q(x k |x k-1 ) represents the forward process probability distribution;

[0050] N represents Gaussian distribution;

[0051] x k is the noise representation added to the embedding vector for the kth time;

[0052] β k ∈(0,1) represents the size of the input noise;

[0053] S2-2-2, reverse process: In the back propagation process, the propagation goal is to embed the representation x that has been iteratively added with noise. z Recover in the implicit vector space and reconstruct the original data from the noise, where k = 1, 2, ..., K. Specifically, the back propagation process is predicted using a trainable deep neural network. The propagation calculation process is as shown in formula (8):

[0054] p(x k-1 |x k )=N(x k-1 ;W1(x k ,k),W2(x k ,k)) (8)

[0055] Among them, p(x k-1 |x k ) represents the probability distribution of the reverse process;

[0056] N represents Gaussian distribution;

[0057] x k-1 is the noise representation added to the embedding vector for the k-1th time;

[0058] W1 and W2 represent Gaussian parameters of a trainable neural network for predicting Gaussian distribution;

[0059] S2-3, the low-dimensional embedding vector reconstructed through the forward and reverse propagation processes of the diffusion model is input into another MLP as a decoder, and the mean and standard deviation of the Gaussian distribution parsed during the encoding process are also input (i.e., μ in Formula 6). φ (e), ), decode and generate a new view.

[0060] The widely used VGAE is introduced on the basis of the diffusion model, which reduces the scale of noise added in data noisification, thereby avoiding the complete noisification of user-item interaction information while retaining the semantic relationship between nodes, thereby improving the problem of loss of user personalized information.

[0061] Preferably, the parametric topology denoising view generator generates contrastive views for contrastive learning. Figure 2 It is expressed as:

[0062] A l =A⊙Z l (10)

[0063] Among them, A l It means that after the l-layer graph neural network, the binary matrix Z is used l The subgraph adjacency matrix after transformation;

[0064] A represents the subgraph adjacency matrix;

[0065] Z l Represents a binary matrix used to generate the subgraph after the l-th layer GNN processing;

[0066] ⊙ represents the element-wise product, so A l Only Z is included l The edge marked with 1 in .

[0067] Preferably, Z l Obtained by:

[0068] The sparse neural network l0 regularization method is used to l To reduce the noise edge, the number of non-zero edges in different GNN layers is directly penalized, as shown in formula (11):

[0069]

[0070] in, is an indicator function, and

[0071] l=1 means starting the calculation from the first layer;

[0072] (u,v)∈ε represents the set of user-item interaction edges of the input subgraph;

[0073] L represents the number of layers of the graph neural network;

[0074] ||·||0 represents the l0 normal form;

[0075] Then we choose to use the graph attention mechanism (GAT) as the denoising layer to learn the parameters that control whether to delete edges. Take a node u in the training subgraph as the central node, and the neighbor nodes of this node represent N u , traversing in the lth layer of GNN, node u and its neighboring nodes v∈N u calculate As shown in formula (12):

[0076]

[0077] in, Represents the quality score of the edge between node u and node v calculated under the graph neural network with l layers;

[0078] is a l Parameterized MLP model;

[0079] They represent the feature vector of the user node and the feature vector of the item node obtained by the l-layer graph neural network respectively;

[0080] θ l It is a trainable parameter, which contains the weight matrix and bias vector of each layer of MLP;

[0081] Through a binary Concrete distribution in the local parameterization Get one The parameterized calculation process is as follows:

[0082]

[0083] Among them, ∈~Uniform(0,1) means random sampling from a binary Concrete distribution;

[0084] ∈ represents a sample randomly drawn from the Concrete distribution;

[0085] Uniform(0,1) means Concrete distribution;

[0086] represents the preliminary calculation result of the retention probability of the edge between node u and node v; the probability value approaches 0 or 1;

[0087] Represents the quality score of the edge between node u and node v under a specific task;

[0088] τ∈R + represents the temperature coefficient;

[0089] σ(·) represents the sigmoid activation function;

[0090] Then use the Hard Sigmoid function to Convert to a binary mask The conversion process is as follows:

[0091]

[0092] Among them, max() means taking the maximum value;

[0093] min() means taking the minimum value.

[0094] Preferably, the loss function of contrastive learning is expressed as:

[0095]

[0096] in, represents the contrast loss function of the user;

[0097] Represents the contrast loss function of the item;

[0098] The calculation formula is the same as The calculation formula is:

[0099]

[0100] in, represents the target loss function;

[0101] u i 、u i′ Represent a user in the user set and a negative sample user in the user set respectively;

[0102] U represents the user node set;

[0103] Represents the cosine similarity of the positive sample pair;

[0104] Represents the cosine similarity of negative sample pairs;

[0105] The hyperparameter τ represents the temperature coefficient in the softmax.

[0106] Preferably, the contrastive supervision loss of hard negative sample sampling is used to optimize the loss function training model:

[0107] First, calculate the hard negative sample pair (s,s - ) to obtain the probability distribution of e i :

[0108]

[0109] Among them, q(s,s - ) represents a hard negative sample pair (s,s - ) probability distribution;

[0110] represents the expected value of selecting category c according to the category distribution ρ

[0111] D c (s - |H(s,e i ,τ)) represents the given samples s, e in category c. i , the probability distribution of negative samples calculated by hard sampling strategy under temperature coefficient

[0112] s - represents negative samples;

[0113] e iRepresents the embedding representation of the corresponding sample;

[0114] τ represents the similarity threshold;

[0115] H(s,e i ,τ) represents the set of hard negative samples;

[0116] Then, according to the embedding representation e of the corresponding sample i Compute feature space similarity:

[0117]

[0118] Among them, h(s,s + ) represents the feature space similarity of positive samples;

[0119] e Τ for e i The transpose of

[0120] e is the embedding representation of all samples;

[0121] e + Embedding representation for all positive samples;

[0122] Represents the feature space similarity of a single negative sample;

[0123] for e i The transpose of

[0124] Represents the feature embedding representation of a single negative sample;

[0125] Υ represents the temperature coefficient. When the similarity is greater than the threshold value, that is, h(s,s - )>τ, and the two samples do not belong to the same probability distribution, then the sample is defined as a hard negative sample;

[0126] Finally, the contrast loss function based on hard negative samples is obtained:

[0127]

[0128] in, represents the hard negative sample contrast loss function;

[0129] It represents the mathematical expectation of positive sample pairs and negative sample pairs according to distribution p;

[0130] k represents the number of negative samples;

[0131] Indicates the similarity between the sample and the negative sample;

[0132] Indicates the similarity between the sample and the positive sample.

[0133] Hard negative samples refer to items that are very similar to positive samples but are actually negative samples. By designing hard negative samples, users can be better personalized and the generalization ability of the model can be improved; distinguishing hard negative samples can better deal with the residual noise data in the original data, thereby improving the robustness of the recommendation system.

[0134] Preferably, in contrastive learning, in order to make the two generated contrastive views Figure 1 and contrast Figure 2 During the training process, they can be aligned with each other and adapt to the CF task. The BPR loss function is introduced to constrain the consistency of the two views:

[0135] For the inline diffusion model view generator, the embedding representation of the user-item interaction encoded by the VGAE encoder is used to calculate the BPR as shown in formula (22):

[0136]

[0137] in, represents the total loss function of the inline diffusion model view generator;

[0138] Represents the KL divergence loss value;

[0139] represents the distribution loss value, ensuring that the node embedding distribution of the generated graph is similar to that of the original graph;

[0140] represents the BPR loss function of the view generator part of the inline diffusion model;

[0141] For the parametric topology denoising view generator, the obtained node embedding representation is used to calculate the BPR, and the loss function is reconstructed as formula (23):

[0142]

[0143] in, represents the total loss function of the parametric topology denoising view generator;

[0144] represents the L0 regularization loss function;

[0145] represents the BPR (Bayesian Personalized Ranking) loss function of the parameterized topology denoising view generator part;

[0146] The total loss function of the model is:

[0147]

[0148] in, represents the total loss function of the inline diffusion model view generator;

[0149] represents the total loss function of the parametric topology denoising view generator;

[0150] represents the hard negative sampling loss function;

[0151] represents the BPR loss function;

[0152] λ1 represents the hyperparameter that controls the strength of the self-supervision signal;

[0153] represents the self-supervised loss function;

[0154] represents the regularization term;

[0155] λ2 is a hyperparameter that controls the regularization strength.

[0156] Preferably, Calculated by the following formula:

[0157]

[0158] Among them, O represents all training data;

[0159] u,i,j represent user, item i, item j respectively, i≠j;

[0160] It means that the calculated difference between the positive and negative sample prediction scores is mapped to between 0 and 1 through the Sigmoid function, and then the negative logarithm is taken to obtain the final loss value;

[0161] σ represents the Sigmoid function;

[0162] Respectively represent the user's prediction scores for positive sample items and negative sample items.

[0163] In summary, due to the adoption of the above technical solution, the present invention combines the signal alignment strategy to build a dual-view generator comparative learning framework, aiming to improve the problems of insufficient model generation ability and large impact of data noise, thereby improving the performance of the recommendation system.

[0164] The specific beneficial effects are:

[0165] (1) Construct a dual-view comparison framework to obtain high-quality image enhancement training signals

[0166] In order to improve the similarity and difference of the comparison view group, the present invention introduces the diffusion model into the VGAE encoder and constructs an inline diffusion model. At the same time, the parameter topology denoising model is introduced, which together with the inline diffusion model constitutes a dual-view comparison framework. This framework can more effectively capture the potential information in the user-item interaction graph, generate high-quality comparison views, and thus obtain higher-quality graph enhancement training signals. This helps to improve the representation ability of the recommendation system, enabling it to capture the complex relationship between users and items more comprehensively and accurately.

[0167] (2) Propose a new hard negative sample sampling strategy for image contrast learning to improve the quality of training signals

[0168] In order to further improve the quality of self-supervised learning training signals, this paper proposes a new graph contrastive learning hard negative sample sampling strategy. This strategy is based on an effective sample selection mechanism of label information and can achieve more accurate sample selection. Combining this sampling strategy with the contrastive loss function of graph generative learning training is conducive to learning the global information representation of the sampled subgraphs in the user-item. This not only improves the quality of the training signal, but also enhances the stability and robustness of the performance of the recommendation system when facing noisy data and sparse data.

[0169] (3) Propose a graph contrastive learning recommendation method GGDHSCL based on self-supervised learning

[0170] In order to comprehensively address the above challenges, this paper proposes a graph contrast learning recommendation method GGDHSCL based on self-supervised learning. This method fully utilizes the advantages of the dual-view contrast framework and the new hard negative sample sampling strategy, and effectively improves the performance of the recommendation system when facing noisy data and sparse data through self-supervised learning. At the same time, the GGDHSCL method also enhances the effectiveness of the collaborative filtering paradigm, enabling the recommendation system to more accurately predict user data and improve the accuracy and relevance of recommendations.

[0171] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0172] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0173] Figure 1 It is a schematic diagram of the overall framework of the GGDHSCL recommended method of the present invention.

[0174] Figure 2This is the performance of the LastFM and Yelp datasets under different sparsity levels of the embodiment of the present invention.

[0175] Figure 3 This is the effect of noise disturbance at different ratios on the performance of the recommended method according to the embodiment of the present invention.

[0176] Figure 4 This is the influence of noise scale and diffusion step length in the diffusion model of Yelp data in an embodiment of the present invention.

[0177] Figure 5 This is a hyperparameter analysis of an embodiment of the present invention on LastFM and Yelp. DETAILED DESCRIPTION

[0178] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0179] This paper proposes a hard negative sampling comparative recommendation method based on a graph generation diffusion model. The user-item interaction graph is encoded through a graph convolutional neural network, and then a comparative view is generated through two different view generators to achieve graph enhancement representation based on self-supervisory signals and improve the performance of the recommendation system. The overall system architecture of the framework is as follows Figure 1 As shown. The two view generators, view generator 1 and view generator 2, work in parallel. View generator 1: First, the sampled user-item interaction graph is input, and the embedded representation is obtained through a multi-layer graph convolutional neural network; then the obtained embedded representation vector is input into the VGAE encoder to obtain a low-dimensional mapping, and at the same time, the mean vector and variance vector inferred by the neural network are obtained, which are used to describe the distribution of each node in the latent space. Then a diffusion model is designed in the implicit space: the forward process is to inject Gaussian noise into the low-dimensional embedding vector obtained by mapping through the principle of multiple small amounts until it obeys the Gaussian distribution; the reverse process is to reconstruct the low-dimensional embedding vector obtained by the variance mean and variance obtained by neural network analysis, and then input the reconstructed embedding vector into a two-layer multi-layer perceptron to obtain a reconstructed graph structure. View generator 2: First, a multi-layer graph neural network is used to obtain the embedded representation of each node, and it is input into the denoising network to obtain a new denoised graph structure. After obtaining the views generated by the two view generators, they are encoded through the graph convolutional neural network, and then the two encoded vectors are compared and learned. At the same time, a BPR loss is calculated during the training of the two views to achieve the signal alignment function. In addition, a hard negative sample contrast loss is calculated to enhance the self-supervisory signal.

[0180] (1) Learning local graph collaborative information representation

[0181] In following the general collaborative filtering paradigm, it is necessary to learn the structural relationship between the different sampled local user-item interaction subgraphs and the semantic relevance between each node, and to segment the interaction nodes into subgraphs centered on the current user node, which is also called an anchor point. The selection of anchor points is to randomly sample the user node set. The global information features of the aggregating subgraph structure require traversing all nodes related to the anchor point in the subgraph to collect high-order information features in the user-item interaction graph. Therefore, the interaction relationship between users and items is encoded, and the interaction graph is embedded in a d-dimensional latent vector space. Specifically, each user u is generated i and itemsv j The information representation embedding vector and (d represents the dimension of the representation embedding). The embedding representation matrix and It is defined as the matrix embedding representation of users and items (I and J represent the number of users and items in the subgraph structure, respectively). In order to capture and aggregate more comprehensive subgraph structure information, an information propagation layer for subgraph representation embedding is constructed as shown in formula (3).

[0182]

[0183] Among them, E (u) 、E (v) They are matrix embedding representations of users and items respectively;

[0184] Respectively represent the aggregate information representation of users and items and the target nodes adjacent to the sampled anchor points;

[0185] Represents the normalized adjacency matrix The i-th row in the normalized adjacency matrix The jth column in .

[0186] It is a normalized adjacency matrix representation calculated by the adjacency matrix A of the user-item interaction graph. Specifically, The calculation of is as shown in formula (4).

[0187]

[0188] in, and Represented as diagonal matrices about users and items respectively;

[0189] Represents the normalized adjacency matrix The element at row i and column j;

[0190] A i,j The element in the i-th row and j-th column of the adjacency matrix A representing the user-item interaction graph;

[0191] L i and L j The normalized matrix can better reflect the interaction signal between users and items, and can better transmit and iterate the information aggregated by the embedding layer.

[0192] In order to more effectively aggregate the semantic relationship feature information of the neighbor nodes of the sampled subgraph, the feature representation of user and item nodes is improved by integrating multiple representation embedding propagation layers. i and v j The representation embedding vectors obtained by traversing the graph neural network at layer l are defined as and The target node information is propagated from the l-1 layer to the l layer. In order to obtain a high-quality node embedding representation, the node embeddings on each propagation layer are summed up, and then the inner product between the final embedding representations of the user and the item is used to predict the user's preference for the item in personalized recommendation. The specific process of embedding representation calculation is shown in formula (5).

[0193]

[0194] in, and u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0195] Indicates the user's preference score for the item;

[0196] They respectively represent the representation embedding vector obtained by passing the aggregate information of the user and the target node adjacent to the sampled anchor through the l-layer graph neural network and the representation embedding vector obtained by passing the aggregate information of the item and the target node adjacent to the sampled anchor through the l-layer graph neural network;

[0197] Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l-1;

[0198] Represent the final node embedding vector of the user and the final node embedding vector of the item respectively;

[0199] Represents the preference score obtained by the inner product of the user and item embedding vectors;

[0200] Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l;

[0201] for The transpose of

[0202] and Represents each user u i and itemsv j The information representation embedding vector of

[0203] L represents the number of layers of the graph neural network;

[0204] (2) Dual-view contrast learning framework

[0205] The self-supervised graph contrastive learning recommendation system not only reduces the reliance on labels in real data sets and reduces the cost of data construction, but also improves the sparsity of user behavior data and the resistance to noise data that always appear in real recommendation scenarios. The additional high-quality training signals generated during the model training process improve the problems of data sparsity and insufficient model robustness. In the existing contrastive recommendation system, the contrastive view construction method generates contrastive views by randomly deleting the connecting edges and nodes of the graph or constructing a hypergraph. The dual-view contrastive learning framework is selected to represent the same set of data through different methods or models. This view construction method can improve a problem that comes from the fact that the contrast enhancement is not accurate enough due to the sharing of the distribution of two views generated by the same view generator, and there is a risk of contrast model collapse. Therefore, it is chosen to use two different view generators to generate views, which can better avoid the above problems and risks, and understand and predict the real needs of users and the real preferences of user groups.

[0206] In terms of view generator selection, we chose to use an inline diffusion model and a parameter topology denoising model as the generators of the two comparison views. The graph generation learning model focuses on reconstructing the graph structure by adding Gaussian noise to the graph to generate a new graph structure; the graph denoising network model focuses on removing the noise data of the user-item interaction graph to form a low-noise comparison view.

[0207] (2.1) Inline Diffusion Model View Generator

[0208] In recent years, the contrast view generator based on graph generation model has been very popular. Generative adversarial network (GAN) can be used to generate different contrast views, and variational autoencoder (VAE) can be used to reconstruct graph structure. The universal diffusion model has a strong generation ability in graph generation tasks, but it needs to make the data noisy until it becomes completely random noise, which will lead to the loss of personalized information in the interaction between users and items. At the same time, although only using graph variational autoencoder (VGAE) as a view generator performs well in recommendation system tasks, it will have the problem of limited representation ability in complex user interaction patterns. Therefore, based on the diffusion model, the widely used VGAE is introduced. When the diffusion model is integrated with VGAE, the scale of adding noise in data noisiness is reduced, thereby avoiding the complete noisiness of user-item interaction information while retaining the semantic relationship between nodes, thereby improving the problem of loss of user personalized information.

[0209] The sampled subgraph is encoded through a multi-layer graph convolutional neural network to obtain a graph representation embedding vector, which is then input into a multi-layer perceptron (MLP) to learn the hidden state variables of each node. The hidden state is then used to estimate the mean and variance representation of the embedding vector to form a multivariate Gaussian distribution parameter. Compared with VAE, VGAE introduces Kullback–Leibler divergence (KL divergence) to reduce the risk of model overfitting. Specifically, the embedding representation vector e of the user-item interaction is l The parameterized graph variation encoder is used to map the low-dimensional image to x, and predicts and learns the mean and variance of the Gaussian distribution. The graph variation encoding process is shown in Equation (6).

[0210]

[0211] Among them, q φ (x|e) represents the probability distribution obtained by graph variation encoding;

[0212] N represents Gaussian distribution;

[0213] x represents the embedding representation vector e of the user-item interaction l Low-dimensional mapping via a parameterized graph variational encoder;

[0214] μ φ (e) represents the mean of the Gaussian distribution inferred by the parameter set of the neural network;

[0215] σ φ (e) represents the variance of the Gaussian distribution inferred by the parameter set of the neural network;

[0216] e is obtained from formula (5): or

[0217] μ represents the mean of the Gaussian distribution. The input data is embedded in the most likely position in the vector space. σ represents the variance of the Gaussian distribution. The variances of each dimension are independent, reflecting the discrete situation around the mean data. φ is the parameter set of the VGAE encoder. e is obtained by formula (5): or In order to increase the distance between positive and negative samples in contrastive learning, maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, and achieve the goal of image contrast enhancement, we use the powerful generation ability of the diffusion model to introduce a diffusion model to generate denoised user-item interaction data encoded by VGAE. Based on the low-dimensional vector x mapped by the VGAE encoder, the forward and reverse processes of diffusion are designed in the latent embedding space. In the forward diffusion process, Gaussian noise increments are added to the original embedding vector x, and in order to improve the missing interaction information between user items, the noise scale is reduced. In the opposite case, the reverse process of diffusion uses a trainable deep neural network to eliminate noise, thereby generating a clean vector representation. By expanding the model's ability to increase features, the distinction between different features is made higher.

[0218] Forward process: During the forward propagation process, the inline diffusion model iteratively applies K steps of adding noise, where K is a hyperparameter and the embedding vector after adding noise at the kth time is represented as x k Each data calculation goes through an iterative process as shown in formula (7).

[0219]

[0220] Among them, q(x k |x k-1 ) represents the forward process probability distribution;

[0221] N represents Gaussian distribution;

[0222] x k is the noise representation added to the embedding vector for the kth time;

[0223] β k ∈(0,1) represents the scale of the input noise, and this parameter is independent in all dimensions.

[0224] Reverse process: In the back propagation process, the propagation goal is to embed the representation x that has been iteratively added with noise z Recover in the implicit vector space and reconstruct the original data from the noise, where k = 1, 2, ..., K. Specifically, the back propagation process uses a trainable deep neural network to predict and propagate the calculation process as shown in formula (8).

[0225] p(x k-1 |x k )=N(x k-1 ;W1(x k ,k),W2(x k ,k)) (8)

[0226] Among them, p(x k-1 |x k ) represents the probability distribution of the reverse process;

[0227] N represents Gaussian distribution;

[0228] x k-1 is the noise representation added to the embedding vector for the k-1th time;

[0229] W1 and W2 represent the Gaussian parameters of the trainable neural network used to predict the Gaussian distribution. The low-dimensional embedding vector reconstructed through the forward and reverse propagation processes of the diffusion model is input into another MLP as a decoder. The mean and standard deviation of the Gaussian distribution parsed during the encoding process are also input for decoding to generate a new view. Finally, the loss function is defined as Equation (9).

[0230]

[0231] in, represents the KL divergence loss value between the node embedding distribution and the standard Gaussian distribution;

[0232] represents the binary cross entropy loss function;

[0233] represents the loss function of the inline diffusion model;

[0234] μ represents the mean of the Gaussian distribution;

[0235] σ represents the variance of Gaussian distribution;

[0236] p pos represents the positive sample edge set;

[0237] p neg represents the negative sample edge set;

[0238] N represents the number of samples taken.

[0239]

[0240] (2.2) Parameter Topology Denoising View Generator

[0241] A high-quality contrast view can improve the performance of the contrast model, considering that the information propagation mechanism of graph convolutional neural networks is to enable each node and edge to effectively aggregate the information of local and global relationships between nodes to update its own embedding representation. However, low-quality graphs with noisy edges will seriously affect the quality of graph embedding and thus the performance of the model. Therefore, the goal of the second view generator in the dual-view generation contrast framework is to build a parametric topological view generator, which uses the graph attention mechanism to reduce the noisy edge data in the graph structure to achieve the effect of contrast enhancement.

[0242] In order to improve the quality of node embedding in each layer of GNN, it is proposed that a denoising network be included in GNN to filter the noise edges in the graph and achieve a denoising effect on the graph structure. The main purpose of this model is to achieve the goal of graph denoising by filtering the noise edges of the input model through a parameterized network. Specifically, the target task is to filter out the noise edges in the input subgraph. Therefore, a binary matrix Z is introduced l ∈{0,1} J×J ,and It represents the edge between two nodes u and v. When the element value is 0, it means there is noise and should be removed. After the information is propagated in GNN, the subgraph adjacency matrix is ​​finally obtained as shown in formula (10).

[0243] A l =A⊙Z l (10)

[0244] Among them, A l It means that after the l-layer graph neural network, the binary matrix Z is used l The subgraph adjacency matrix after transformation;

[0245] A represents the subgraph adjacency matrix;

[0246] Z l Represents a binary matrix used to generate the subgraph after the l-th layer GNN processing.

[0247] ⊙ represents the element-wise product, so A l Only Z is included l In order to improve the probability of removing the noise edges with high probability in the subgraph, the sparse neural network l0 regularization method is used to A l The method to minimize the noise edges is to directly penalize the number of non-zero edges in different GNN layers as shown in Equation (11).

[0248]

[0249] in, is an indicator function, and

[0250] l=1 means starting the calculation from the first layer;

[0251] (u,v)∈ε represents the set of user-item interaction edges of the input subgraph;

[0252] L represents the number of layers of the graph neural network;

[0253] ||·||0 represents the l0 normal form.

[0254] This regularization function is used to encourage the generation of sparse matrices at all layers of the parametric topology denoising view generator, and then remove those edges that may be irrelevant or noisy by minimizing this regularization term. Due to the compositionality and non-differentiability of the l0 norm, this regularization term cannot be directly optimized using traditional gradient descent methods. However, each 0-1 element They are all obtained from the parameterized Bernoulli distribution, denoted as: Then the number of edges used can be re-regularized as: Then we introduce the reparameterization method and extract 0-1 elements from the Bernoulli distribution Optimize into parameters and a deterministic function of independent random variables, thereby helping the gradient optimization method to optimize the subgraph more effectively.

[0255] Then we choose to use the graph attention mechanism (GAT) as the denoising layer to learn the parameters that control whether to delete the edge. Take a node u in the training subgraph as the central node, and the neighbor nodes of this node represent N u . Traversing in the l-th GNN layer, node u and its neighbor nodes v∈N u calculate As shown in formula (12).

[0256]

[0257] in, Represents the quality score of the edge between node u and node v calculated under the graph neural network with l layers;

[0258] is a l Parameterized MLP model;

[0259] They respectively represent the feature vector of the user node and the feature vector of the item node obtained through the l-layer graph neural network.

[0260] θ lis a trainable parameter, which contains the weight matrix and bias vector of each layer of MLP. Then, through the binary Concrete distribution and Hard Sigmoid function, according to Dynamically decide whether to keep each edge. Through a binary Concrete distribution in the local parameterized Get one The parameterized calculation process is shown in formula (13).

[0261]

[0262] Among them, ∈~Uniform(0,1) means random sampling from a binary Concrete distribution;

[0263] ∈ represents a sample randomly drawn from the Concrete distribution;

[0264] Uniform(0,1) means Concrete distribution;

[0265] represents the preliminary calculation result of the retention probability of the edge between node u and node v; the probability value approaches 0 or 1;

[0266] Represents the quality score of the edge between node u and node v under a specific task;

[0267] τ∈R + represents the temperature coefficient;

[0268] σ(·) denotes the sigmoid activation function.

[0269] Then use the Hard Sigmoid function to Convert to a binary mask The conversion process is as shown in formula (14).

[0270]

[0271] Calculate After that, the problem of formula (11) can be solved, thus achieving the effect of parameterized network noise reduction. Then according to formula (13), we can get The retention probability of the edge can be derived Then, the expected value of the retention probability of each edge is calculated through the L0 regularized loss. The loss calculation process is as shown in formula (15).

[0272]

[0273] in, Cumulative distribution function The probability that the edge is less than 0 is then mapped to between 0 and 1 using σ(·), thereby obtaining the probability that the edge in the subgraph is retained.

[0274]

[0275] (3) Model optimization

[0276] (3.1) Comparative self-supervised loss optimization

[0277] According to the previous article, we can know that the entire model framework chooses to use a dual-view generator to generate contrast views, and applies two different learnable view generators to generate contrast views. Based on the existing self-supervised collaborative filtering (CF) paradigm, this method allows nodes to automatically distinguish between positive sample pairs and negative sample pairs. Specifically, the model regards the representation of the same node in different views as a positive sample pair (such as the representation of the embedding vector of a user in two different views); and the representation of different nodes is regarded as a negative sample pair (such as the embedding vector of any two different users). The contrast loss objective function is to maximize the consistency of positive sample pairs and minimize the consistency of negative sample pairs. The objective loss function is calculated as shown in formula (16).

[0278]

[0279] in, represents the target loss function;

[0280] u i 、u i′ Represent a user in the user set and a negative sample user in the user set respectively;

[0281] U represents the user node set;

[0282] Represents the cosine similarity of the positive sample pair;

[0283] Represents the cosine similarity of negative sample pairs;

[0284] The hyperparameter τ represents the temperature coefficient in softmax. Calculate the contrast loss of the item In the same way, therefore, the loss function of self-supervised learning can be expressed as

[0285] (3.2) Contrastive Supervision Loss Optimization for Hard Negative Sample Sampling

[0286] The effectiveness of the hard negative sampling strategy (HSCL) in contrastive learning and the practicality of label information in self-supervised learning, therefore, the HSCL strategy is introduced into the self-supervised recommendation framework. In the process of training the model in the contrastive learning framework, in order to improve the learning performance of the self-supervised learning recommendation algorithm, HSCL is used to fully combine label information to select positive and negative samples. Negative samples usually refer to items that users have not interacted with. However, not all negative samples are equally helpful for training models. Hard negative samples refer to items that are very similar to positive samples but are actually negative samples. These samples are more difficult to distinguish and are therefore more critical to the learning of the model during training. H-SCL selects hard negative samples by calculating the similarity between samples. If the similarity between a negative sample and the anchor sample is higher than a certain threshold, but they belong to different categories, then the negative sample is considered to be a hard negative sample. First, a sample seed set is constructed and positive samples are randomly selected. If the label of the sample belongs to the current set, it is selected as a positive sample; otherwise, it is considered a negative sample. A simple random probability sample selection strategy is designed to calculate the distribution probability of positive and negative samples in the sample set. Specifically, the positive sample pair (s,s + ) is from the same category distribution D c The category c is selected independently from the global category distribution ρ, and the probability distribution of positive samples is as shown in formula (17).

[0287]

[0288] Among them, p(s,s + ) represents the positive sample (s,s + ) probability distribution;

[0289] represents the expected value of selecting category c according to the category distribution ρ;

[0290] D c (s) represents the probability distribution of sample s in category c;

[0291] D c (s + ) represents the probability distribution of sample s+ in category c;

[0292] D c represents the probability distribution associated with class c.

[0293] The key point here is that for a given anchor sample s, its positive sample s + should come from the same category to ensure that they are similar. This similarity is achieved by assuming that s and s + All come from the same category distribution D cTo ensure that, category c is randomly selected according to the distribution ρ of each category in the entire data set. Negative sample selection is based on different negative probability distributions. An auxiliary function is designed to calculate the similarity between samples in the feature space, thereby optimizing the final loss function. Similar to unsupervised contrastive learning, negative samples are randomly selected in the input space as shown in formula (18).

[0294]

[0295] Among them, q(s,s - ) represents a hard negative sample pair (s,s - ) probability distribution;

[0296] represents the expected value of selecting category c according to the category distribution ρ

[0297] D c (s - |H(s,e i ,τ)) represents the given samples s, e in category c. i , the probability distribution of negative samples calculated by hard sampling strategy under temperature coefficient

[0298] s - represents negative samples;

[0299] e i Represents the embedding representation of the corresponding sample;

[0300] τ represents the similarity threshold;

[0301] H(s,e i ,τ) represents the set of hard negative samples, which consists of two parts: sample s - and the anchor sample s belong to different categories; sample s - The anchor point samples s in the feature space are very close. The feature space similarity calculation is as shown in formula (19).

[0302]

[0303] Among them, h(s,s + ) represents the feature space similarity of positive samples;

[0304] e Τ for e i The transpose of

[0305] e is the embedding representation of all samples;

[0306] e + Embedding representation for all positive samples;

[0307] Represents the feature space similarity of a single negative sample;

[0308] for e i The transpose of

[0309] e i Represents the embedding representation of the corresponding sample;

[0310] Represents the feature embedding representation of a single negative sample;

[0311] Υ represents the temperature coefficient. When the similarity is greater than the threshold value, that is, h(s,s - )>τ, and the two samples do not belong to the same probability distribution, then the sample is defined as a hard negative sample. When constructing the contrast loss function, e i It is used to calculate the representation of positive samples and negative samples, and then calculate the loss. The loss function is shown in formula (20).

[0312]

[0313] in, represents the hard negative sample contrast loss function;

[0314] It represents the mathematical expectation of positive sample pairs and negative sample pairs according to distribution p;

[0315] k represents the number of negative samples;

[0316] Indicates the similarity between the sample and the negative sample;

[0317] Indicates the similarity between the sample and the positive sample;

[0318] By summing up the similarities of all negative samples, that is, Used to minimize the similarity of all negative samples.

[0319] (3.3) View generation optimization for signal alignment

[0320] Selecting two different generators to generate contrasting views on the contrasting view generator requires designing a specific strategy to align the training signals of the two views to adapt to the CF (collaborative filtering) task and optimize the final effect. In order to achieve this optimization goal, BPR (Bayesian Personalized Ranking Loss) is selected to optimize the alignment of view training signals as shown in formula (21).

[0321]

[0322] Among them, O represents all training data;

[0323] It means that the calculated difference between the positive and negative sample prediction scores is mapped to between 0 and 1 through the Sigmoid function, and then the negative logarithm is taken to obtain the final loss value;

[0324] σ represents the Sigmoid function;

[0325] Respectively represent the user's prediction scores for positive sample items (indicating items that have been interacted with) and negative sample items (indicating items that have not been interacted with);

[0326] In order to optimize the inline diffusion model view generator, the embedding representation of the user-item interaction encoded by the VGAE encoder is used to calculate the BPR as shown in Equation (22).

[0327]

[0328] in, represents the total loss function of the inline diffusion model view generator;

[0329] Represents the KL divergence loss value;

[0330] represents the distribution loss value, ensuring that the node embedding distribution of the generated graph is similar to that of the original graph;

[0331] represents the BPR loss function of the view generator part of the inline diffusion model; It is obtained by formula (21) that in order to distinguish the diffusion model view and the denoising network view, Rather than express.

[0332] represents the regularization term, which is calculated by the square of the Frobenius norm of the model parameter set, and λ2 is a hyperparameter used to control the regularization strength. The training signal of the topological denoising view generator can be used to calculate the BPR by embedding the obtained node, and the loss function is reconstructed as Equation (23).

[0333]

[0334] in, represents the total loss function of the parametric topology denoising view generator;

[0335] represents the L0 regularization loss function;

[0336] represents the BPR (Bayesian Personalized Ranking) loss function of the parameterized topology denoising view generator part; It is obtained by formula (21) that in order to distinguish the diffusion model view and the denoising network view, Rather than express.

[0337] Finally, training is performed by combining multiple objectives to maximize the user-item interaction predictions of all positive samples and minimize the user-item interaction predictions of all negative samples. Finally, model training is completed by calculating loss optimization as shown in formula (24).

[0338]

[0339] in, represents the total loss function of the inline diffusion model view generator;

[0340] represents the total loss function of the parametric topology denoising view generator;

[0341] represents the hard negative sampling loss function;

[0342] represents the BPR loss function;

[0343] λ1 represents the hyperparameter that controls the strength of the self-supervision signal;

[0344] represents the self-supervised loss function;

[0345] represents the regularization term;

[0346] λ2 is a hyperparameter that controls the regularization strength.

[0347] (4) Experimental results analysis

[0348] In order to evaluate the performance and effectiveness of GGDHSCL, the experiment was compared with various mainstream recommendation methods on three different real-world datasets. The experiment was designed to verify the following research questions:

[0349] Question 1: How does the proposed recommendation method perform compared with the current mainstream recommendation methods?

[0350] Question 2: How do the key modules of the proposed recommendation method contribute to the method on different datasets?

[0351] Question 3: How does the proposed recommendation method perform on the problem of data sparsity?

[0352] Question 4: How effective is the proposed method in alleviating the data noise problem?

[0353] Question 5: How do the key hyperparameters affect the performance of the proposed recommendation method?

[0354] (4.1) Dataset evaluation

[0355] To evaluate the performance of GGDHSCL, three public datasets from different real-world web platforms were selected to evaluate the performance of different recommendation tasks for different locations, venues, and products. The datasets include: LastFM, Yelp, and BeerAdvocate. LastFM: This dataset contains labels, musicians, and social network listening information collected from a group of users collected from LastFM. This is mainly a study centered around music recommendation. Yelp: This mainstream dataset is user ratings of commercial places in life collected from the Yelp platform. BeerAdvocate: This dataset comes from beer lovers’ comments on beer on the BeerAdvocate platform. It is oriented towards the field of beer recommendation and diversifies the recommendation scenarios. The statistical results of these datasets are shown in Table 1.

[0356] Table 1: Statistics of the experimental dataset.

[0357] Dataset #Users #Items #Interactions Density LastFM 1,889 15,376 51,987 <![CDATA[2.8e -3 ]]> Yelp 42,712 26,822 182,357 <![CDATA[1.6e -4 ]]> BeerAdvocate 10,456 13,845 1,381,094 <![CDATA[9.5e -3 ]]>

[0358] (4.2) Evaluation indicators

[0359] In order to ensure the fairness of the experimental results, the ratio of 7:2:1 is chosen in the experiment to divide the interactions observed in each dataset into training set, validation set and test set. The sampling full-rank evaluation scheme, for each test user, tests and ranks the positive items in the test set together with all non-interaction items to infer the user's preferences. Two representative indicators, Recall@N and NDCG@N, are used to evaluate the performance of the proposed GGDHSCL recommendation method. By default, the value of N is set to 20 for evaluating Top-N recommendations. The calculation methods of Recall and NDGC are as follows:

[0360] 1) Recall: Recall reflects the proportion of relevant items that the recommendation system can successfully find. R(u) represents the Top-N recommendation list provided to the user based on the user's behavior on the training set; T(u) indicates the set of items that the user actually selects after the system recommends items to the user:

[0361]

[0362] 2) CG (Cumulative Gain): Cumulative gain, which adds up each recommendation result (relevance score) in the recommendation list. But there may be a list with high scores at the bottom and low scores at the top.

[0363]

[0364] 3) DCG (Discounted Cumulative Gain): The lost cumulative gain introduces the position influence factor based on CG, and the recommendation effect of the lower-ranked recommendation results is "discounted".

[0365]

[0366] 4) NDCG (Normalized Discounted Cumulative Gain): Normalized loss cumulative gain, evaluates the recommendation system for the entire test set of user recommendation lists, and normalizes the evaluation scores of recommendation lists for different users. IDCG is the list of the best recommendations returned by the recommendation system for a specific user.

[0367]

[0368] (4.3) Comparison with baseline methods

[0369] Compare the performance of GGDHSCL with 14 other recommender system baseline methods on Recall@20@40 and NDCG@20@40

[0370] a) Traditional collaborative filtering recommendation method

[0371] BiasMF: This is a matrix factorization based method that transforms users and items into vector representations in the latent space while considering their respective biases to reflect user preferences.

[0372] NCF: This is a neural network approach that uses multiple layers of neural networks to encode nonlinear features in user-item interactions.

[0373] AutoRec: This model uses an autoencoder architecture to learn embedding representations by reconstructing observed interaction data.

[0374] b) GNN-based recommendation method

[0375] GCMC: This is a method that uses GCN to introduce GCN for interaction matrix and considers the problem of matrix completion of recommendation system from the perspective of link prediction.

[0376] PinSage: This is a GCN-based recommendation method that generates node embeddings that combine graph structure and node feature information through efficient random walks and graph convolutions, significantly improving the quality of recommendations.

[0377] NGCF: This method explicitly incorporates collaborative signals into the embedding process by propagating embeddings on the user-item bipartite graph, thereby effectively capturing the high-order connectivity in the user-item graph and improving the recommendation effect.

[0378] STGCN: This method combines the advantages of graph convolutional encoders and graph autoencoders, improves the robustness of the model in dealing with sparse data and cold start problems, and enhances its performance in collaborative filtering tasks.

[0379] LightGCN: This recommendation method uses linear propagation and weighted summation to learn the embedding of users and items, which simplifies the design of GCN. It is not only easy to implement and train, but also shows significant performance improvement over advanced models such as NGCF under the same experimental settings.

[0380] GCCF: This is GCN-based collaborative filtering, which eliminates nonlinear activation and introduces a residual network structure to alleviate the over-smoothing problem.

[0381] c) Recommendation method based on self-supervision enhancement

[0382] HCCF: By introducing a hypergraph-enhanced cross-view self-supervised contrastive learning architecture, it captures local and global collaborations simultaneously, effectively solving the over-smoothing problem of deep graph structures and the problem of scarce supervisory signals, and significantly improving the representation quality and robustness of the recommendation system.

[0383] SHT: This method is an innovative self-supervised hypergraph transformer framework that improves the performance of recommendation systems by enhancing user representation and capturing global collaboration relationships, especially performing well in dealing with data sparsity and noise.

[0384] SLRec: A multi-task self-supervised learning framework for large-scale item recommendation that addresses the label sparsity problem by learning latent relations between item features and proposes a new data augmentation method that exploits feature correlations.

[0385] SGL: A graph self-supervised learning framework for user-item bipartite graph recommendation systems. It performs data augmentation by performing node and edge dropout on the input bipartite graph. The augmented graph can be viewed as a subview of the original graph.

[0386] NCL: This method is a neighborhood-enhanced contrastive learning method that significantly improves the performance of graph collaborative filtering recommendation systems by explicitly incorporating potential neighbors into contrast pairs and combining neighbor information in structural and semantic spaces.

[0387] AdaGCL: This is an adaptive graph contrastive learning framework that enhances the collaborative filtering paradigm through two trainable view generators to improve the robustness and generalization ability of recommender systems.

[0388] Table 2: Comparison of Recall@20@40 and NDCG@20@40 performance of 14 baseline methods on LastFM, Yelp and Beer datasets.

[0389]

[0390]

[0391] (4.4) Overall performance comparison (RQ1)

[0392] Table 2 gives the performance evaluation of GGDHSCL on three real datasets (Recall@20, NDCG@20, Recall@40 and NDCG@40) of Yelp, LastFM and Beer with 14 benchmark methods such as AdaGCL and SHT. From the data, the following observations are made:

[0393] ·Evaluation results show that GGDHSCL outperforms all other baseline methods in both top-20 and top-40 settings, and significant performance improvements are observed. The excellent performance of GGDHSCL may be attributed to the effective performance of the graph contrastive learning framework of the method in user-item interaction graph data enhancement. The inline diffusion model is combined with the parameter topology denoising model to construct a dual-view contrastive learning framework, and the view generator is automatically adapted to the data characteristics in combination with the view signal alignment optimization to ensure that the view information generated by the view generator is rich and diverse. In addition, hard negative sample contrast loss optimization is introduced to enhance the self-supervised contrastive learning signal, thereby producing higher quality recommendation results. The above conclusions prove that the graph contrastive learning framework is effective and interpretable in the recommendation system of the collaborative filtering paradigm, and the data enhancement view generation technology in contrastive learning is the core issue. High-quality data enhancement technology can effectively improve the performance of recommendation methods.

[0394] ·From the results, it can be seen that compared with other baseline methods, the self-supervised learning-based method is better than the pure GNN-based recommendation system in the recommendation task of the collaborative filtering paradigm. Therefore, the performance gap of the GNN-based CF model exposes the problems of this type of method: it is difficult to solve the noise problem in the data embedding representation and the model overfitting problem caused by data sparsity. However, the recommendation data in real scenarios often have noise and data sparsity problems. However, the improvement of introducing self-supervised signals is attributed to the task of data enhancement to improve the representation ability of the model. GGDHSCL shows two advantages here. First, an inline diffusion model view generator and a parameter topology denoising model view generator are constructed to form a dual-view comparison framework, which can effectively solve the problem of model collapse in graph comparison learning. Secondly, the two view generators are automatically adapted to create effective and reasonable information through signal alignment optimization loss, thereby avoiding information loss and instability caused by data enhancement techniques such as random edge deletion and random mask nodes.

[0395] (4.5) Ablation experiment (RQ2)

[0396] In order to further verify the performance of the method, ablation experiments are conducted to further explore the effectiveness of each submodule in the proposed GGDHSCL method. Table 3 shows the NDGC and Recall indicators of different modules on the top-20 and top-40 LastFM and Yelp datasets.

[0397] Table 3: Ablation experiment results

[0398]

[0399] (4.5.1) Study on the impact of inline diffusion model The diffusion model is inline embedded in the VGAE encoder as a generative view generator in the dual-view generation comparison framework, which shows the advantage of taking into account the unstable representation ability in generative tasks. In the ablation experiment, the variant w / o-DF adopts the VGAE encoder-decoder that only retains the inline diffusion model. The results show that the proposed GGDHSCL recommendation method has a significant improvement over w / o-DF, indicating that the inline diffusion model has a strong and stable generation ability for generative recommendation tasks, thereby improving the performance of the recommendation system.

[0400] (4.5.2) Impact of hard negative sampling strategy The excellent performance of the proposed recommendation method is partly due to the fact that the supervised contrast signal of the introduced hard negative samples enhances the training signal of self-supervised learning. In the ablation experiment, the variant w / o-HS abandons the hard negative sampling strategy. The results show that the hard negative sampling strategy can be effectively combined with the enhanced self-supervised optimization loss, which is beneficial to enhance the recommendation system based on the GNN CF paradigm. In addition, the variant w / o-DH abandons the diffusion model and hard negative sampling strategy in the inline diffusion model at the same time, retaining the dual-view contrast framework. The results show that the combination of the diffusion model and the hard negative sampling strategy is very effective in combating the noise in the data and enhancing the self-supervised signal to effectively perform recommendations.

[0401] (4.5.3) Influence of parameter topology denoising model In order to further study the reasons for the excellent performance of the proposed recommendation method, the influence of the denoised view in the dual-view generation comparison framework on the overall framework is considered. In the ablation experiment, the variant w / o-VG replaces the denoised view generator with the VGAE view generator in the dual-view framework. The results show that although data augmentation using VGAE still has good performance, it is difficult to solve the problem of data noise. The construction of the denoised view generator can effectively avoid this problem, thereby significantly improving the performance of the recommendation method.

[0402] (4.6) Data sparseness experiment (RQ3)

[0403] The performance of the proposed GGDHSCL recommendation method on data with different levels of sparsity is studied to verify the method's good robustness to data sparsity. Specifically, by reducing the interactive information in the dataset at different proportions (10%, 20%, 30%), the function of evaluating the sparse robustness of the recommendation method is realized. This recommendation method is compared with the other three mainstream methods AdaGCL, SGL and LightGCN on the LastFM and Yelp datasets, respectively. In order to intuitively demonstrate the performance of different recommendation methods at different proportions of sparsity, four line graphs are used to show the impact of sparse data on performance trends. Figure 2 middle.

[0404] Figure 2Middle: The proposed GGDHSCL recommendation method shows excellent performance in different datasets with different degrees of sparsity, and is compared with three mainstream recommendation methods, AdaGCL, SGL and LightGCN, and obtains consistent superiority on two different datasets. It can be observed that with the increase of sparsity, the performance of the recommendation method decreases more slowly. After analysis, the results are attributed to the strong generation ability of the inline diffusion model view generator used in this method, combined with the hard negative sampling strategy, which strengthens the contrastive self-supervision signal, thereby alleviating the negative impact of data sparsity; by reducing the sparsity of the interactive information of the data, the four recommendation methods in the figure are significantly affected. In general, the proposed GGDHSCL recommendation method has good robustness in dealing with user-item interaction relationships with different degrees of sparsity.

[0405] (4.7) Noise impact experiment (RQ4)

[0406] In order to better evaluate the impact of noise on the performance of recommendation methods, the proposed GGDHSCL recommendation method is demonstrated to have good robustness to data noise. In the specific experiment, random false edges are used to replace 5%, 10%, 15%, 20% and 25% of the interaction edges, and then this method is compared with the other three baseline mainstream methods AdaGCL, SGL and LightGCN on the LastFM and Yelp datasets respectively. In order to clearly show the performance of different methods under different proportions of noise, four histograms are used to show the impact of noise data on the performance degradation trend. The performance results of NDCG and Recall indicators of the proposed GGDHSCL recommendation method under different proportions of noise data are shown in Figure 2. Figure 3 shown. Figure 3 Middle: Compared with the other three methods AdaGCL, SGL, and LightGCN, the proposed GGDHSCL method performs best in different proportions of noisy data, and the downward trend is more gentle, with stronger robustness. This observation is attributed to the proposed inline diffusion model as a generative comparison framework, which not only retains strong generation capabilities but also has good ability to handle data noise; the denoising view in the dual-view comparison framework of the proposed method is also a crucial part in addressing the data noise problem. Therefore, both view generators used in this method are able to generate comparison views with less noise and more task-related information. By observing the changing trends of each recommendation method under two different datasets, because the Yelp dataset is more sparse, the impact of noise on all methods is more obvious. In summary, it can be concluded from the experimental results that GGDHSCL is a recommendation method with good robustness.

[0407] (4.8) Hyperparameter sensitivity analysis (RQ5)

[0408] In order to study the impact of different hyperparameters on the performance changes of the recommendation method, Figure 4 and Figure 5 Experiments conducted on the LastFM and Yelp datasets are presented. The GGDHSCL recommendation method has four important parameters, of which the diffusion model has two important parameters: the scale of Gaussian noise injected in the forward propagation of the diffusion model and the number of time steps of the Markov chain of the diffusion model; in addition, there are the weights of the sampling hard negative sample strategy and the weight of the InfoNCE loss. Hyperparameter experiments were conducted on the Yelp dataset. By adjusting the scale of Gaussian noise injected by the diffusion model during the forward propagation process to the range of (0.0001, 0.0005, 0.001, 0.005, 0.01, 0.1), and controlling the diffusion step size to 5, a line graph was drawn as shown below. Figure 4 (a); Then, by controlling the injection noise scale to 0.005 and the diffusion step length to the interval of (5, 10, 20, 50, 100), a histogram is drawn as Figure 4 (b).

[0409] Figure 4 Middle: The results show the performance on the Yelp dataset. When the diffusion step size is 5 and the injected Gaussian noise scale is 0.005, the proposed GGDHSCL recommendation method achieves the best performance in terms of NDGC@40 and Recall@40. When the injected Gaussian noise scale is less than 0.005, the performance of the recommendation method increases with the increase of the noise scale. This is attributed to the fact that injecting a higher scale of Gaussian noise is conducive to destroying the internal structure, so that the structure generated by the view generator is more conducive to widening the distance between negative sample pairs, enhancing the effect of contrastive learning, and thus improving the recommendation performance. However, according to the optimal noise level principle in diffusion theory, there is an optimal noise level in the propagation process. Exceeding this level will introduce too much unnecessary complexity and cause performance degradation. This view just explains the reason for the downward trend after the noise scale exceeds 0.005. As the diffusion step size increases, the performance of the proposed GGDHSCL recommendation method also decreases. This is attributed to the fact that too many steps will cause the diffusion model to experience too much noise injection, making it difficult to restore high-quality graph structures during the back-propagation process.

[0410] In order to explore the influence of the weight of the sampling hard negative sample strategy and the weight change trend of the InfoNCE loss on the performance of the proposed GGDHSCL recommendation method, the recommendation method was evaluated on the NDGC@40 and Recall@40 indicators using the LastFM and Yelp datasets. Specifically, the H-SCL weight is searched in the interval of (0.2, 0.4, 0.6, 0.8, 1.0) and the InfoNCE weight is searched in the interval of (1, 0.1, 0.01, 0.001, 0.0001), and the performance change trend of the proposed recommendation method is as follows: Figure 5 shown.

[0411] Figure 5 Middle: The results show the impact of changing the H-SCL loss weight and InfoNCE loss weight on the recommendation method for the LastFM and Yelp datasets. Figure 5 The results in (a) and 5(b) show that the NDCG@40 and Recall@40 indicators perform best when the H-SCL weight is 1 on the LastFM and Yelp datasets. Figure 5 (c) and 5(d), it can be seen that the NDCG@40 and Recall@40 indicators have the best performance on the LastFM and Yelp datasets when the InfoNCE weight is also 1.

[0412] In summary, this paper proposes an innovative graph-based hard negative sampling contrastive recommendation method (GGDHSCL), which aims to address the challenges of existing recommender systems in terms of representation ability, data noise, and data sparsity. To achieve this goal, we introduce a dual-view contrast framework, which includes two key components: an inline diffusion model and a parametric topological denoising model. The inline diffusion model, as a generative view generator, can capture the latent information in the user-item interaction graph and generate high-quality contrast views. The parametric topological denoising model, as a denoised view generator, further refines the generated contrast views and reduces the impact of data noise. To further improve the effect of generative contrast enhancement and reduce the risk of model collapse, we introduce a signal alignment loss strategy that combines task-aware information to improve the quality of the minimum generation. This strategy makes the generated contrast views more accurate and stable, which helps to improve the performance of the recommender system. In addition, we combine the H-SCL loss function to more effectively allow the user-item interaction graph to propagate information by enhancing the self-supervisory signal. This improvement enables the graph-based self-supervised recommender system to more accurately capture the relationship between users and items, thereby improving the accuracy and relevance of recommendations. In order to verify the rationality and effectiveness of the GGDHSCL method, we conducted extensive experiments on three real-world datasets such as LastFM. The experimental results show that compared with 14 popular recommendation algorithms, GGDHSCL performs well in both Recall@20@40 and NDCG@20@40 indicators. In particular, on the BeerAdvocate dataset, GGDHSCL's NDCG@40 is improved by 3.6% and Recall@40 is improved by 3.3%. This result fully demonstrates the effectiveness of GGDHSCL in solving noisy data and sparse data. We also conducted four supplementary experiments to further verify the contribution of each module to the performance of the recommendation method, the robustness perturbation experiment of noise and data sparsity, and the hyperparameter sensitivity analysis experiment. These experimental results further prove the rationality and effectiveness of the GGDHSCL method, and provide strong support for the recommendation method we proposed.

[0413] In summary, the GGDHSCL method effectively solves the challenges faced by existing recommendation systems and improves the performance of recommendation systems by introducing a dual-view comparison framework, a signal alignment loss strategy, and an H-SCL loss function. Experimental results show that the GGDHSCL method performs well on multiple datasets and has made significant improvements in solving noisy and sparse data. Therefore, we can believe that the GGDHSCL method is a reasonable and effective recommendation method.

[0414] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A hard negative sampling comparison recommendation method based on a graph generation diffusion model, characterized in that: The following steps are involved: A user-item interaction graph is constructed, where users and items are nodes and their interactions are edges. Then, the user-item interaction graph is encoded through a graph convolutional neural network to obtain user and item node embedding vectors, and then the inner product of the user and item node embedding vectors is performed to obtain the recommendation probability. Finally, the top N items with the highest recommendation probabilities are recommended to the user. The model parameters of the graph convolutional neural network are obtained through the following training: S1, generating a comparison view 1 and a comparison view 2 by respectively passing the node embedding vector through an inline diffusion model view generator and a parameter topology denoising view generator; the inline diffusion model view generator includes an inline diffusion model, and the parameter topology denoising view generator includes a parameter topology denoising network; S2, compare and learn the comparison view 1 and the comparison view 2.

2. According to claim 1, a hard negative sampling comparison recommendation method based on a graph generation diffusion model is characterized in that: The user-item interaction graph is encoded through a graph convolutional neural network to obtain a node embedding vector, including: improving the feature representation of user and item nodes by integrating multiple representation embedding propagation layers: will u i and v j The representation embedding vectors obtained by traversing the graph neural network at layer l are defined as and The target node information is propagated from the l-1 layer to the l layer. In order to obtain a high-quality node embedding representation, the node embeddings on each propagation layer are summed, and then the inner product between the final embedding representations of the user and the item is used to predict the user's preference for the item in personalized recommendation. The specific process of calculating the node embedding vector is shown in formula (5): in, and u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l; They respectively represent the representation embedding vector obtained by passing the aggregate information of the user and the target node adjacent to the sampled anchor through the l-layer graph neural network and the representation embedding vector obtained by passing the aggregate information of the item and the target node adjacent to the sampled anchor through the l-layer graph neural network; Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l-1; Represent the final node embedding vector of the user and the final node embedding vector of the item respectively; Represents the preference score obtained by the inner product of the user and item embedding vectors; Respectively represent u i and v j The representation embedding vector obtained by traversing the graph neural network at layer l; and Represents each user u i and itemsv j The information representation embedding vector of L represents the number of layers of the graph neural network; The inner product of the node embedding vector is performed to obtain the recommendation probability; it is expressed as: in, for The transpose of Represents the user's preference score for an item.

3. According to claim 2, a hard negative sampling comparison recommendation method based on a graph generation diffusion model is characterized in that: The calculation formula is as follows: Among them, E (u) 、E (v) They are matrix embedding representations of users and items respectively; Respectively represent the aggregate information representation of users and items and the target nodes adjacent to the sampled anchor points; Represents the normalized adjacency matrix The i-th row in the normalized adjacency matrix The jth column in ; It is the normalized adjacency matrix representation calculated by the adjacency matrix A of the user-item interaction graph. The calculation derivation of is as follows: in, and Represented as diagonal matrices about users and items respectively; Represents the normalized adjacency matrix The element at row i and column j; A i,j The element in the i-th row and j-th column of the adjacency matrix A representing the user-item interaction graph; L i and L j are the sets of all neighbor nodes for users and items respectively; || is the absolute value symbol.

4. According to the hard negative sampling comparison recommendation method based on the graph generation diffusion model of claim 1, it is characterized in that: Inline Diffusion Model View Generator Generating contrastive views for contrastive learning 1 includes the following steps: S2-1, embedding representation vector e of user-item interaction l Input VGAE encoder, the process is as follows: Among them, q φ (x|e) represents the probability distribution obtained by graph variation encoding; N represents Gaussian distribution; x represents the embedding representation vector e of the user-item interaction l Low-dimensional mapping via a parameterized graph variational encoder; μ φ (e) represents the mean of the Gaussian distribution inferred by the parameter set of the neural network; σ φ (e) represents the variance of the Gaussian distribution inferred by the parameter set of the neural network; e is obtained from formula (5): or μ represents the mean of the Gaussian distribution. The input data is embedded in the most likely position in the vector space. σ represents the variance of the Gaussian distribution, and the variances between dimensions are independent, reflecting the discrete situation around the mean data. φ is the parameter set of the VGAE encoder. S2-2, low-dimensionally map the output of the VGAE encoder to x and input it into the diffusion model; S2-2-1, Forward process: In the forward propagation process, the inline diffusion model iteratively applies K steps of adding noise, where K is a hyperparameter and the embedding vector is represented by x after adding noise for the kth time. k , each data calculation goes through an iterative process as shown in formula (7): Among them, q(x k |x k-1 ) represents the forward process probability distribution; N represents Gaussian distribution; x k is the noise representation added to the embedding vector for the kth time; β k ∈(0,1) represents the size of the input noise; S2-2-2, reverse process: In the back propagation process, the propagation goal is to embed the representation x that has been iteratively added with noise. z Recover in the implicit vector space and reconstruct the original data from the noise, where k = 1, 2, ..., K, and the propagation calculation process is as shown in formula (8): p(x k-1 |x k )=N(x k-1 ;W1(x k ,k),W2(x k ,k)) (8) Among them, p(x k-1 |x k ) represents the probability distribution of the reverse process; N represents Gaussian distribution; x k-1 is the noise representation added to the embedding vector for the k-1th time; W1 and W2 represent Gaussian parameters of a trainable neural network for predicting Gaussian distribution; S2-3, the low-dimensional embedding vector reconstructed through the forward and reverse propagation processes of the diffusion model is input into another MLP serving as a decoder. The mean and standard deviation of the Gaussian distribution parsed during the encoding process are also input for decoding to generate a new view.

5. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 1, characterized in that: The parametric topology denoising view generator generates contrastive views 2 for contrastive learning as: A l =A⊙Z l (10) Among them, A l It means that after the l-layer graph neural network, the binary matrix Z is used l The subgraph adjacency matrix after transformation; A represents the subgraph adjacency matrix; Z l Represents a binary matrix used to generate the subgraph after the l-th layer GNN processing; ⊙ represents the element-wise product.

6. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 5, characterized in that: Z l Obtained by: The sparse neural network l0 regularization method is used to l To reduce the noise edge, the number of non-zero edges in different GNN layers is directly penalized, as shown in formula (11): in, is an indicator function, and l=1 means starting the calculation from the first layer; (u,v)∈ε represents the set of user-item interaction edges of the input subgraph; L represents the number of layers of the graph neural network; ||·||0 represents the l0 normal form; Then we choose to use the graph attention mechanism as the denoising layer to learn the parameters that control whether to delete edges. Take a node u in the training subgraph as the central node, and the neighbor nodes of this node represent N u , traversing in the lth layer of GNN, node u and its neighboring nodes v∈N u calculate As shown in formula (12): in, Represents the quality score of the edge between node u and node v calculated under the graph neural network with l layers; is a l Parameterized MLP model; They represent the feature vector of the user node and the feature vector of the item node obtained by the l-layer graph neural network respectively; θ l It is a trainable parameter, which contains the weight matrix and bias vector of each layer of MLP; Through a binary Concrete distribution in the local parameterization Get one The parameterized calculation process is as follows: Among them, ∈~Uniform(0,1) means random sampling from a binary Concrete distribution; Represents a sample randomly drawn from a Concrete distribution; Uniform(0,1) means Concrete distribution; represents the preliminary calculation result of the retention probability of the edge between node u and node v; the probability value approaches 0 or 1; Represents the quality score of the edge between node u and node v under a specific task; τ∈R + represents the temperature coefficient; σ(·) represents the sigmoid activation function; Then use the Hard Sigmoid function to Convert to a binary mask The conversion process is as follows: Among them, max() means taking the maximum value; min() means taking the minimum value.

7. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 1 is characterized in that: The loss function of contrastive learning is expressed as: in, represents the contrast loss function of the user; Represents the contrast loss function of the item; The calculation formula is the same as The calculation formula is: Among them, u i 、u i′ Represent a user in the user set and a negative sample user in the user set respectively; U represents the user node set; Represents the cosine similarity of the positive sample pair; Represents the cosine similarity of negative sample pairs; The hyperparameter τ represents the temperature coefficient in the softmax.

8. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 1, characterized in that: The model is trained using the contrastive supervision loss optimization loss function with hard negative sample sampling: First, calculate the hard negative sample pair (s,s - ) to obtain the probability distribution of e i : Among them, q(s,s - ) represents a hard negative sample pair (s,s - ) probability distribution; represents the expected value of selecting category c according to the category distribution ρ D c (s - |H(s,e i ,τ)) represents the given samples s, e in category c. i , the probability distribution of negative samples calculated by hard sampling strategy under temperature coefficient s - represents negative samples; e i Represents the embedding representation of the corresponding sample; τ represents the similarity threshold; H(s,e i ,τ) represents the set of hard negative samples; Then, according to the embedding representation e of the corresponding sample i Compute feature space similarity: Among them, h(s,s + ) represents the feature space similarity of positive samples; e Τ for e i The transpose of e is the embedding representation of all samples; e + Embedding representation for all positive samples; Represents the feature space similarity of a single negative sample; for e i The transpose of Represents the feature embedding representation of a single negative sample; Υ represents the temperature coefficient. When the similarity is greater than the threshold value, that is, h(s,s - )>τ, and the two samples do not belong to the same probability distribution, then the sample is defined as a hard negative sample; Finally, the contrast loss function based on hard negative samples is obtained: in, represents the hard negative sample contrast loss function; It represents the mathematical expectation of positive sample pairs and negative sample pairs according to distribution p; k represents the number of negative samples; Indicates the similarity between the sample and the negative sample; Indicates the similarity between the sample and the positive sample.

9. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 1, characterized in that: In contrastive learning, in order to make the two generated contrast views 1 and 2 aligned with each other during training and adapt to the CF task, the BPR loss function is introduced to constrain the consistency of the two views: For the inline diffusion model view generator, the embedding representation of the user-item interaction encoded by the VGAE encoder is used to calculate the BPR as shown in formula (22): in, represents the total loss function of the inline diffusion model view generator; Represents the KL divergence loss value; represents the distribution loss value, ensuring that the node embedding distribution of the generated graph is similar to that of the original graph; represents the BPR loss function of the view generator part of the inline diffusion model; For the parametric topology denoising view generator, the obtained node embedding representation is used to calculate the BPR, and the loss function is reconstructed as formula (23): in, represents the total loss function of the parametric topology denoising view generator; represents the L0 regularization loss function; represents the BPR loss function of the parameterized topology denoising view generator part; The total loss function of the model is: in, represents the total loss function of the inline diffusion model view generator; represents the total loss function of the parametric topology denoising view generator; represents the hard negative sampling loss function; represents the BPR loss function; λ1 represents the hyperparameter that controls the strength of the self-supervision signal; represents the self-supervised loss function; represents the regularization term; λ2 is a hyperparameter that controls the regularization strength.

10. The hard negative sampling comparison recommendation method based on the graph generation diffusion model according to claim 9, characterized in that: Calculated by the following formula: Among them, O represents all training data; u,i,j represent user, item i, and item j respectively, i≠j; It means that the calculated difference between the positive and negative sample prediction scores is mapped to between 0 and 1 through the Sigmoid function, and then the negative logarithm is taken to obtain the final loss value; σ represents the Sigmoid function; Respectively represent the user's prediction scores for positive sample items and negative sample items.

Images