Article recommendation method based on hidden space condition diffusion

Abstract

The invention discloses an article recommendation method based on hidden space condition diffusion, and relates to the technical field of computers. The method comprises the following steps: reducing the dimension of a matrix corresponding to a social relation graph into low-dimensional social relation embedding; performing singular value decomposition on a result of transpose multiplication of the matrix corresponding to the user item interaction diagram and the matrix corresponding to the user item interaction diagram to obtain inter-user common interaction article embedding; reconstruction is carried out from noise-added low-dimensional social relation embedding in combination with inter-user common interaction article embedding, and de-noised social relation embedding is obtained; learning the user item interaction diagram to obtain user item interaction embedding, and splitting the user item interaction embedding into user embedding and article embedding; and fusing the de-noised social relation embedding into the user embedding to obtain a fused user embedding, calculating an inner product of the fused user embedding and the article embedding, and determining a recommended article according to the size of the inner product. The method can improve the accuracy of article recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an item recommendation method based on hidden space conditional diffusion. BACKGROUND

[0002] In the field of recommendation methods, the research of social recommendation is concerned. Graph Convolutional Network (GNN) is widely used to encode social connection graph as auxiliary information to enhance user recommendation for items. However, when GNN encodes social connection graph into latent space, it will cause distortion of node embedding due to average aggregation of neighbor node information in user social connection graph. In addition, there is noise in the graph data which belongs to non-key features.

[0003] Diffusion model also has important application in the field of social recommendation, which can be used to denoise social connection embedding. Diffusion model gradually adds Gaussian noise to embedding until it completely changes into pure noise, and then gradually reconstructs the embedding through the reverse process. However, relying only on diffusion model to denoise pure noise embedding will make the generated embedding have low alignment with the recommendation task, thereby affecting the performance of item recommendation.

[0004] Therefore, there is an urgent need for a method that can improve the accuracy of item recommendation. SUMMARY

[0005] Therefore, it is necessary to provide an item recommendation method based on hidden space conditional diffusion in view of the above technical problems. The method can improve the accuracy of item recommendation.

[0006] The present application adopts the following technical solutions: The present application provides an item recommendation method based on hidden space conditional diffusion, comprising: orthogonal triangular decomposition and singular value decomposition to reduce the matrix corresponding to the social relationship graph to a low-dimensional social relationship embedding in the hidden space; the hidden space is a latent space; determining the multiplication result of the matrix corresponding to the user-item interaction graph and the transpose of the matrix corresponding to the user-item interaction graph as the inter-user common interaction item matrix, and obtaining a low-dimensional inter-user common interaction item embedding through singular value decomposition of the inter-user common interaction item matrix; adding noise to the low-dimensional social relationship embedding, and combining the low-dimensional inter-user common interaction item embedding to reconstruct from the low-dimensional social relationship embedding with added noise to obtain a denoised social relationship embedding; learning the user-item interaction graph through a graph autoencoder to obtain a user-item interaction embedding, and splitting the user-item interaction embedding into a user embedding and an item embedding; The denoised social relationship embedding is integrated into the user embedding to obtain the fused user embedding. The inner product between the positive sample of the fused user embedding and the positive sample of the item embedding, as well as the inner product between the negative sample of the fused user embedding and the negative sample of the item embedding are calculated respectively. The inner product of the positive sample is subtracted from the inner product of the negative sample. The result of the subtraction is determined as the user's predicted score for the item, and the predicted scores are sorted in descending order. The items corresponding to the preset number of prediction scores are determined as recommendation results.

[0007] Preferably, the user-item interaction graph is learned by a graph autoencoder to obtain a user-item interaction embedding, specifically including: Perform multiple iterative generation operations on the user-item interaction graph to obtain multiple embedding matrices of the user-item interaction graph, where the first l The embedding matrix generated by the iterative generation operation is: ; in, For the l The embedding matrix generated in the iterations is D is the degree matrix, is the normalization function, W is the weight matrix, Interaction diagram for user projects The adjacency matrix of For the l- The embedding matrix generated in 1 iteration; Aggregate the embedding matrices generated by multiple iterations to obtain the user-item interaction embedding; the formula corresponding to the aggregation operation is: ; in, Embed for user project interaction, L is the number of iterations, For the l The embedding matrix generated by the iteration.

[0008] Preferably, the matrix corresponding to the social relationship graph is reduced to a low-dimensional social relationship embedding in the latent space by orthogonal triangular decomposition and singular value decomposition, specifically including: The matrix corresponding to the social relationship graph Multiplying with the random matrix to obtain a first matrix; the first matrix is ​​used to project the column space of the matrix corresponding to the social relationship graph into the low-dimensional subspace spanned by the random matrix; Based on the first matrix, a power iteration generation result is obtained by performing multiple power iteration operations; Decompose the result generated by the last power iteration using the orthogonal triangular decomposition method to obtain an orthogonal matrix and an upper triangular matrix; Multiply the transpose of the orthogonal matrix with the matrix corresponding to the social relation graph to obtain a second matrix; Perform singular value decomposition on the second matrix and retain the first k principal components; Determine the retained first k principal components as a low-dimensional social relation embedding.

[0009] Preferably, noise is added to the low-dimensional social relation embedding, specifically including: Step 0, no noise is added; Step t- 1 to step t The noise addition process is: ; Wherein, is a standard normal distribution, is the low-dimensional social relation embedding with added noise obtained in step t is the low-dimensional social relation embedding with added noise obtained in step , t- , is a unit matrix, I is a scale parameter in step is a scale parameter in step t -1; By the additivity of the standard normal distribution, the low-dimensional social relation embedding with added noise obtained in step t is simplified to obtain the low-dimensional social relation embedding with added noise in step Step t is: t ; Wherein, , , and represent independent Gaussian distributions, is the variance of the Gaussian noise added to control the diffusion in step t is a standard Gaussian distribution, is a Gaussian variable that combines multiple steps of noise, is the residual signal strength of the low-dimensional social relation embedding after diffusion in step . t

[0010] Preferably, the corresponding formula for reconstruction is: ; Wherein, and represent a neural network with learnable parameters . t ​​is a time step, z is a guided embedding, represents a standard normal distribution, is a node embedding vector in the first diffusion step, t- is a node embedding vector in the first diffusion step, is a low-dimensional social relationship embedding with added noise obtained in the first step, t is a low-dimensional social relationship embedding with added noise obtained in the first step, is a low-dimensional social relationship embedding with added noise obtained in the first step, t- is a low-dimensional social relationship embedding with added noise obtained in the first step; the guided embedding is an encoding result of a low-dimensional inter-user common interaction item embedding.

[0011] Preferably, the construction process of the user-item interaction graph specifically comprises: constructing a user set and an item set ; representing the interaction behavior of users and items by a matrix , each element in the matrix R , represents that the user has interaction with the item , represents that the user has no interaction with the item ; based on the user set, the item set and the matrix R , constructing a collaborative bipartite graph , and determining the collaborative bipartite graph as the user-item interaction graph.

[0012] Preferably, the construction process of the social relationship graph specifically comprises: constructing a social relationship matrix between users , each element in the matrix S , represents that there is a social connection between the user and the user , represents that there is no social connection between the user and the user ; based on the user set and the social relationship matrix, constructing a social relationship graph .

[0013] The application provides an item recommendation device based on hidden space conditional diffusion, comprising: a first dimension reduction module, configured to reduce the matrix corresponding to the social relationship graph to a low-dimensional social relationship embedding in a hidden space by orthogonal triangular decomposition and singular value decomposition; the hidden space is a latent space; The second dimension reduction module is configured to multiply the user-item interaction graph corresponding matrix and a transpose of the user-item interaction graph corresponding matrix to obtain a user-common interaction item matrix, and to obtain a low-dimensional user-common interaction item embedding by singular value decomposition of the user-common interaction item matrix. The reconstruction module is configured to add noise to the low-dimensional social relationship embedding, and to reconstruct the low-dimensional social relationship embedding with noise added to obtain a denoised social relationship embedding. The learning module is configured to learn the user-item interaction graph by the graph autoencoder to obtain a user-item interaction embedding, and to split the user-item interaction embedding into a user embedding and an item embedding. The calculation module is configured to fuse the denoised social relationship embedding into the user embedding to obtain a fused user embedding, to calculate inner products between positive samples of the fused user embedding and positive samples of the item embedding and between negative samples of the fused user embedding and negative samples of the item embedding, to subtract the inner products of the positive samples from the inner products of the negative samples, to determine a result of the subtraction as a predicted score of the user for the item, and to sort the predicted scores in descending order. The determination module is configured to determine items corresponding to the first preset number of predicted scores as the recommendation result.

[0014] The present application provides a computer readable storage medium, the storage medium stores a computer program, the computer program is executed by a processor to realize the above-mentioned social recommendation method based on hidden space conditional diffusion.

[0015] The present application provides a computer device, including memory, processor and computer program stored in the memory and running on the processor, the processor executes the program to realize the above-mentioned social recommendation method based on hidden space conditional diffusion.

[0016] The above-mentioned at least one technical scheme adopted by the present application can achieve the following beneficial effects: The noise is added to the low-dimensional social relationship embedding, and the low-dimensional common interaction item embedding between users is combined, the social relationship embedding after denoising is obtained by reconstruction from the low-dimensional social relationship embedding after adding noise, the social noise is corrected by using the common interaction item embedding between users, and the credibility is improved; the user-item interaction embedding is obtained by learning the user-item interaction graph through the graph autoencoder, and the user-item interaction embedding is split into user embedding and item embedding, feature decoupling is realized, and targeted fusion is facilitated; the social relationship embedding after denoising is integrated into the user embedding to obtain the fusion user embedding, which can supplement the social influence feature; the inner product between the positive sample of the fusion user embedding and the positive sample of the item embedding and the inner product between the negative sample of the fusion user embedding and the negative sample of the item embedding are calculated respectively, the inner product of the positive sample is subtracted from the inner product of the negative sample, the result of the subtraction is determined as the prediction score of the user to the item, the negative sample interference is eliminated, the effective recommendation signal is highlighted, and the prediction scores are sorted in descending order; the items corresponding to the first preset number of prediction scores are determined as the recommendation result. The method can improve the accuracy of item recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application. In the drawings:

[0018] Figure 1 A flowchart of an item recommendation method based on hidden space conditional diffusion provided by the present application; Figure 2 A general architecture diagram of the CGDLS model provided by the present application; Figure 3 A conditional guided diffusion model diagram of the CGDLS model provided by the present application; Figure 4 A schematic diagram of an item recommendation device based on hidden space conditional diffusion provided by the present application; Figure 5 A schematic diagram of a computer device for implementing an item recommendation method based on hidden space conditional diffusion provided by the present application. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in conjunction with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] A device such as a desktop, a notebook, etc. that executes the present scheme. For the convenience of explanation, a server will be explained as an execution subject below.

[0021] Significant progress has been made in early social recommendation research. For example, a social-aware item-aware sequence is generated using random walk, and social information is mined from distant neighbors for deep collaborative filtering recommendation, while a social embedding is learned using a deep autoencoder, and is combined with a collaborative neural network to model the shallow linear interaction and deep nonlinear relationship between users and items. Recently, node embeddings in the collaborative and social fields are jointly encoded through feature enhancement and bidirectional transfer mechanisms to align the features of the same positive sample pair. In contrast, social networks and user community interaction graphs are integrated through global propagation, and community subgraph learning is enhanced through local propagation. However, these social recommendation methods ignore the potential noise information in user social relationships, such as fake social relationships. Such fake social relationships as noise information can reduce the performance of social recommendation methods. In addition, for social users with limited interests, over-reliance on social information can mislead the recommendation system, especially when users randomly add others as friends or follow them just to increase their exposure or popularity.

[0022] In recent research on recommendation systems, diffusion models have shown significant effects in generating denoised embeddings. Diffusion models gradually add Gaussian noise to the embedding until it is completely transformed into pure Gaussian noise, and then gradually reconstruct the embedding through the reverse process. Recently proposed architectures, such as DiffRec, explicitly handle noise in social interactions through diffusion models, combining temporal dynamics and latent space techniques to achieve optimized social relationship modeling. Unlike before, Diff4Rec effectively extends interaction data by constructing a user behavior generation model in latent space and adopting a progressive curriculum reinforcement strategy. In addition, RecDiff uses diffusion models to denoise the embedding of social connections between users. However, when using the reverse process of the diffusion model to denoise the pure noise embedding, the denoising direction may be weakly associated with the recommendation task. Relying solely on diffusion models to denoise pure noise embeddings, embedding denoising can make it challenging to generate embeddings that are highly aligned with the recommendation task, which can affect recommendation performance.

[0023] For social recommendation, to generate more task-relevant social relation embeddings, recent studies proposed conditional guided diffusion models that guide the reverse process of diffusion models to generate denoised social relation embeddings that are highly consistent with the recommendation task. For example, Composite Split Denoising (CSD) and Conditional Guided Diffusion (CGD) are used to remove redundant social relations, dynamically adjust preference weights, and reduce popularity bias by optimizing preferences. Neighbor preferences are used to guide the generation of high-quality social relations during the reverse process of the diffusion model. However, since users who interact with the same item tend to exhibit higher similarity in preferences, such users not only have a higher likelihood of establishing social connections but also have similar preferences, possibly belonging to the same user category. Furthermore, users in the same category are more likely to have common interests in the same item. Therefore, these methods ignore the use of embeddings of common interaction items extracted from the user-item interaction graph as conditional guidance.

[0024] Furthermore, most existing methods directly apply diffusion models to graph data. Specifically, numerous methods directly utilize the forward process of diffusion models to introduce noise into graph data. This not only makes the diffusion process difficult to control but also incurs significant computational costs. For example, a social relation matrix is introduced and combined with an inverse logarithmic weight penalty to directly diffuse on the user-item graph. Subsequently, Denoising Diffusion Probabilistic Models (DDPM) are integrated into a collaborative filtering recommendation system by modeling the user-item interaction matrix using diffusion-based methods. However, these methods do not consider encoding graph data in the latent space for use by diffusion models. Encoding social connections between users in the latent space can enhance control over the diffusion process. Furthermore, graph data itself is challenging due to its sparsity, discreteness, and exponential complexity of data space, which can hinder effective diffusion.

[0025] To compress social relation graphs into latent space, some researchers utilize Graph Neural Networks (GNNs) as encoders to compress graphs into low-dimensional embeddings in latent space. Recently, Graph Convolutional Networks (GCN) are used to encode the original social graph into a low-dimensional social embedding in latent space, followed by the use of a stochastic differential equation-based diffusion model to generate social embeddings consistent with preferences. In contrast, GCN is used to encode social connection graphs into latent space, and diffusion models are directly applied to social connection embeddings.

[0026] However, traditional graph neural network methods that encode social connection graphs into latent spaces can distort node embeddings by averaging and aggregating neighbor node information in a user's social connection graph. Furthermore, due to noise in graph data, which is not a key feature, these methods struggle to encode the social connections corresponding to key features in the social connection graph into a low-dimensional latent space and may inadvertently aggregate noisy information, thereby amplifying the impact of noise on the encoded embedding.

[0027] The shortcomings of the prior art are: 1. Ignoring the importance of conditional guidance based on the items of common interactions between users: Existing methods often ignore the fact that the direction of social connection embedding denoising by the diffusion model may be weakly correlated with the direction of the recommendation task during the social recommendation process. The generated social connection embedding is not strongly related to the recommendation task, which affects the effectiveness of using the social connection graph to assist recommendations.

[0028] 2. Ignoring the importance of using SVD to encode social graphs into a low-dimensional latent space: Existing methods directly apply noise and denoise to social graphs using the forward and reverse processes of the diffusion model. This makes the diffusion model computational process difficult to control and is computationally intensive. Furthermore, some methods use graph neural networks to encode into a low-dimensional latent space, resulting in node distortion and amplified noise effects during embedding.

[0029] The purpose of this invention is to propose a conditional guided diffusion model in latent space for social recommendation (CGDLS) to address the limitations of existing social recommendation methods.

[0030] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0031] Figure 1 The figure is a flowchart of an item recommendation method based on latent space conditional diffusion in the present invention, which specifically includes the following steps: S101: Reduce the dimension of the matrix corresponding to the social relationship graph into a low-dimensional social relationship embedding in a latent space through orthogonal triangular decomposition and singular value decomposition; the latent space is a latent space.

[0032] In an exemplary embodiment, the process of constructing the user-project interaction graph specifically includes: constructing a user set and project sets ; Through the matrix Represents the interaction between users and items, matrix R For each element in Represents a user with project has interaction, represents user with project has no interaction; based on the user set, the project set and the matrix R , a collaborative bipartite graph is constructed , and the collaborative bipartite graph is determined as the user-project interaction graph.

[0033] In an exemplary embodiment, the construction process of the social relationship graph specifically includes: constructing a social relationship matrix between users , each element S in the matrix represents that there is a social contact between user and user , represents that there is no social contact between user and user ; based on the user set and the social relationship matrix, a social relationship graph is constructed .

[0034] Specifically, the user set and the project set are defined. The interaction behavior of users with projects is represented by a matrix , wherein represents that user has interacted with project , represents that user has no interaction with project . Similarly, the social relationship between users is represented by a matrix , wherein represents that there is a social contact (such as friendship) between user and user , represents that there is no social contact between user and user . In order to reveal deeper association patterns, these data are converted into a graph structure: based on the interaction behavior of users with projects, a collaborative bipartite graph named user-project interaction graph is constructed, and the edge set of the user-project interaction graph is ; based on the social relationship between users, a social relationship graph is constructed, and the edge set of the social relationship graph is .

[0035] In an exemplary embodiment, the matrix corresponding to the social relationship graph is reduced to a low-dimensional social relationship embedding in a hidden space by orthogonal triangular decomposition and singular value decomposition, specifically including: reducing the matrix corresponding to the social relationship graph to a low-dimensional social relationship embedding in a hidden space by orthogonal triangular decomposition and singular value decomposition Multiply it with the random matrix to obtain the first matrix; the first matrix is ​​used to project the column space of the matrix corresponding to the social relationship graph to the low-dimensional subspace spanned by the random matrix; based on the first matrix, multiple power iteration operations are performed to obtain the power iteration generation result; the last power iteration generation result is decomposed by the orthogonal triangular decomposition method to obtain an orthogonal matrix and an upper triangular matrix; the transpose of the orthogonal matrix is ​​multiplied with the matrix corresponding to the social relationship graph to obtain the second matrix; the second matrix is ​​calculated to perform singular value decomposition, and the previous k principal components; the first k principal components, which are determined as low-dimensional social relationship embeddings.

[0036] Specifically, as a graph encoder, singular value decomposition (SVD) can not only convert high-dimensional graphs into low-dimensional compact embeddings, but also effectively extract key features of the data while significantly reducing computational complexity. Figure 2 The overall architecture of the CGDLS model is shown in Figure 2 In the present invention, SVD is used as a social relationship graph encoder, SVD encodes the social relationship graph, and iterates multiple times to obtain higher quality social relationship embedding.

[0037] Multiply the matrix corresponding to the social relationship graph by the random matrix to obtain the first matrix. The calculation formula of the first matrix is ​​shown in formula (1): (1) in, is the first matrix, is the matrix corresponding to the social relationship graph, is a random matrix. The purpose of the first matrix is ​​to The column space of On Zhang Cheng's low-dimensional subspace.

[0038] By performing multiple iterations to obtain higher quality social relationship embeddings, the multiple iterations are defined as shown in formula (2): (2) in, For the q The result of the iterations, For the q+1 The result of the iteration.

[0039] right Applies a power iteration operation. Represents the orthogonalization process of the matrix. The function of the orthogonalization process is to The column vectors of the matrix are converted into a set of standard orthogonal bases. The orthogonalization process is shown in formula (3):

[0040] (3) After the next iteration, the present application obtains After the next iteration, the present application obtains Then, QR decomposition is performed on The corresponding calculation formula of the QR decomposition is shown in formula (4): (4) Wherein, is an orthogonal matrix, and the column vectors thereof constitute an orthogonal basis of the column space of is an upper triangular matrix. Subsequently, the exact singular value decomposition is calculated for the second matrix Substitute into

[0041] to obtain .

[0042] Wherein, represents the SVD operation on the second matrix , represents the number of principal components reserved after the SVD operation is performed, represents the social relationship embedding obtained by decomposing the second matrix . Since is less than and , the approximate SVD result finally obtained is , which represents the low-dimensional social relationship embedding obtained after the entire calculation process is completed. The corresponding calculation is shown in formula (5): (5).

[0043] S102: Determine the multiplication result of the matrix corresponding to the user-item interaction graph and the transpose of the matrix corresponding to the user-item interaction graph as the user-common interaction item matrix, and obtain the low-dimensional user-common interaction item embedding through singular value decomposition of the user-common interaction item matrix.

[0044] Determine the multiplication result of the matrix corresponding to the user-item interaction graph and the transpose of the matrix corresponding to the user-item interaction graph as the user-common interaction item matrix; encode the user-common interaction item matrix through SVD to obtain the low-dimensional user-common interaction item embedding in the latent space; calculate the average value of the features in the low-dimensional user-common interaction item embedding to obtain the guide embedding. The user-common interaction item graph is constructed based on the user-item interaction graph, and this process can be expressed as shown in formula (6):

[0045] Determine the multiplication result of the matrix corresponding to the user-item interaction graph and the transpose of the matrix corresponding to the user-item interaction graph as the user-common interaction item matrix; encode the user-common interaction item matrix through SVD to obtain the low-dimensional user-common interaction item embedding in the latent space; calculate the average value of the features in the low-dimensional user-common interaction item embedding to obtain the guide embedding. The user-common interaction item graph is constructed based on the user-item interaction graph, and this process can be expressed as shown in formula (6): ​​​

[0046] (6) wherein, z is the guided embedding, is the co-interaction item matrix between users, is the transpose of the co-interaction item matrix between users, is the co-interaction item matrix between users, is the low-dimensional co-interaction item embedding between users, is the average feature vector.

[0047] First, the co-interaction item matrix between users is obtained by matrix multiplication , wherein the off-diagonal elements of the co-interaction item matrix between users represent the number of co-interaction items between each pair of users, second, the graph is encoded using SVD to obtain a low-dimensional co-interaction item embedding between users in the latent space . Finally, the average value of the features in the low-dimensional co-interaction item embedding between users is calculated to obtain , wherein, corresponds to the batch data involved in the current calculation. as the guided embedding, and is integrated into the latent space of the diffusion model backpropagation process to enhance the influence of collaborative recommendation.

[0048] S103: Add noise to the low-dimensional social relationship embedding, combine the low-dimensional co-interaction item embedding between users, and reconstruct from the low-dimensional social relationship embedding with added noise to obtain a denoised social relationship embedding.

[0049] The present application illustrates the denoising condition guiding mechanism in the diffusion model backpropagation process within the CGDLS framework, as shown in Figure 3 . First, the co-interaction item graph between users is extracted from the user-item interaction graph; then, the co-interaction item graph between users is encoded into a low-dimensional embedding in the latent space by SVD; next, the average feature vector of the embedding corresponding to the co-interaction item is calculated; finally, the calculated average feature vector is integrated into the noisy social connection embedding before the denoising step of the diffusion model backpropagation process, thereby guiding the diffusion model backpropagation process to generate a highly task-related social connection embedding.

[0050] In order to extract the key features of the social relationship graph, the CGDLS framework introduces an SVD encoder for encoding the social relationship graph into an embedding in a low-dimensional latent space, thereby better controlling the diffusion model on the social relationship embedding. This process is shown in equation (7):

[0051] (7) wherein, and Represents a bidirectional mapping between the graph data domain and the embedding domain, and is used to encode the SVD encoder for social connection graphs. Corresponding to SVD encoding, and represent the forward and reverse processes of the diffusion model, respectively. Towards Noise is introduced into Used to eliminate The added noise.

[0052] In an exemplary embodiment, the user-item interaction graph is learned by a graph autoencoder to obtain a user-item interaction embedding, which specifically includes: the first step, no noise is added; the second step t- Step 1 to t The noise adding process of the first step is shown in formula (8): (8) in, is the standard normal distribution, For the t The low-dimensional social relationship embedding with added noise obtained by step For the t- The low-dimensional social relationship embedding with added noise obtained in step 1, , I is the identity matrix, For the t The scale parameter of the step, For the t -1 step scale parameter.

[0053] By the additivity of the standard normal distribution, t The low-dimensional social relationship embedding with noise obtained in the first step is simplified to obtain the t The low-dimensional social relationship embedding with added noise is shown in formula (9): (9) in, , , and represents independent Gaussian distributions, To control the t The variance of the Gaussian noise added during step diffusion, is a standard Gaussian distribution, is a Gaussian variable that incorporates multi-step noise, is the residual signal strength of the low-dimensional social relationship embedding after diffusion in the tth step.

[0054] Specifically, since social connections among users can contain a large amount of false connections, i.e., noise. To mitigate the adverse effects of noise on recommendation performance, the present application employs a conditionally guided diffusion model. The conditionally guided diffusion model aims to denoise the social connection embedding and generate highly relevant task social connection embedding.

[0055] In the forward process, Gaussian noise is gradually added to the original social connection embedding , which is gradually transformed into a pure noise embedding. In the reverse process, inspired by the success of various conditionally guided diffusion models, the CGDLS framework of the present application introduces a conditionally guided diffusion model guided by the common interaction item embedding to generate denoised and highly task-relevant social connection embedding. The core idea is that users who interact with more common items have higher similarity and are more likely to establish social connections. Users with social connections are more likely to belong to the same category, and users in the same category often have similar user preferences.

[0056] Therefore, by integrating the embedding of social relationships with the embedding of items that users have interacted with, the present application can better capture the preferences of users in the social network. Based on this, the present application proposes a method that gradually incorporates the embedding of items that users have interacted with in the reverse process of the diffusion model, thereby generating highly relevant social relationship embedding and improving recommendation performance.

[0057] Specifically, in the reverse process, the goal of the present application is to reconstruct the social relationship embedding from the low-dimensional social relationship embedding with added noise. To achieve higher quality reconstruction of the social relationship embedding, the present application encodes the co-interaction graph of items among users using SVD. Subsequently, in the model training phase, the mean value of the encoded embedding is calculated to obtain the mean embedding, which serves as a guide signal in the reverse sampling process. This process involves using a learnable neural network to estimate the following conditional probability, as shown in equation (10):

[0058] (10) wherein and denote a neural network with learnable parameters for estimating a Gaussian distribution.

[0059] The embedding vector at time step is concatenated with the embedding vector corresponding to the time step to form the input. The detailed process is shown in equation (11):

[0060] (11) wherein denotes the node embedding vector in the th diffusion step, denotes an embedding vector obtained by embedding the collaborative interaction item between users into the node embedding of the second diffusion, denotes the embedding of the first time step, denotes two consecutive fully connected layers, in addition, and denote an activation function, a linear transformation matrix and a bias vector, respectively.

[0061] S104: learning the user-item interaction graph by the graph autoencoder to obtain a user-item interaction embedding, and splitting the user-item interaction embedding into a user embedding and an item embedding.

[0062] In an exemplary embodiment, the user-item interaction graph is learned by the graph autoencoder to obtain a user-item interaction embedding, specifically comprising: performing multiple iteration generation operations on the user-item interaction graph to obtain multiple embedding matrices of the user-item interaction graph, wherein the embedding matrix generated by the first iteration generation operation is as shown in formula (12): l (12) wherein, is the embedding matrix generated in the first iteration, l is a degree matrix, D is a normalization function, is a weight matrix, W is an adjacency matrix of the user-item interaction graph, is the embedding matrix generated in the first iteration. l- The embedding matrices generated by multiple iterations are aggregated to obtain a user-item interaction embedding; the formula corresponding to the aggregation operation is as shown in formula (13):

[0063] (13) wherein, is the user-item interaction embedding, is the number of iterations, L is the embedding matrix generated by the first iteration. l

[0064] ​​​​​​​​S105: Integrate the denoised social relationship embedding into the user embedding to obtain a fused user embedding, calculate the inner product between the positive sample of the fused user embedding and the positive sample of the item embedding, and the inner product between the negative sample of the fused user embedding and the negative sample of the item embedding, and subtract the inner product of the positive sample from the inner product of the negative sample. The result of the subtraction is determined as the user's predicted score for the item, and the predicted scores are sorted in order from large to small.

[0065] The calculation formula corresponding to the inner product is shown in formula (14): (14) in, d For the dimension, i For the i item embeddings and the fused user embeddings i elements, , is the positive sample vector of the item embedding, is the negative sample vector of the item embedding, Pred is the inner product.

[0066] To extract key features from the social graph and encode the social graph and user-item interaction graph into a low-dimensional latent space, we use an SVD encoder and a GNN encoder to learn user-user embeddings on the social graph and user-item embeddings on the user-item interaction graph, respectively. These learned embeddings can capture user preferences and thus enhance the prediction of user-item interactions. This process is shown in Equation (15):

[0067] (15) in, represents the social relationship embedding obtained by encoding the social relationship graph through the SVD-based encoding method, and Represents user-item interaction embedding. Social relationship embedding By embedding The forward and reverse processes of the diffusion model are performed on the

[0068] Will (social relationship embedding) and (user-item interaction embedding) aggregated to obtain the combined embedding .

[0069] In an exemplary embodiment, the present invention integrates the denoised social relationship embedding with the user-item interaction embedding to obtain the final embedding for prediction, as shown in formula (16): (16) in, Represents a user The denoising sampling step. By incorporating the predicted value , and jointly minimize the recommendation loss and the loss function of the diffusion model to optimize CGDLS. The loss function is shown in formula (17):

[0070] (17) In the denoising sampling step Diffusion losses are calculated at . Indicates the expected calculation loss using mean square error, indicating go through Step and adopt guided In the parameter The denoised social connection embeddings generated by the neural network are compared with the real data The learning process of CGDLS includes graph encoding, forward process, backward process and loss calculation.

[0071] S106: Determine the items corresponding to the preset number of predicted scores as recommendation results.

[0072] Specifically, the preset number is set according to specific engineering practices.

[0073] The proposed CGDLS method introduces a social connection encoder and a conditional guided diffusion model to improve the performance of social recommendation methods. It uses SVD to encode the social connection graph into a low-dimensional latent space, alleviating node distortion, reducing computational costs, and making the diffusion process more controllable. Conditional guided diffusion is also used to denoise the social connection embeddings while generating more task-relevant social connection embeddings.

[0074] This paper introduces a diffusion model to address the issue of noise in graph data. It adds conditional guided embedding to the diffusion model, guiding the reverse process of the diffusion model toward generating denoised social connection embeddings relevant to the recommendation task. The present invention uses the embedding of commonly interacted items between users as the condition for the conditional probability. This guidance is derived from the user-item interaction graph. Commonly interacted items reflect the consistency of user preferences. Users with similar preferences are more likely to form social connections. Therefore, using this embedding as a conditional guide, the generated social connection embedding is more relevant to the recommendation task.

[0075] The present invention performs a denoising process and a denoising process of a diffusion model on social connection embeddings in a low-dimensional latent space, which not only makes it possible to control the diffusion process using the embedding dimension but also reduces the amount of computation.

[0076] The SVD is adopted to encode the social connection graph, and by extracting key features, the problems of node distortion caused by average aggregation neighbors and the problem of noise amplification in embedding caused by multi-layer propagation are alleviated.

[0077] Compared with the prior art, the user social connection graph encoder is proposed, the key features of the social connection graph are extracted by SVD, and the social connection graph corresponding to the key features is encoded into a social connection embedding in a low-dimensional latent space. This not only reduces the computational cost, but also enhances the controllability of the diffusion process, reduces the noise influence in the encoding process, and reduces the influence of node distortion on the embedding quality. In addition, the conditional guided diffusion model is introduced, which uses the embedding of the common interaction items between users to guide the reverse process of the diffusion model, generates highly task-related social connection embedding, and improves the recommendation performance. The experimental results on three publicly available real-world datasets show that the CGDLS and its core components of the application are superior to the state-of-the-art methods.

[0078] The conditional guided diffusion model is introduced to generate social connection embedding with strong relevance to the recommendation task: the conditional guided diffusion model is introduced in the social recommendation of the application, which effectively denoises the social connection embedding and generates denoised social connection embedding with strong relevance to the recommendation task, enhances the ability to assist recommendation using social connection graph, and improves the performance of social recommendation.

[0079] The SVD encoder is used to encode the social connection graph into an embedding in a low-dimensional latent space: the social connection graph is encoded into a low-dimensional latent space in the application, which is used to enhance the control of the diffusion model and reduce the computational cost; the key features of the social connection graph are extracted by using the SVD encoder, and the embedding corresponding to the key features is used as the encoded social connection embedding, which avoids node distortion and reduces the influence of noise on multi-layer propagation compared with GNN.

[0080] When the social recommendation method based on hidden space conditional diffusion provided by the application is applied, the user can not be required to Figure 1 The order of each step shown in the figure can be determined as required, and the application does not limit the execution order of each step.

[0081] The above is the social recommendation method based on hidden space conditional diffusion provided by one or more embodiments of the application. Based on the same idea, the application also provides a corresponding social recommendation device based on hidden space conditional diffusion, as shown in Figure 4 .

[0082] Figure 4 A kind of item recommendation device based on hidden space conditional diffusion provided by the application is shown in the schematic diagram, comprising: The first dimension reduction module 401 is configured to reduce dimension of a matrix corresponding to the social relationship graph into a low-dimensional social relationship embedding in a hidden space through orthogonal triangular decomposition and singular value decomposition; and the hidden space is a latent space.

[0083] The second dimension reduction module 402 is configured to multiply the matrix corresponding to the user-item interaction graph by a transpose of the matrix corresponding to the user-item interaction graph to obtain a user-common interaction item matrix, and to obtain a low-dimensional user-common interaction item embedding through singular value decomposition of the user-common interaction item matrix.

[0084] The reconstruction module 403 is configured to add noise to the low-dimensional social relationship embedding, and to reconstruct the low-dimensional social relationship embedding with noise added to obtain a denoised social relationship embedding in combination with the low-dimensional user-common interaction item embedding.

[0085] The learning module 404 is configured to learn the user-item interaction graph through a graph autoencoder to obtain a user-item interaction embedding, and to split the user-item interaction embedding into a user embedding and an item embedding.

[0086] The calculation module 405 is configured to integrate the denoised social relationship embedding into the user embedding to obtain a fused user embedding, to calculate inner products between positive samples of the fused user embedding and positive samples of the item embedding, and between negative samples of the fused user embedding and negative samples of the item embedding, respectively, to subtract the inner products of the positive samples from the inner products of the negative samples, to determine a result of the subtraction as a predicted score of the user for the item, and to sort the predicted scores in descending order.

[0087] The determination module 406 is configured to determine items corresponding to a preset number of largest predicted scores as a recommendation result.

[0088] The specific limitations of the item recommendation device based on hidden space conditional diffusion can be referred to the limitations of the item recommendation method based on hidden space conditional diffusion in the above, which will not be described here again. Each module in the above item recommendation device based on hidden space conditional diffusion can be realized by software, hardware and combinations thereof, in whole or in part. The above each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0089] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the above Figure 1 The application provides an item recommendation method based on hidden space conditional diffusion.

[0090] The application further provides an item recommendation device based on hidden space conditional diffusion. Figure 5 The structure of the computer device is shown in FIG. 1.Figure 5 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and of course can also include other hardware required by the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to implement the above-mentioned Figure 1 The method for item recommendation based on hidden space conditional diffusion is provided.

[0091] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments can be included. In each embodiment of the present application, any reference to memory, storage, database or other medium can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0092] Each technical feature of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of each technical feature in the above embodiments are not described, but as long as the combination of these technical features does not exist, it should be considered as the scope of the present application.

Claims

1. An item recommendation method based on latent space conditional diffusion, characterized in that: include: The matrix corresponding to the social relationship graph is reduced to a low-dimensional social relationship embedding in the latent space through orthogonal triangular decomposition and singular value decomposition; The hidden space is a latent space; Determine the inter-user common interaction item matrix by multiplying the matrix corresponding to the user-item interaction graph by the transpose of the matrix corresponding to the user-item interaction graph, and perform singular value decomposition on the inter-user common interaction item matrix to obtain a low-dimensional inter-user common interaction item embedding; Adding noise to the low-dimensional social relationship embedding, and combining it with the low-dimensional embedding of commonly interacted items between users, reconstructing it from the low-dimensional social relationship embedding with the added noise to obtain a denoised social relationship embedding; Learning the user-item interaction graph through a graph autoencoder to obtain a user-item interaction embedding, and splitting the user-item interaction embedding into a user embedding and an item embedding; Integrate the denoised social relationship embedding into the user embedding to obtain a fused user embedding, calculate the inner product between the positive sample of the fused user embedding and the positive sample of the item embedding, and the inner product between the negative sample of the fused user embedding and the negative sample of the item embedding, and subtract the inner product of the positive sample from the inner product of the negative sample. The result of the subtraction is determined as the predicted score of the user for the item, and the predicted scores are sorted in descending order; The items corresponding to the preset number of prediction scores are determined as recommendation results.

2. The method according to claim 1, wherein The learning of the user-item interaction graph by the graph autoencoder to obtain the user-item interaction embedding specifically includes: Perform multiple iterative generation operations on the user-item interaction graph to obtain multiple embedding matrices of the user-item interaction graph, where the first l The embedding matrix generated by the iterative generation operation is: ； in, For the l The embedding matrix generated in the iterations is D is the degree matrix, is the normalization function, W is the weight matrix, Interaction diagram for user projects The adjacency matrix of For the l- The embedding matrix generated in 1 iteration; The embedding matrices generated by multiple iterations are aggregated to obtain the user-item interaction embedding; the formula corresponding to the aggregation operation is: ； in, Embed for user project interaction, L is the number of iterations, For the l The embedding matrix generated by the iteration.

3. The method according to claim 1, wherein The matrix corresponding to the social relationship graph is reduced to a low-dimensional social relationship embedding in the latent space by orthogonal triangular decomposition and singular value decomposition, specifically including: The matrix corresponding to the social relationship graph Multiplying the random matrix to obtain a first matrix; the first matrix is ​​used to project the column space of the matrix corresponding to the social relationship graph into the low-dimensional subspace spanned by the random matrix; Based on the first matrix, performing multiple power iteration operations to obtain a power iteration generation result; Decompose the result generated by the last power iteration using the orthogonal triangular decomposition method to obtain an orthogonal matrix and an upper triangular matrix; Multiply the transpose of the orthogonal matrix by the matrix corresponding to the social relationship graph to obtain a second matrix; Calculate the second matrix and perform singular value decomposition, and keep the first k principal components; The previous k principal components are determined as the low-dimensional social relationship embedding.

4. The method according to claim 1, wherein The adding noise to the low-dimensional social relationship embedding specifically includes: In step 0, no noise is added; No. t- Step 1 to t The noise adding process of the first step is: ; in, is the standard normal distribution, For the t The low-dimensional social relationship embedding with added noise obtained by step For the t- The low-dimensional social relationship embedding with added noise obtained in step 1, , I is the identity matrix, For the t The scale parameter of the step, For the t -1 step scale parameter; By the additivity of the standard normal distribution, t The low-dimensional social relationship embedding with noise obtained in the first step is simplified to obtain the t The low-dimensional social relationship embedding with added noise is: ； in, , , and represents independent Gaussian distributions, To control the t The variance of the Gaussian noise added during step diffusion, is a standard Gaussian distribution, is a Gaussian variable that incorporates multi-step noise, Embed the low-dimensional social relationship in the t The residual signal intensity after the first diffusion step.

5. The method according to claim 1, wherein The corresponding formula for the reconstruction is: ； in, and Represents a learnable parameter neural network, t is the time step, z To guide embedding, represents the standard normal distribution, For the t- Node embedding vector in 1 diffusion step, For the t The low-dimensional social relationship embedding with added noise obtained by step For the t- The low-dimensional social relationship embedding with added noise obtained in step 1; the guided embedding is the encoding result of the low-dimensional embedding of the common interaction items between users.

6. The method according to claim 1, wherein The process of constructing the user-project interaction graph specifically includes: Building User Sets and project sets ; Through the matrix Represents the interaction between users and items, matrix R For each element in Represents a user With the project There is interaction, Represents a user With the project No interaction; Based on the user set, the item set and the matrix R , constructing a collaborative bipartite graph , and determining the collaboration bipartite graph as the user-item interaction graph.

7. The method according to claim 6, wherein The process of constructing the social relationship graph specifically includes: Build a social relationship matrix between users ,matrix S Each element Represents a user and users There are social connections between Represents a user and users There is no social connection between them; Based on the user set and the social relationship matrix, the social relationship graph is constructed. .

Images