Guide diffusion model enhancement-based graph encoder recommendation method

Abstract

The invention discloses a graph encoder recommendation method based on guide diffusion model enhancement, and the method comprises the specific steps: S1, initializing a user-article interaction adjacency matrix, carrying out the normalization processing, and generating a Laplacian matrix; s2, adaptively calculating a mask probability for each click behavior, generating a partially visible mask bipartite graph, taking the mask bipartite graph as a mask graph encoder input in a model training process, and enhancing a supervision signal from a structure level; and S3, generating representations of the user and the article by using a mask graph encoder, performing de-noising processing by adding directed noise and introducing a diffusion model guided by a cooperative signal, and enhancing a supervision signal from a semantic level. And S4, generating user and article representations by using a graph encoder, optimizing semantic information of the representations, and finally predicting articles which may be clicked by each user. According to the method, by combining a structure and semantics dual enhancement strategy, the accuracy and recall rate of a recommendation system are remarkably improved, and meanwhile, the robustness of the system to noise information is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation algorithms and self-supervised learning, and in particular to a recommendation method based on a graph encoder enhanced by a guided diffusion model. Background Art

[0002] In the context of the rapid development of digitalization and information technology, the application of recommendation systems in online platforms has become increasingly common and has become an important tool for users to obtain information and provide personalized services. From the perspective of graph structure, the recommendation system can be modeled as a bipartite graph containing nodes and edges, where nodes represent users and items (such as goods, articles, movies, etc.), and edges reflect the interactive relationship between users and items, such as clicks, purchases, or ratings. This graph structure not only reveals the potential patterns of user interests, but also provides rich contextual information for recommendation algorithms. The recommendation system based on graph structure can effectively mine the relationships and propagation paths in the network, thereby improving the accuracy of recommendations while providing strong support for diversified network services and personalized applications.

[0003] In practical applications, recommendation systems need to process large-scale interaction data between users and items, such as Amazon's product recommendations and Douyin video recommendations. These data are often sparse and noisy. For example, most users only browse a small number of products or like a small number of content, and many clicks may not reflect real interests. To solve these problems, contrastive learning was introduced to help the model learn the implicit relationship between users and items by constructing contrast views (such as node subgraphs, randomly perturbed graph structures) and maximizing the mutual information between views. However, the view construction of contrastive learning requires careful design, and the slightest carelessness may lead to view distortion; at the same time, the computational cost of negative sample sampling is high, which limits its application in large-scale scenarios. Generative self-supervised learning shows significant advantages by directly reconstructing the original graph data. For example, in e-commerce, generative methods can predict the feature distribution of products that users have not interacted with, thereby better capturing user preferences; in social recommendations, user relationship graphs can be reconstructed to denoise and mine potential associations.

[0004] However, the potential of existing generative self-supervised learning methods in recommendation systems has not been fully explored. The main problem is that they focus too much on structural reconstruction and ignore the importance of semantic reconstruction. Specifically, existing methods often focus on restoring the structural information of the graph, but do not effectively utilize the rich semantic information in the graph. This one-sidedness leads to the inability to fully explore the two-dimensional information (structure and semantics) of the graph representation. In addition, traditional denoising autoencoders and variational graph encoders have insufficient semantic reconstruction capabilities in collaborative filtering tasks. Due to the limited ability of single-step hidden state compression, they are easily disturbed when processing noisy or erroneous interaction data, resulting in deviations in the optimization of the semantic space, which in turn affects the recommendation effect. Summary of the invention

[0005] The purpose of the present invention is to provide a recommendation method for a graph encoder based on guided diffusion model enhancement to solve the above defects.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A recommendation method for a graph encoder enhanced by a guided diffusion model comprises the following steps:

[0008] S1. Model the user-item interaction network as a bipartite graph G = (V, E), where the node set V represents the users and items in the interaction network, and the edge set E represents the interaction between users and items; initialize the user-item interaction adjacency matrix and normalize it to generate the Laplacian matrix for the model inference stage;

[0009] S2. Enhance the supervision signal from the structural level. Specifically, based on the user's click behavior information and the popularity of the items, the mask probability is adaptively calculated for each click behavior, and a partially visible mask bipartite graph is generated as the input of the mask graph encoder in the model training process.

[0010] S3. Enhance the supervision signal from the semantic level. Specifically, use the mask map encoder to generate the representation of users and items, and perform denoising by adding directed noise and introducing a diffusion model guided by collaborative signals.

[0011] S4. In the reasoning stage, the graph encoder is used to generate user and item representations, and the semantic information of the representations is optimized through the trained diffusion model, and finally the items that may be clicked are predicted for each user.

[0012] Preferably, in step S1, the specific steps are as follows:

[0013] S11. Determine the user-item interaction adjacency matrix A:

[0014]

[0015] Where R is the user-item interaction matrix, T represents the matrix transpose, 0 represents the all-zero matrix, M and N represent the number of users and items respectively;

[0016] S12. Normalize the user-item interaction adjacency matrix and calculate the Laplacian matrix L used in the graph convolution process:

[0017]

[0018] Where A is the interaction adjacency matrix, and D is the degree diagonal matrix of the adjacency matrix.

[0019] Preferably, in step S2, the supervisory signal is enhanced from a structural level, and the specific steps are as follows:

[0020] S21. By perturbing the interaction edges of the original user-item interaction bipartite graph G = (V, E), a masked interaction graph is obtained, and a masked graph encoder is constructed:

[0021]

[0022] In the formula, ε represents the edge set, is a mask vector composed of 0 and 1, o, r, and m represent the original edge set, the remaining edge set, and the masked edge set respectively; do a dot product operation on the original edge set and the mask vector to get the remaining edge set;

[0023] S22, mask vector The adaptive masking strategy yields:

[0024]

[0025] In the formula, u represents a user, i represents an item, (u, i) represents an interaction edge; m is the mask vector The value range of one bit in is {0, 1}, which indicates whether the corresponding edge is masked or retained; p represents the adaptive probability of each edge in the original graph, and Bernoulli is the Bernoulli distribution, that is, the mask m is sampled from the Bernoulli distribution using the adaptive probability of each edge;

[0026] S23. Calculate the adaptive probability p of each edge:

[0027]

[0028] In the formula, scr represents the adaptive score calculated by the edge, and degree represents the degree of the user or item node, that is, the adaptive score is calculated by the degree of the nodes at both ends of the edge;

[0029]

[0030] In the formula, l represents the lowest probability hyperparameter, h represents the highest probability hyperparameter, min represents the minimum value of scr, and max represents the maximum value of scr. That is, the adaptive probability is obtained by normalizing the adaptive score, and the upper and lower bounds of the probability are controlled by the hyperparameters.

[0031] S24. After using the mask vector obtained in step S22, a mask map can be obtained. The mask map is used as a map encoder in the training phase, and the embedded representation of users and items is obtained through the neighbor aggregation formula:

[0032]

[0033] In the formula, u (l) represents the embedding representation of the user at layer l, represents the mask image, u (l-1) represents the embedding representation of the user at the l-1th layer, and N represents the neighboring nodes of the node. Through this formula, the embedding representation of each layer is obtained;

[0034] S25. After obtaining the convolutional embedding representation of each layer of the user and item through the mask graph encoder, the final user and item representation is obtained by aggregating each layer of representation:

[0035]

[0036] In the formula, u e is the final embedding representation of the user, and L represents the number of graph convolution layers.

[0037] Preferably, in step S3, the supervisory signal is enhanced from a semantic level, and the specific steps are as follows:

[0038] S31, after the final user and item embedding is obtained by the mask image encoder in step S25, the semantic level reconstruction task is performed through the scaled guided diffusion model, and the forward process of the diffusion model is performed using a directed noise addition strategy:

[0039]

[0040] In the formula, u t represents the user representation of the t-th step of adding noise, α represents a value between (0,1), which controls the size of the noise. yes And ∈ d is to add directed random noise;

[0041] S32. In order to adapt to the anisotropy of user and item embedding, the random noise of the forward diffusion process is restricted to directed noise, and the formula is:

[0042]

[0043] Where sign is the sign function, which aligns the noise direction with the direction of the user or object embedding. is anisotropic Gaussian noise, as follows:

[0044]

[0045] Where ∈ is random Gaussian noise, μ and σ are the mean and variance of the user and item embedding representations, respectively;

[0046] S33. After step S32, the user and item embeddings at time t are obtained, and the reverse process of the diffusion model is performed:

[0047]

[0048] In the formula, It is the user embedding representation at time t after reconstruction, represents the normal distribution, μ θ and ∑θ are Gaussian distribution parameters determined by the denoising neural network, and the learnable parameter θ is shared between users and items;

[0049] S34. In step S33, in the reverse process of the diffusion model, a denoising model needs to be learned to fit the mean and variance of the distribution, so that the denoising model can denoise from time t to time t-1; the mean fitting formula is as follows:

[0050]

[0051] In the formula, u e t It is directly obtained through the forward process formula of step S31, α is a hyperparameter specified artificially, and the model fitting u e 0 , that is, the model needs to fit the initial representation of the user or item at time 0;

[0052] S35. Use the guided denoising model to complete the reconstruction task under the following conditions:

[0053]

[0054] In the formula, MLP is a multi-layer perceptron, u obtained after MLP reconstruction e 0 , is the user embedding at time t, where t is the input representing the time. and They are conditional embedding and unconditional embedding, respectively. The unconditional embedding is represented by a full 0 vector, while the conditional embedding formula is as follows:

[0055]

[0056] In the formula, I u is the interaction record of user u, i e It is the representation of the item, that is, the conditional embedding of reconstructing the user representation is the user's historical interaction, and the user's preferences are used to guide the denoising and reconstruction process.

[0057] Preferably, the step S4 is specifically as follows:

[0058] S41. Through steps S21-S25, the final embedding representation of the user and the item is obtained, and then the embedding parameters are updated through back propagation through the loss function of structural reconstruction:

[0059]

[0060] In the formula, represents the recommendation loss, u represents the user, i + with i - They represent items that users have interacted with and items that they have not interacted with (positive and negative samples), σ represents the sigmoid function, and represents the user's rating of the positive and negative sample pairs, Represents the original user-item interaction graph;

[0061]

[0062] In the formula, u e and i e Represent the embedding representation of users and items respectively, T represents the matrix transpose; score Obtained by the dot product of the transposed embedding of the user and item, used to calculate the loss and the final predicted ranking;

[0063] S42, execute steps S31-S35, perform semantic reconstruction through a guided diffusion model, and then calculate the reconstruction loss and back-propagate to update the parameters:

[0064]

[0065] In the formula, and They are conditional reconstruction loss and unconditional reconstruction loss; reconstructed by MLP with u e Using mean square error as the loss function, the final semantic level reconstruction loss is:

[0066]

[0067] In the formula, is the final semantic-level loss, which is the average of the conditional reconstruction loss and the unconditional reconstruction loss;

[0068] S43, repeat steps S41-S42 until the parameters of the entire recommendation model converge; in the inference stage of the model, use the scaling guidance strategy to further optimize the embedding representation of users and items, and finally predict the score and sort by the dot product of the user and item embeddings:

[0069]

[0070] In the formula, and They are the user embedding representations under conditional reconstruction and unconditional reconstruction respectively. The final restored user embedding is obtained by scaling-guided denoising.

[0071] The beneficial effects of the present invention are:

[0072] (1) The present invention proposes a recommendation method based on a graph encoder enhanced by a guided diffusion model, which comprehensively considers the structural level and semantic level of the graph. By designing a guided diffusion enhanced mask graph autoencoder, it realizes the deep mining of collaborative signals in the graph recommendation system. This new paradigm makes up for the problem of insufficient balance between structural reconstruction and semantic reconstruction in existing methods.

[0073] (2) The present invention effectively reconstructs the structural information of the graph through an adaptive mask encoder, deeply explores the collaborative relationship between users and items, enhances the recommendation model's ability to capture user behavior patterns, and provides more accurate structural features for recommendation tasks.

[0074] (3) To address the problem of insufficient semantic reconstruction, the present invention designs a dedicated guided diffusion model to optimize the collaborative filtering process from a semantic level. This method accurately reconstructs the semantic space by gradually adding and removing noise, fully utilizes self-supervisory signals, and improves the performance and robustness of the recommendation system from different dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of the method of the present invention;

[0076] Figure 2 A visualization diagram of embedding clustering of a real data set in an embodiment;

[0077] Figure 3 It is a visualization diagram of the embedding distribution of a real data set in the embodiment. DETAILED DESCRIPTION

[0078] The present invention is further described below in conjunction with the embodiments. It should be noted that these are merely examples and illustrations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should be deemed to fall within the protection scope of the present invention.

[0079] Embodiment 1:

[0080] Figure 1 Flow chart of the method of the present invention. Figure 1 As shown, a recommendation method for a graph encoder enhanced by a guided diffusion model includes the following steps:

[0081] S1. Model the user-item interaction network as a bipartite graph G = (V, E), where the node set V represents the users and items in the interaction network, and the edge set E represents the interaction behavior between users and items. Initialize the user-item interaction adjacency matrix and perform Laplace normalization on it to generate a Laplace matrix for the model inference stage.

[0082] The user-item interaction network is modeled as a bipartite graph G = (V, E), where the node set V represents the users and items in the interaction network, and the edge set E represents the interaction between users and items. The user-item interaction adjacency matrix is ​​initialized. The specific steps are as follows:

[0083] S11. Determine the user-item interaction adjacency matrix A:

[0084]

[0085] Where R is the user-item interaction matrix, T is the matrix transpose, 0 is the all-zero matrix, M and N are the number of users and items;

[0086] S12. Normalize the user-item interaction adjacency matrix and calculate the Laplacian matrix used in the graph convolution process

[0087]

[0088] Where A is the interaction adjacency matrix, and D is the degree diagonal matrix of the adjacency matrix.

[0089] S2. Enhance the supervision signal from the structural level. Specifically, based on the user's click behavior information and the popularity of items, the mask probability is adaptively calculated for each click behavior, and a partially visible masked bipartite graph is generated as the input of the mask graph encoder during the model training process.

[0090] like Figure 1 As shown in the adaptive mask of the UI diagram in , the supervision signal is enhanced from the structural level. The specific steps are as follows:

[0091] S21. By perturbing the interaction edges of the original user-item interaction bipartite graph G = (V, E), we obtain the masked interaction graph and construct the masked graph encoder:

[0092]

[0093] In the formula, ε represents the edge set, is a mask vector composed of 0 and 1, o, r, and m represent the original edge set, the remaining edge set, and the masked edge set, respectively; that is, the original edge set is composed of the remaining edges and the mask edge set, and the dot product operation is performed on the original edge set and the mask vector to obtain the remaining edge set.

[0094] S22, mask vector The adaptive masking strategy yields:

[0095]

[0096] In the formula, u represents a user, i represents an item, (u,i) represents an interaction edge; m is the mask vector The value range of one bit in is {0,1}, which indicates whether the corresponding edge is masked or retained; p represents the adaptive probability of each edge in the original graph, and Bernoulli is the Bernoulli distribution, that is, the mask m is sampled from the Bernoulli distribution using the adaptive probability of each edge.

[0097] S23. Calculate the adaptive probability p of each edge:

[0098]

[0099] In the formula, scr represents the adaptive score calculated by the edge, and degree represents the degree of the user or item node, that is, the adaptive score is calculated by the degree of the nodes at both ends of the edge;

[0100]

[0101] In the formula, l represents the lowest probability hyperparameter, h represents the highest probability hyperparameter, min represents the minimum value of scr, and max represents the maximum value of scr. That is, the adaptive probability is obtained by normalizing the adaptive score, and the upper and lower bounds of the probability are controlled by hyperparameters.

[0102] S24. After using the mask vector obtained in step S22, a mask map can be obtained. The mask map is used as a map encoder in the training phase, and the embedded representation of users and items is obtained through the neighbor aggregation formula:

[0103]

[0104] In the formula, u (l) represents the embedding representation of the user at layer l, represents the mask image, u (l-1) represents the embedding representation of the user at the l-1th layer, N represents the neighboring nodes of the node, and through this formula, the embedding representation of each layer is obtained.

[0105] S25. After obtaining the convolutional embedding representation of each layer of the user and item through the mask graph encoder, the final user and item representation is obtained by aggregating each layer of representation:

[0106]

[0107] In the formula, u e is the final embedding representation of the user, and L represents the number of graph convolution layers.

[0108] S3. Further enhance the supervision signal from the semantic level. Specifically, the mask map encoder is used to generate the representation of users and items, and denoising is performed by adding directed noise and introducing a diffusion model guided by collaborative signals.

[0109] like Figure 1 As shown in the diffusion model process and denoising model in the figure, the supervision signal is enhanced from the semantic level. The specific steps are as follows:

[0110] S31, after the final user and item embedding is obtained by the mask image encoder in step S25, the semantic level reconstruction task is performed through the scaled guided diffusion model, and the forward process of the diffusion model is performed using a directed noise addition strategy:

[0111]

[0112] In the formula, u t represents the user representation of the t-th step of adding noise, α represents a value between (0,1), which controls the size of the noise. yes And ∈ d is to add directed random noise.

[0113] S32. In order to adapt to the anisotropy of user and item embedding, the random noise of the forward diffusion process is restricted to directed noise, and the formula is:

[0114]

[0115] Where sign is the sign function, which aligns the noise direction with the direction of the user or object embedding. is anisotropic Gaussian noise, as follows:

[0116]

[0117] Where ∈ is random Gaussian noise, μ and σ are the mean and variance of the user and item embedding representations, respectively, so that the Gaussian noise conforms to anisotropy.

[0118] S33. After step S32, the user and item embeddings at time t are obtained, and the reverse process of the diffusion model is performed:

[0119]

[0120] In the formula, It is the user embedding representation at time t after reconstruction, represents the normal distribution, μθ and ∑θ are the Gaussian distribution parameters determined by the denoising neural network, and the learnable parameter θ is shared between users and items.

[0121] S34. In step S33, in the reverse process of the diffusion model, a denoising model needs to be learned to fit the mean and variance of the distribution, so that the denoising model can denoise from time t to time t-1; for the convenience of fitting, here we only fit the mean, and the mean fitting formula is as follows:

[0122]

[0123] In the formula, u e t It is directly obtained through the forward process formula of step S31. α is a hyperparameter specified artificially, so only the model needs to fit u e 0 , that is, the model needs to fit the initial representation of users or items at time 0.

[0124] S35. Use the guided denoising model to complete the reconstruction task under the following conditions:

[0125]

[0126] In the formula, MLP is a multi-layer perceptron, After MLP reconstruction, is the user embedding at time t, where t is the input representing the time. and They are conditional embedding and unconditional embedding, respectively. The unconditional embedding is represented by a full 0 vector, while the conditional embedding formula is as follows:

[0127]

[0128] In the formula, I u is the interaction record of user u, i e It is the representation of the item, that is, the conditional embedding of reconstructing the user representation is the user's historical interaction, and the user's preferences are used to guide the denoising and reconstruction process.

[0129] S4. In the inference phase, the graph encoder is used to generate user and item representations, and the semantic information of the representations is optimized through the trained diffusion model, and finally the items that each user may click are predicted, as follows:

[0130] S41. Through steps S21-S25, the final embedding representation of the user and the item is obtained through the mask image encoder, and then the embedding parameters are updated through back propagation through the loss function of structural reconstruction:

[0131]

[0132] In the formula, represents the recommendation loss, u represents the user, i + with i - They represent items that users have interacted with and items that they have not interacted with (positive and negative samples), σ represents the sigmoid function, and represents the user's rating of the positive and negative sample pairs, Represents the original user-item interaction graph;

[0133]

[0134] In the formula, u e and i e Represent the embedding representation of users and items respectively, T represents the matrix transpose; score It is obtained by transposing the dot product of user and item embeddings and is used to calculate the loss and the final predicted ranking.

[0135] S42, execute steps S31-S35, perform semantic reconstruction through a guided diffusion model, and then calculate the reconstruction loss and back-propagate to update the parameters:

[0136]

[0137] In the formula, and They are conditional reconstruction loss and unconditional reconstruction loss; reconstructed by MLP and Using mean square error as the loss function, the final semantic level reconstruction loss is:

[0138]

[0139] In the formula, is the final semantic-level loss, which is the average of the conditional reconstruction loss and the unconditional reconstruction loss.

[0140] S43, repeat steps S41-S42 until the parameters of the entire recommendation model converge; in the inference stage of the model, use the scaling guidance strategy to further optimize the embedding representation of users and items, and finally predict the score and sort by the dot product of the user and item embeddings:

[0141]

[0142] In the formula, and They are the user embedding representations under conditional reconstruction and unconditional reconstruction respectively. The final restored user embedding is obtained by scaling-guided denoising.

[0143] like Figure 2and Figure 3 As shown in the figure, 1000 users and 1000 items were randomly selected from the real data. The embeddings were clustered by k-means clustering, and the distribution of user embeddings was visualized in two dimensions. When the denoising module was removed, the embedding representation showed a uniform distribution and lacked the fine capture of user behavior characteristics. After the denoising module was introduced, the embeddings of users and items were more closely clustered into their respective categories, showing a clearer synergistic relationship. This shows that the denoising module can optimize the embedding representation through synergistic signals and reduce the impact of noise. In practical scenarios, such as e-commerce platforms, the denoising module can help the model more accurately identify the core interests of users and avoid the interference of accidental clicks or invalid data on the recommendation results. This method not only improves the accuracy of recommendations, but also enhances the robustness of the system in the face of complex data distribution, thereby better meeting the needs of practical applications. The visualization results of the embedding further verify the potential of this method in improving the performance of recommendation systems.

[0144] The invention discloses a recommendation method based on a graph encoder enhanced by a guided diffusion model, which comprehensively considers the structural level and semantic level of the graph, and realizes in-depth mining of collaborative signals in the graph recommendation system by designing a guided diffusion enhanced mask graph autoencoder. This new paradigm makes up for the problem of insufficient balance between structural reconstruction and semantic reconstruction in existing methods.

[0145] The present invention effectively reconstructs the structural information of the graph through an adaptive mask encoder, deeply explores the collaborative relationship between users and items, enhances the recommendation model's ability to capture user behavior patterns, and provides more accurate structural features for recommendation tasks.

[0146] In order to solve the problem of insufficient semantic reconstruction, this paper designs a dedicated guided diffusion model to optimize the collaborative filtering process from the semantic level. This method accurately reconstructs the semantic space by gradually adding and removing noise, makes full use of self-supervisory signals, and improves the performance and robustness of the recommendation system from different dimensions.

[0147] The invention discloses a recommendation method based on a graph encoder enhanced by a guided diffusion model. The method adopts an adaptive mask encoder to reconstruct the structural information of a graph and fully explore the collaborative relationship between users and items. The method focuses on optimizing the collaborative filtering process from the semantic level of the graph by designing a diffusion model for semantic reconstruction. This two-pronged strategy can better utilize self-supervisory signals and enhance the effect of graph recommendation from two different perspectives: structure and semantics.

[0148] The above is an exemplary description of the invention. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as such non-substantial improvements are made by adopting the method concept and technical solution of the present invention, or the concept and technical solution of the invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A recommendation method for a graph encoder enhanced by a guided diffusion model, characterized in that: The following steps are involved: S1. Model the user-item interaction network as a bipartite graph G = (V, E), where the node set V represents the users and items in the interaction network, and the edge set E represents the interaction between users and items; initialize the user-item interaction adjacency matrix and normalize it to generate the Laplacian matrix for the model inference stage; S2. Enhance the supervision signal from the structural level. Specifically, based on the user's click behavior information and the popularity of the items, the mask probability is adaptively calculated for each click behavior, and a partially visible mask bipartite graph is generated as the input of the mask graph encoder in the model training process. S3. Further enhance the supervision signal from the semantic level. Specifically, use the mask map encoder to generate the representation of users and items, and perform denoising by adding directed noise and introducing a diffusion model guided by collaborative signals. S4. In the reasoning stage, the graph encoder is used to generate user and item representations, and the semantic information of the representations is optimized through the trained diffusion model, and finally the items that may be clicked are predicted for each user.

2. The method for recommending a graph encoder based on guided diffusion model enhancement according to claim 1, characterized in that: In step S1, the user-item interaction network is modeled as a bipartite graph G = (V, E), where the node set V represents the users and items in the interaction network, and the edge set E represents the interaction behavior between users and items. The user-item interaction adjacency matrix is ​​initialized. The specific steps are as follows: S11. Determine the user-item interaction adjacency matrix A: Where R is the user-item interaction matrix, T represents the transpose, 0 represents the all-zero matrix, M and N represent the number of users and items respectively; S12. Normalize the user-item interaction adjacency matrix and calculate the Laplacian matrix used in the graph convolution process Where A is the interaction adjacency matrix, and D is the degree diagonal matrix of the adjacency matrix.

3. The method for recommending a graph encoder based on guided diffusion model enhancement according to claim 1, characterized in that: In step S2, the supervisory signal is enhanced from the structural level. The specific steps are as follows: S21. By perturbing the interaction edges of the original user-item G = (V, E), we obtain the masked interaction graph and construct the masked graph encoder: e o =e r +e m , In the formula, ε represents the edge set, is a mask vector composed of 0 and 1, o, r, and m represent the original edge set, the remaining edge set, and the masked edge set respectively; do a dot product operation on the original edge set and the mask vector to get the remaining edge set; S22, mask vector The adaptive masking strategy yields: In the formula, u represents a user, i represents an item, (u,i) represents an interaction edge; m is the mask vector One bit in , with a value range of {0,1}, represents whether the corresponding edge is masked or retained; p represents the adaptive probability of each edge in the original graph, and Bernoulli is the Bernoulli distribution, that is, the mask m is sampled from the Bernoulli distribution using the adaptive probability of each edge; S23. Calculate the adaptive probability p of each edge: In the formula, scr represents the adaptive score calculated by the edge, and degree represents the degree of the user or item node, that is, the adaptive score is calculated by the degree of the nodes at both ends of the edge; In the formula, l represents the lowest probability hyperparameter, h represents the highest probability hyperparameter, min represents the minimum value of scr, and max represents the maximum value of scr. That is, the adaptive probability is obtained by normalizing the adaptive score, and the upper and lower bounds of the probability are controlled by the hyperparameters. S24. After using the mask vector obtained in step S22, a mask map can be obtained. The mask map is used as a map encoder in the training phase, and the embedded representation of users and items is obtained through the neighbor aggregation formula: In the formula, u (l) represents the embedding representation of the user at layer l, represents the mask image, u (l-1) represents the embedding representation of the user at the l-1th layer, and N represents the neighboring nodes of the node. Through this formula, the embedding representation of each layer is obtained; S25. After obtaining the convolutional embedding representation of each layer of the user and item through the mask graph encoder, the final user and item representation is obtained by aggregating each layer of representation: In the formula, u e is the final embedding representation of the user, and L represents the number of graph convolution layers.

4. The method for recommending a graph encoder based on guided diffusion model enhancement according to claim 3, characterized in that: In step S3, the supervision signal is enhanced from the semantic level. The specific steps are as follows: S31, after the final user and item embedding is obtained by the mask image encoder in step S25, the semantic level reconstruction task is performed through the scaled guided diffusion model, and the forward process of the diffusion model is performed using a directed noise addition strategy: In the formula, u t represents the user representation of the t-th step of adding noise, α represents a value between (0,1), which controls the size of the noise. yes And ∈ d is to add directed random noise; S32. In order to adapt to the anisotropy of user and item embedding, the random noise of the forward diffusion process is restricted to directed noise, and the formula is: Where sign is the sign function, which aligns the noise direction with the direction of the user or object embedding. is anisotropic Gaussian noise, as follows: Where ∈ is random Gaussian noise, μ and σ are the mean and variance of the user and item embedding representations, respectively; S33. After step S32, the user and item embeddings at time t are obtained, and the reverse process of the diffusion model is performed: In the formula, It is the user embedding representation at time t after reconstruction, represents the normal distribution, μ θ and ∑θ are Gaussian distribution parameters determined by the denoising neural network, and the learnable parameter θ is shared between users and items; S34. In step S33, in the reverse process of the diffusion model, a denoising model needs to be learned to fit the mean and variance of the distribution, so that the denoising model can denoise from time t to time t-1; the mean fitting formula is as follows: In the formula, u e t It is directly obtained through the forward process formula of step S31, α is a hyperparameter specified artificially, and the model fitting u e 0, that is, the model needs to fit the initial representation of the user or item at time 0; S35. Use the guided denoising model to complete the reconstruction task under the following conditions: Where MLP is a multi-layer perceptron. After MLP reconstruction, we get is the user embedding at time t, where t is the input representing the time. and They are conditional embedding and unconditional embedding, respectively. The unconditional embedding is represented by a full 0 vector, while the conditional embedding formula is as follows: In the formula, I u is the interaction record of user u, i e It is the representation of the item, that is, the conditional embedding of reconstructing the user representation is the user's historical interaction, and the user's preferences are used to guide the denoising and reconstruction process.

5. The method for recommending a graph encoder based on guided diffusion model enhancement according to claim 4, characterized in that: The S4 step is specifically as follows: S41. Through steps S21-S25, the final embedding representation of the user and the item is obtained, and then the embedding parameters are updated through back propagation through the loss function of structural reconstruction: In the formula, represents the recommendation loss, u represents the user, i + with i - They represent items that users have interacted with and items that they have not interacted with (positive and negative samples), σ represents the sigmoid function, and represents the user's rating of the positive and negative sample pairs, Represents the original user-item interaction graph; In the formula, u e and i e Represent the embedding representation of users and items respectively, T represents transposition; score Obtained by the dot product of the transposed embedding of the user and item, used to calculate the loss and the final predicted ranking; S42, execute steps S31-S35, perform semantic reconstruction through a guided diffusion model, and then calculate the reconstruction loss and back-propagate to update the parameters: In the formula, and They are conditional reconstruction loss and unconditional reconstruction loss; reconstructed by MLP with u e Using mean square error as the loss function, the final semantic level reconstruction loss is: In the formula, is the final semantic-level loss, which is the average of the conditional reconstruction loss and the unconditional reconstruction loss; S43, repeat steps S41-S42 until the parameters of the entire recommendation model converge; in the inference stage of the model, use the scaling guidance strategy to further optimize the embedding representation of users and items, and finally predict the score and sort by the dot product of the user and item embeddings: In the formula, and They are the user embedding representations under conditional reconstruction and unconditional reconstruction respectively. The final restored user embedding is obtained by scaling-guided denoising.

Images