A user-item recommendation method, system, computer device and storage medium

Abstract

The application provides a user-item recommendation method and system, computer equipment and a storage medium, belonging to the field of data recommendation, comprising obtaining user and item interaction data, constructing a user-item interaction graph from the user and item interaction data; reparameterization sampling of the user-item interaction graph, obtaining latent semantic representation and reconstructing to obtain an enhanced graph structure; extracting node embedding features from the enhanced graph structure; determining the recommendation score of the item relative to the user according to the inner product of the user embedding and the item embedding in the node embedding features; sorting multiple items according to the recommendation score of the multiple items relative to the user, and determining the item recommendation result of the user based on the sorting result. The method generates latent semantic representation by reparameterization sampling from the learned node distribution, and then reconstructs it into an enhanced graph structure, thereby supplementing the sparse interaction relationship while maintaining the true similarity between the user and the item, and further ensuring the reliability of the recommendation score.

Description

Technical Field

[0001] This invention belongs to the field of data recommendation technology, specifically relating to a user-item recommendation method, system, computer equipment, and storage medium. Background Technology

[0002] User preference calculation is the foundation and input of the item recommendation process. By quantifying users' potential interest in items, it provides a basis for decision-making. With the development of Graph Neural Networks (GNNs), researchers have used graph convolution operations to aggregate high-order neighbor information, such as NGCF and LightGCN models, which have improved item recommendation performance to some extent. However, GNN-based collaborative filtering methods usually rely heavily on sufficient supervision signals, but real-world user-item interaction data is generally implicit feedback and highly sparsity, making it difficult for models to learn high-quality node representations, and the effectiveness of user-oriented item recommendation remains limited.

[0003] Building upon this, contrastive learning (CL) offers an effective approach to alleviate data sparsity. Researchers have combined graph neural networks (GNNs) with contrastive learning, constructing multiple contrastive views through data augmentation and maximizing mutual information to enhance consistency across these views, thereby mitigating data sparsity in recommender systems. However, existing methods combining GNNs and contrastive learning often introduce unreasonable perturbations into data augmentation. For example, structural augmentation by discarding edges or nodes can fail to accurately reflect the true similarity between users and items, thus affecting the accuracy of recommendations. Similarly, feature augmentation using uniform noise can introduce redundant noise, generating meaningless views that increase the learning difficulty of the model and also impact the accuracy of recommendations. Summary of the Invention

[0004] To address the problem that data augmentation in existing project recommendation methods introduces unreasonable disturbances, leading to a decrease in recommendation accuracy, this invention provides a user-project recommendation method, system, computer device, and storage medium.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A user-item recommendation method, comprising: Obtain the interaction data between the users to be recommended and the projects, and construct the interaction graph between the users to be recommended and the projects based on the interaction data; Based on the preset node distribution parameters, the interaction graph of the user-item to be recommended is reparameterized and sampled to obtain the latent semantic representation. The latent semantic representation is then reconstructed to obtain the augmented graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distribution of user nodes and item nodes in the interaction graph of the user-item to be recommended. The embedding features of the user to be recommended and multiple item embedding features are extracted from the augmented graph structure. The recommendation scores of multiple items relative to the user to be recommended are determined based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; the multiple items are ranked according to their recommendation scores relative to the user to be recommended, and the item recommendation results for the user to be recommended are determined based on the ranking results.

[0006] Optionally, in a user-item recommendation method provided by the present invention, the user-item interaction graph to be recommended is processed by the GGDCL model to obtain recommendation scores of multiple items relative to the user to be recommended; wherein, the GGDCL model includes a variational graph generation module, a dual-graph encoder contrastive learning module and a prediction module connected in sequence, the variational graph generation module is constructed based on a variational autoencoder, and the dual-graph encoder contrastive learning module includes a master encoder; The variational graph generation module reparameterizes each node in the user-item interaction graph to be recommended, based on its node distribution parameters, to obtain the latent semantic representation of the node. The semantic similarity between multiple latent semantic representations is calculated, and the enhanced graph structure is reconstructed from the semantic similarity between each node. Node embedding features are extracted from the augmented graph structure using the main encoder; The recommendation score is determined by the prediction module from the inner product of the user embedding to be recommended and multiple item embeddings in the node embedding features.

[0007] Optionally, the user-item recommendation method provided by the present invention further includes: Construct a node set based on the user set and item set in the interaction data between the users and items to be recommended; Construct a user-project interaction matrix based on the node set and the interaction records in the user-project interaction data; Construct the adjacency matrix of the bipartite graph based on the user-item interaction matrix, and determine the user-item interaction graph to be recommended by the node set and the adjacency matrix of the bipartite graph.

[0008] Optionally, the dual-image encoder contrastive learning module also includes a twin encoder running in parallel with the main encoder. The parameters of the twin encoder are determined by the current parameters of the main encoder based on a momentum update mechanism. The user-item recommendation method provided by this invention also includes: Obtain the interaction data between training users and projects, and construct a training user-project interaction graph from the training user-project interaction data; perform semantic reconstruction on the training user-project interaction graph to obtain the training augmented graph structure; The GGDCL model is trained by extracting main view node embedding features from the augmented graph structure using a master encoder and a twin view node embedding feature from the augmented graph structure using a twin encoder. Alignment loss and uniformity loss are calculated based on the main view node embedding features, and contrastive learning loss is calculated based on the main view node embedding features and the twin view node embedding features. The GGDCL model is then optimized based on the alignment loss, uniformity loss, and contrastive learning loss to obtain the trained GGDCL model.

[0009] Optionally, the main view node embedding features include user node embedding features and item node embedding features. The user-item recommendation method provided by this invention further includes: The alignment loss is calculated based on the Euclidean distance between the user node embedding features and the item node embedding features; The uniformity loss is calculated based on the negative log-expected distance of the user node embedding feature distribution and the negative log-expected distance of the item node embedding feature distribution.

[0010] Optionally, the user-item recommendation method provided by the present invention further includes: Calculate the variational graph generation loss based on the user-item interaction graph used for training and the augmentation graph structure used for training. The total loss of the GGDCL model is determined by variational graph generation loss, alignment loss, uniformity loss, and contrastive learning loss. The parameters of the GGDCL model are optimized based on the total loss to obtain the trained GGDCL model.

[0011] Optionally, the user-item recommendation method provided by the present invention further includes: The current parameters of the twin encoder are obtained by updating the momentum based on the current parameters of the main encoder; The current parameters of the master encoder are updated by backpropagation based on the total loss.

[0012] The present invention also provides a user-item recommendation system, comprising: The interaction graph construction module is used to obtain the interaction data between the users to be recommended and the projects, and to construct the interaction graph between the users to be recommended and the projects. The graph structure enhancement module is used to reparameterize the user-item interaction graph to be recommended according to preset node distribution parameters to obtain a latent semantic representation. The latent semantic representation is then reconstructed to obtain an enhanced graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distribution of user nodes and item nodes in the user-item interaction graph to be recommended. The module also extracts embedded features of the user to be recommended and multiple item embedded features from the enhanced graph structure. The recommendation result determination module is used to determine the recommendation scores of multiple items relative to the user to be recommended based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; to rank the multiple items according to the recommendation scores of the multiple items relative to the user to be recommended, and to determine the item recommendation results for the user to be recommended based on the ranking results.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of a user-item recommendation method.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, is capable of executing any step of a user-item recommendation method.

[0015] The user-item recommendation method provided by this invention has the following beneficial effects: The user-item recommendation method provided by this invention reparameterizes the user-item interaction graph to be recommended using preset node distribution parameters to generate a latent semantic representation, then reconstructs it into an enhanced graph structure, and finally calculates the recommendation score based on the enhanced graph structure to generate the user's item recommendation result.

[0016] This invention employs a generative enhancement mechanism to obtain a semantically consistent and informationally complete enhanced graph structure. While supplementing sparse interaction relationships, it maintains the genuine similarity between users and items, providing a high-quality foundation for subsequent feature extraction and avoiding the introduction of unreasonable perturbations and redundant noise. The node embedding features extracted based on this enhanced graph structure can more accurately reflect the essential relationship between users and items, thus making the recommendation score calculated through inner product more reliable. Ultimately, the item recommendation suggestions generated based on the ranking results show a significant improvement in accuracy, effectively solving the problem of decreased recommendation performance caused by inappropriate enhancement strategies in existing technologies. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a user-item recommendation method provided in an embodiment of the present invention; Figure 2 This is an example of a variational graph generation architecture in the GGDCL model provided in this embodiment of the invention; Figure 3 This is an example of dual-graph comparison learning in the GGDCL model provided in this embodiment of the invention; Figure 4 Examples of optimization and recommendation in the GGDCL model provided in this embodiment of the invention; Figure 5 This is an example of the impact of contrastive learning weights on the performance of a model based on the Yelp 2018 dataset, provided in an embodiment of the present invention. Figure 6 An example of the impact of contrastive learning weights on the performance of a model based on the Amazon-Book dataset, provided as an embodiment of the present invention; Figure 7 This is an example of the impact of contrastive learning weights on the performance of a model based on the Alibaba-ifashion dataset, provided in an embodiment of the present invention. Figure 8 This is an example of the impact of momentum coefficient on model performance based on the Yelp 2018 dataset, provided in an embodiment of the present invention. Figure 9 This is an example of the impact of the momentum coefficient on the performance of a model based on the Amazon-Book dataset, provided as an embodiment of the present invention. Figure 10 This is an example of the impact of the momentum coefficient on the performance of a model based on the Alibaba-ifashion dataset, provided in an embodiment of the present invention. Figure 11 An example of the impact of uniformity loss weights on the performance of a model based on the Yelp 2018 dataset, provided as an embodiment of the present invention; Figure 12 An example of the impact of uniformity loss weights on the performance of a model based on the Amazon-Book dataset, provided for embodiments of the present invention; Figure 13 This is an example of the impact of uniformity loss weights on the performance of a model based on the Alibaba-ifashion dataset, provided as an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0020] Early techniques used collaborative filtering to provide recommended items to users through graph neural network-based recommendation systems. For example, NGCF extracts higher-order connections between users and items by propagating node embeddings in the graph; LightGCN simplifies NGCF by removing nonlinear activations and feature transformations; UltraGCN further simplifies LightGCN by directly approximating the limit of infinite-layer graph convolutions using a constrained loss, without explicit message passing. CAGCN proposes a topology metric for recommendation tasks, aiming to improve the accuracy of link prediction. PolyCF proposes a flexible graph signal processing model based on multinomial graph filters. However, due to insufficient graph structural information caused by sparse interactions, graph-based collaborative filtering methods struggle to learn high-quality node representations through message passing.

[0021] To address the issue of data sparsity, existing technologies utilize graph-based contrastive learning-based recommendation systems to effectively alleviate this problem and better model user interests. For example, SGL pioneered the introduction of contrastive learning into recommendation systems, proposing three suitable operators: random edge discarding, random node discarding, and random walks. SIMGCL generates contrast pairs by adding perturbations. While these random augmentation strategies are simple and effective, unreasonable perturbations can disrupt the true semantic relationships between users and items. To construct more semantically meaningful contrast pairs, NCL explores building contrast pairs between structural neighbors and semantic neighbors. CGCL investigates the semantic associations between users and candidate items in different layers of embedding representations and constructs contrast pairs through semantically similar embeddings. ADAMCL fuses embedding information learned from user-item graphs and user-user graphs, using the fused higher-order information to construct contrast pairs with the original information. REDCL proposes a dual contrastive learning task based on graph structure and feature space. Furthermore, DIMCL proposes methods to eliminate redundancy, false positives, and irrelevant noise in the dimensional representations of contrastive learning and recommendation systems. However, the above methods are all limited to transforming or mining the original observation data, and fail to fundamentally complete or enhance the graph structure to provide richer self-supervised signals.

[0022] Furthermore, existing technologies also combine Variational Auto-Encoders (VAEs) with collaborative filtering for item recommendation. For example, Mult-VAE proposes a multinomial distribution suitable for modeling user-item interaction data and enhances representational power by parameterizing user representations through neural networks. CVGA combines graph neural networks with VAEs to propose a collaborative graph autoencoder recommendation method, reconstructing the user-item bipartite graph through variational inference. DR-VAE alleviates exposure bias by introducing a debiased estimator and uses a flow model-based representation enhancement module to directly fit complex posterior distributions to address the posterior collapse problem; however, this approach focuses on subsequent ratings, and noise interference with the model still exists.

[0023] In summary, existing graph-based collaborative filtering methods, at the data augmentation strategy level, suffer from the problem that the addition of noise leads to meaningless views, increasing the difficulty of model learning. Furthermore, at the optimization objective level, existing collaborative filtering methods typically jointly optimize the Bayesian Personalized Ranking (BPR) loss and the contrastive learning loss. The BPR loss heavily relies on the quality of negative sampling and is susceptible to popularity bias. Moreover, this joint optimization scheme fails to fully exploit the potential value of representation alignment and uniformity, making it difficult to ensure that positive samples are closely packed while negative samples are sufficiently dispersed on the hypersphere.

[0024] To address the aforementioned shortcomings, this invention provides a user-item recommendation method that further improves upon collaborative filtering, abandoning traditional graph structure or random embedding space augmentation. First, a Variational Graph Auto-Encoder (VGAE) is employed to reconstruct and complete the original user-item graph. This module learns the probability distribution of node embeddings and generates augmented graphs with richer semantic information and more complete structure, thus providing a high-quality foundation for downstream contrastive learning. Second, a dual-graph encoder is introduced, comprising a master encoder and a momentum-updated Siamese encoder. The Siamese encoder smoothly tracks the parameter evolution of the master encoder through a momentum mechanism, naturally generating stable and diverse contrastive views without data augmentation, effectively avoiding semantic damage caused by inappropriate augmentation strategies. Furthermore, by introducing alignment and uniformity losses, the recommendation task is directly optimized without negative sampling. By optimizing the alignment and uniformity of user and item embeddings, the model accelerates convergence while improving its resistance to overfitting.

[0025] Example 1 This application provides a user-item recommendation method, specifically as follows: Figure 1 As shown, it includes the following steps: Step 11: Obtain the interaction data between the user and the items to be recommended, and construct a user-item interaction graph from this data. Specifically, the user-item interaction graph is processed using the GGDCL model to obtain recommendation scores for multiple items relative to the user. The GGDCL model includes a variational graph generation module, a dual-graph encoder contrastive learning module, and a prediction module connected in sequence. The variational graph generation module is built based on a variational autoencoder, and the dual-graph encoder contrastive learning module includes a master encoder. The variational graph generation module performs semantic reconstruction on the user-item interaction graph to obtain an enhanced graph structure. The master encoder extracts node embedding features from the enhanced graph structure. The prediction module determines the recommendation score by the inner product of the user embedding and multiple item embeddings in the node embedding features.

[0026] The interaction diagram of the user-project to be recommended can be constructed through the following steps: Step 111: Construct a node set based on the user set and item set in the interaction data between the users and items to be recommended.

[0027] Step 112: Construct a user-project interaction matrix based on the node set and the interaction records in the user-project interaction data.

[0028] Step 113: Construct the adjacency matrix of the bipartite graph based on the user-item interaction matrix. The user-item interaction graph to be recommended is determined by the node set and the adjacency matrix of the bipartite graph.

[0029] Specifically, in collaborative filtering, there are two types of entities: user set U and item set V. For recommendation scenarios with implicit feedback, this embodiment uses a matrix... Describe the user-project interaction relationship, where M represents the number of users and N represents the number of projects. When a user and a project interact, the corresponding... =1, otherwise =0, thus modeling the data as a user-item bipartite graph based on graph collaborative filtering methods. ,in The adjacency matrix represents the set of nodes. The definition is shown in formula (1): (1) Then, based on the given initialization node embedding The graph-based collaborative filtering method updates the node representation through multi-layer graph convolution operations as shown in formula (2): (2) Where D is the diagram The degree of a matrix and They represent the first and the Node embedding in layer graph convolution. After stacking L layers of graph convolution operations, the final node representation... Through read operation ( The result is obtained as shown in formula (3): (3) Step 12: Obtain the interaction data between the training users and the projects, and construct the training user-project interaction graph from the interaction data; perform semantic reconstruction on the training user-project interaction graph to obtain the training augmented graph structure.

[0030] Specifically, once the user-project interaction graph or user-project bipartite graph is obtained, it can be obtained through methods such as... Figure 2 The GGDCL model shown calculates the relative user recommendation score for each item. The GGDCL model consists of a variational graph generation module and a dual-graph encoder contrastive learning module. The variational graph generation module augments the original user-item graph to generate rich graph structure data. The dual-graph contrastive learning module uses a pair of GCN encoders—a master encoder and a momentum-updated twin encoder—to generate node embeddings for the two views respectively, and performs contrastive learning using InfoNCE loss to obtain more robust representations. During training, the GGDCL model introduces alignment and uniformity losses to create a synergistic effect with the dual-graph encoder: for the dual encoder, alignment and uniformity optimization helps the master encoder obtain better parameters. The entire framework organically combines generative augmentation and contrastive learning to jointly optimize the final recommendation task.

[0031] For example, in the variational graph generation module, a graph neural network (GNN) can be used as an encoder to infer the probability distribution parameters of user and item nodes in the latent space, namely the mean and variance or standard deviation (std). Then, the enhanced graph structure can be sampled and reconstructed using this distribution. Given the observed user-item interaction graph and initial node embedding The goal of graph generation is to learn a probability distribution Z that can reconstruct the structure of the input graph: Variational inference can be used during the model training phase. To approximate the true posterior To better utilize the higher-order information of the user-item graph structure, a GNN is used to estimate the parameters of the potential distribution of node i, as shown in formula (4): (4) in, and These are two learnable parameters. Let be the mean of the potential distribution of node i. Let be the standard deviation in the latent distribution of node i. The above graph generation process can also be implemented using LightGCN as the encoder. Specifically, for node i∈U∪V, the mean vector update method of its l-th layer is shown in formula (5): (5) in, and Let represent the mean vectors of node i in the l-th and (l-1)-th graph convolutional layers, respectively. and Let represent the sets of connected neighbors of node i and node j, respectively. After stacking L layers of graph convolutional layers, L+1 outputs are obtained. The outputs of all layers are fused to obtain the final mean, and the variance is calculated based on the mean, as shown in formula (6): (6) The variance is learned through a multilayer perceptron. Alternatively, using a single-layer MLP can achieve optimal performance, i.e., σ = exp(μW + b), where W ∈ and b∈ These are two learnable parameters. After obtaining approximate mean and variance or standard deviation, this is achieved by analyzing the distribution... Mid-sampling to generate the latent representation of node i Considering that the sampling process is non-differentiable, this method cannot be directly optimized. In this embodiment, a reparameterization technique is used to replace the original sampling process, as shown in formula (7): (7) in, It is standard Gaussian noise. After obtaining the probability distribution of the latent variable Z, the original user item graph representation is reconstructed as shown in Equation (8): (8) in, This is the probability matrix representation of the generated new graph. At this point, the probability of each edge existing in the generated graph is independent, and modeling is performed conditioned on latent variables. Then, the link prediction scores for nodes i and j are calculated using the inner product method, as shown in formula (9): (9) in, It's the sigmoid function. During model training, the loss function for graph generation loss is... The calculation is shown in formula (10): (10) in, These are the decoder parameters. These are the encoder parameters. It is the reconstruction error between the original image and the generated image. It is the KL (Kullback-Leibler) divergence between the node embedding distribution and the standard Gaussian distribution.

[0032] Step 13: Extract the main view node embedding features from the training augmented graph structure using the master encoder, and extract the twin view node embedding features from the augmented graph structure using the twin encoder; calculate the alignment loss and uniformity loss based on the main view node embedding features, and calculate the contrastive learning loss based on the main view node embedding features and the twin view node embedding features; optimize the parameters of the GGDCL model based on the alignment loss, uniformity loss, and contrastive learning loss to obtain the trained GGDCL model.

[0033] The main view node embedding features include user node embedding features and item node embedding features. The alignment loss and uniformity loss calculations include: Step 131: Calculate the alignment loss based on the Euclidean distance between the user node embedding features and the item node embedding features.

[0034] Step 132: Calculate the uniformity loss based on the negative log-expected distance of the user node embedding feature distribution and the negative log-expected distance of the project node embedding feature distribution.

[0035] The loss function can be calculated using the following steps: Step 133: Calculate the variational graph generation loss based on the user-item interaction graph used for training and the augmentation graph structure used for training.

[0036] Step 134: Determine the total loss of the GGDCL model from the variational graph generation loss, alignment loss, uniformity loss, and contrastive learning loss.

[0037] Step 135: Optimize the parameters of the GGDCL model based on the total loss to obtain the trained GGDCL model. For example, update the current parameters of the twin encoder based on the momentum of the current parameters of the master encoder; and update the current parameters of the master encoder through backpropagation based on the total loss.

[0038] Specifically, such as Figure 3 As shown, after generating the enhanced graph structure based on the variational graph autoencoder, this invention performs contrastive learning based on a dual-graph encoder framework. The dual-graph encoder framework achieves dynamic contrastive learning without compromising semantics through the collaborative work of two encoders, specifically including a main graph encoder. and momentum update twin encoder The main encoder Represented as parameters Twin encoder The parameters are expressed as This corresponds to the learnable user and item embedding representations within the encoder. The two encoders achieve dynamic optimization through a collaborative mechanism: the twin encoder, based on parameter updates from the main encoder, adaptively adjusts the generation strategy of the comparison view through a momentum update mechanism.

[0039] Among them, the main encoder Parameters are updated via conventional backpropagation. Siamese encoder. This aims to provide a stable contrast target, whose parameter φ tracks the evolution of the master encoder parameter θ through a momentum update mechanism. The parameters of the twin encoder at step t are... The update mechanism is shown in formula (11): (11) in, φ is the momentum coefficient. During model training, the parameters θ of the master encoder are updated through backpropagation, while the parameters φ of the twin encoder are copied from θ by momentum updates, ignoring gradient updates during the update process. For example, the expanded representation of the twin encoder parameters at the nth iteration is shown in Equation (12): (12) As shown in formula (12), the parameters of the twin encoder at step n are... In reality, it is an exponentially weighted moving average of the historical parameters of the main encoder. Among them, The term represents the weight of the historical parameter θ(t−1) decaying over time: the further away the parameter is from the current step n, the smaller its weight.

[0040] The main encoder and its twin encoder Extracting node-level representations using graph convolution respectively and This allows for comparative learning, and the specific graph convolution operation can also be performed using LightGCN as the encoder. (Main Encoder) Graph convolution is performed using the current parameters θ updated via backpropagation to generate the user's embedding. Embedding of projects As shown in formulas (13) and (14): (13) (14) Where A is the user-item normalized adjacency matrix reconstructed by the variational graph autoencoder. Main encoder The initial node embedding matrix, The sequence of node embedding matrices output from the Lth layer graph convolution of the master encoder. The twin encoder... Use its momentum update parameters Generate another set of user embeddings Embedding of projects As shown in formulas (15) and (16): (15) (16) like Figure 4 As shown, after obtaining the representations of the master encoder and the twin encoder, in order to maximize the performance of the dual encoder... and The similarity between the generated pairs of representations, while encouraging consistency in the embedding space, is used for model updates through contrastive learning, where the user-specific InfoNCE loss formula is used. As shown in formula (17): (17) in, and It is user expression, respectively by the main encoder and its twin encoder generate, It is a comparison of temperature over-parameters. It's the training batch, the main encoder. and its twin encoder InfoNCE loss formula for projects This can be calculated and determined in a similar manner. Finally, the InfoNCE loss for the user is added to the InfoNCE loss for the project to obtain the master encoder. and its twin encoder Contrast loss As shown in formula (18): (18) Furthermore, considering that alignment and uniformity are two key attributes closely related to representation quality, this invention also requires the calculation of alignment and uniformity losses during the model training phase. Specifically, the alignment loss... Semantic aggregation is achieved by increasing the similarity of positively correlated user-item pairs, while uniformity loss is used. The degree of dispersion on the hypersphere is then quantified, as shown in formulas (19) and (20) in the dual encoder architecture of this invention: (19) (20) in, This represents the distribution of positive sample pairs embedded in user and item data. Represents the distribution of user embeddings. Indicates the distribution of project embeddings. This is the embedding of the x-th item node obtained by random sampling from the item embedding distribution. The embedding of the y-th item node is obtained by randomly sampling from the item embedding distribution. For the embedding of the x-th user node obtained by random sampling from the user embedding distribution, Let y be the embedding of the y-th user node randomly sampled from the user embedding distribution. Then, based on the above four types of losses, the total loss of the model is determined, and the hyperparameters of the model are optimized, as shown in formula (21): (twenty one) in, To control the hyperparameters of uniformity loss, To control the hyperparameters of the contrastive learning loss, To generate hyperparameters for the loss in the control chart, To control The hyperparameter of regularization strength, These are the model parameters.

[0041] Step 14: Reparameterize the user-item interaction graph to be recommended according to the preset node distribution parameters to obtain the latent semantic representation. Reconstruct the latent semantic representation to obtain the enhanced graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distribution of user nodes and item nodes in the user-item interaction graph to be recommended. Extract the embedded features of the user to be recommended and multiple item embedded features from the enhanced graph structure.

[0042] Step 15: Determine the recommendation scores of multiple items relative to the user to be recommended based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; rank the multiple items according to the recommendation scores of multiple items relative to the user to be recommended, and determine the item recommendation results for the user to be recommended based on the ranking results.

[0043] Specifically, when the user embedding is determined through the main encoder and project embedding Then, it is embedded by the user. and project embedding The inner product is used to calculate the predicted score for the user based on the project. As shown in formula (22): (twenty two) in, and All originated from the main encoder The final output is as follows: When there are multiple items, the predicted score of each item for the user to be recommended is determined by the inner product of the embedding of the user to be recommended and the embeddings of multiple different items. Then, these predicted scores are sorted in order of priority, and the items with the highest predicted scores are recommended to the user to be recommended.

[0044] Example 2 Building upon Example 1, this example compares and verifies the GGDCL model provided by this invention with existing methods based on DNN, GNN, and GCL using three publicly available datasets to determine the performance of the GGDCL model. The three publicly available datasets are Yelp2018, Amazon-Book, and Alibaba-ifashion, three public benchmark datasets used in e-commerce solutions. The Yelp2018 dataset, derived from publicly available data from the Yelp platform in 2018, is a commonly used benchmark in the field of local life service recommendations. Its interaction data consists of user reviews of restaurants, hotels, and other businesses. Compared to the other two datasets, Yelp2018 interactions are typically sparser, effectively testing the model's performance in data-scarce scenarios. The Amazon-Book dataset, a large subset of Amazon product review data, focuses on books and contains rich user historical behavior data. It is one of the standard benchmarks for evaluating the accuracy of recommendation systems on large e-commerce platforms, with relatively high interaction density, helping to test the model's ability to learn effective patterns from dense interactions. The Alibaba-iFashion dataset is a fashion recommendation dataset from real-world industrial scenarios, released by Alibaba Group. Unlike the previous two datasets based on explicit ratings, iFashion primarily includes implicit feedback such as user clicks, making it closer to the practical applications of modern online recommendation systems. For example, using Yelp 2018 as a training user, a user node might represent a diner named "Zhang San," and an item node might be "a Sichuan restaurant." If Zhang San has written a review for the restaurant, there is an edge in the interaction graph. In the Amazon-Book training user dataset, the edge between the user node "Li Si" and the item node "The Three-Body Problem" indicates that Li Si has purchased or reviewed the book. In the Alibaba-iFashion training user dataset, the edge between the user node "Wang Wu" and the item node "a pair of sneakers" indicates that Wang Wu has clicked or viewed the product details page.

[0045] The baseline models for the experiments include: the BPRMF model, which uses Bayesian ranking loss to optimize matrix factorization for personalized ranking predictions; the NGCF model, which propagates embeddings in the user-item interaction graph to establish higher-order connections between nodes; the LightGCN model, a simplified graph neural network that removes nonlinear variations and activation functions; the DirectAU model, which abandons the Bayesian Personalized Ranking (BPR) loss and optimizes representation quality by directly improving the alignment and uniformity of representations on the hypersphere; the CVGA model, which learns user-item collaborative relationships through information propagation and aggregation paradigms to estimate parameters for inferring the probability distribution of user behavior, rather than directly learning user or item embeddings; the SGL graph contrastive learning method, which performs data augmentation by discarding nodes or edges in the graph structure; the NCL method, which utilizes prototype contrastive targets and integrates structural and semantic neighbor relationships; and the GLSCL framework, which enhances the uniformity of embeddings and constructs contrast pairs based on local similarity.

[0046] In this embodiment, PyTorch is used to implement the model and all baselines. Adam is selected as the model optimizer, with a learning rate of 0.001, a regularization coefficient of 0.0001, a mini-batch size of 1024, and an embedding size of 64. Furthermore, the hyperparameters controlling the uniformity loss are set to a range of values. The range of values ​​for the hyperparameters controlling the weights of the contrastive learning loss is... The hyperparameter value for the loss weight in the control chart generation is fixed at 0.1. Empirically, the temperature coefficient hyperparameter of InfoNCE is set to... =0.2. Furthermore, Recall@k and NDCG@k, commonly used in collaborative filtering models based on graph convolutional networks, were chosen as evaluation metrics. Recall@k represents the coverage rate of correctly recommended items, serving as the result of top-K recommendations; NDCG@k is the normalized depreciation cumulative gain, a measure of ranking quality. For each metric, the performance of each model was calculated based on the top 20 results.

[0047] First, the overall performance of the GGDCL model is compared with the baseline model mentioned above on three datasets, as shown in Table 1: Table 1. Overall performance comparison of the GGDCL model with baselines based on DNN, GNN, and GCL.

[0048] Among these, the GGDCL model used in this invention achieved the best performance on all three datasets, surpassing state-of-the-art methods based on MF, GNN, and GCL. Specifically, compared to GNN-based methods such as NGCF and LightGCN, traditional methods such as BPRMF performed the worst. This indicates that graph neural networks can effectively capture high-order relationships between user items. CVGA outperformed LightGCN, demonstrating that completing and enhancing the original sparse graph structure through generative models can provide richer and more robust supervision signals for recommendation models. DirectAU outperformed LightGCN, indicating that using alignment and uniformity losses can effectively enhance the representation quality in collaborative filtering. GCL-based methods such as SGL, NCL, and GLSCL significantly improved the performance of LightGCN on the three datasets, demonstrating that introducing self-supervised contrastive learning as an auxiliary task can effectively alleviate the data sparsity problem and learn more robust representations. The GGDGCL model used in this invention achieves the best performance compared to all baselines because the GGDGCL model completes the graph structure through the variational graph generation module and uses a momentum-update-based dual encoder architecture for comparative learning. At the same time, it uses alignment and uniformity as the core optimization objectives, which solves the problems of data sparsity, unreasonable perturbations caused by random data augmentation, and negative sampling bias in BPR loss.

[0049] Subsequently, ablation experiments were conducted on the GGDGCL model provided by this invention to analyze the contributions of the variational graph generation module and the dual-graph encoder contrastive learning module to the model, as shown in Table 2: Table 2 Comparison of Ablation Experiment Results

[0050] In this table, GGDCL / GG represents the removal of the variational graph generation module, and GGDCL / D represents the removal of the dual-graph encoder contrastive learning module. As shown in Table 2, GGDCL / GG outperforms LightGCN. This means that even without the variational graph generation module, the proposed dual-encoder contrastive learning framework learns more robust user and item representations than traditional supervised GCNs through effective view comparison. This demonstrates that the momentum-update-based contrastive learning mechanism can effectively enrich data representations, thereby improving recommendation performance. The GGDCL / D variant also surpasses LightGCN on all metrics, indicating that by reconstructing and completing the original graph structure, the graph neural network can learn based on a more semantically complete graph structure, generating higher-quality node embeddings. The GGDCL model achieves the best performance, exceeding all variant models, proving a significant synergistic effect between the variational graph generation module and the dual-graph encoder contrastive learning module. The two are not simply additive but complementary: variational graph generation provides a high-quality enhanced graph structure, laying a solid foundation for contrastive learning; while dual-encoder contrastive learning can fully utilize the structural information of the generated graph to further optimize node representations.

[0051] Furthermore, this invention also experiments with the hyperparameters of the model training to determine the weights of the contrastive learning loss. Momentum coefficient Impact and uniformity loss weights Impact on model performance.

[0052] Among them, the contrastive learning loss weights The impact on the model is as follows Figure 5 , Figure 6 and Figure 7 As shown, where Figure 5 Weights for different contrastive learning loss patterns when training a model based on the Yelp 2018 dataset. Impact on the model results Figure 6 Weights for different contrastive learning loss when training a model based on the Amazon-Book dataset Impact on the model results Figure 7 Different contrastive learning loss weights when training a model based on the Alibaba-ifashion dataset The impact on the model results. Figure 5 , Figure 6 and Figure 7 It can be seen that model performance is quite sensitive to this hyperparameter, and its optimal value varies depending on the characteristics of the dataset. It is best performed on the Yelp2018, Amazon-Book, and Alibaba-ifashion datasets. The values ​​are 0.3, 1.0, and 0.1 respectively. When When the value is too high, the model may overemphasize bringing positive sample pairs closer together, potentially forcing semantically distinct item embeddings to be compressed into excessively close spatial distances, thus impairing the discriminative power of the representation. Conversely, when... When the value is too low, the effect of contrastive learning is weakened, and the model struggles to effectively utilize self-supervised signals to alleviate the data sparsity problem.

[0053] The momentum coefficient β is a key hyperparameter controlling the parameter update rate of the twin encoder in the dual encoder framework, and its impact on the model is as follows: Figure 8 , Figure 9 and Figure 10 As shown, where Figure 8 This section presents the impact of different momentum coefficients β on the model when training it using the Yelp 2018 dataset. Figure 9 This shows the effect of different momentum coefficients β on the model when training the model based on the Amazon-Book dataset. Figure 10 This paper presents the impact of different momentum coefficients β on the model when training on the Alibaba-ifashion dataset. When β is below 0.8, the Siamese encoder updates too quickly, making it difficult to effectively maintain historical parameter information and weakening its role as a stable learning objective. Conversely, when β exceeds 0.95, parameter updates are too slow, potentially leading to a decrease in model convergence speed. The impact of the momentum coefficient on model performance is relatively small; the performance of the GGDCL model on the three datasets mentioned above is not significantly affected. This phenomenon indicates that as long as the β value is within the effective range of 0.8-0.95, the momentum update mechanism can fully play its key role in maintaining training stability and providing a consistent learning objective. This range ensures that moderate differences between encoders are maintained in the early stages of training to provide diverse comparative perspectives. Therefore, β is uniformly set to 0.9 for all datasets. This demonstrates that the GGDCL model exhibits robustness related to hyperparameter selection.

[0054] To study the uniformity loss weight The effect of adjusting the uniformity of the representation distribution on the hypersphere was investigated on three datasets. Parameter tuning experiments were conducted, and the results are as follows: Figure 11 , Figure 12 and Figure 13 As shown, where Figure 11 Weights for loss with different uniformity when training a model based on the Yelp 2018 dataset Impact on the model results Figure 12 Weights for loss with different uniformity when training a model based on the Amazon-Book dataset Impact on the model results Figure 13 Different uniformity loss weights when training a model based on the Alibaba-ifashion dataset The impact on the model results. In different... At this value, the model performance fluctuates less and the trend is more gradual. This phenomenon can be explained by the design of the loss function: the InfoNCE loss itself, while bringing positive sample pairs closer together, also promotes the uniform distribution of negative samples in the representation space, that is, it implicitly optimizes uniformity. For the Yelp2018, Amazon-Book, and Alibaba-fashion datasets, this embodiment selects final values ​​of 0.5, 1.0, and 0.5, respectively.

[0055] In summary, this invention addresses the limitations of existing graph contrastive learning-based recommendation methods in terms of data augmentation and optimization objectives. It innovatively proposes the GGDCL framework, organically combining generative augmentation with dual-encoder contrastive learning, and optimizing it through alignment and uniformity losses. First, a variational graph autoencoder performs semantic reconstruction and completion on the original sparse interaction graph, fundamentally avoiding the semantic damage that random augmentation might cause, providing a higher-quality training foundation for the model. Second, a momentum-updated dual-encoder mechanism generates stable and diverse contrastive views without explicit data augmentation, effectively improving the robustness of the representation. Finally, abandoning the traditional BPR loss, it adopts alignment and uniformity as the core optimization objectives, enabling the model to directly optimize the quality of the embedding space, avoiding negative sampling bias and accelerating convergence, thereby improving the accuracy of item recommendations.

[0056] Example 3 This application also provides a user-item recommendation system, including: The interaction graph construction module is used to obtain the interaction data between the users to be recommended and the projects, and to construct the interaction graph between the users to be recommended and the projects.

[0057] The graph structure enhancement module is used to reparameterize the user-item interaction graph to be recommended according to preset node distribution parameters to obtain a latent semantic representation. The latent semantic representation is then reconstructed to obtain an enhanced graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distribution of user nodes and item nodes in the user-item interaction graph to be recommended. The module also extracts embedded features of the user to be recommended and multiple item embedded features from the enhanced graph structure.

[0058] The recommendation result determination module is used to determine the recommendation scores of multiple items relative to the user to be recommended based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; to rank the multiple items according to the recommendation scores of the multiple items relative to the user to be recommended, and to determine the item recommendation results for the user to be recommended based on the ranking results.

[0059] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps in an embodiment of a user-item recommendation method. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0060] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in a user-item recommendation method embodiment. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0061] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A user-item recommendation method, characterized in that, include: The interaction data between training users and projects is obtained, and a training user-project interaction graph is constructed from the interaction data. The training user-project interaction graph is semantically reconstructed to obtain a training augmented graph structure. The main view node embedding features are extracted from the training augmented graph structure by the main encoder of the GGDCL model, and the twin view node embedding features are extracted from the training augmented graph structure by the twin encoder of the GGDCL model. Alignment loss and uniformity loss are calculated based on the embedding features of the main view nodes; contrastive learning loss is calculated based on the embedding features of the main view nodes and the embedding features of the twin view nodes; variational graph generation loss is calculated based on the training user-item interaction graph and the training augmentation graph structure; the total loss of the GGDCL model is determined by the variational graph generation loss, alignment loss, uniformity loss, and contrastive learning loss; the parameters of the GGDCL model are optimized based on the total loss to obtain the trained GGDCL model, wherein the GGDCL model includes a variational graph generation module, a twin-graph encoder contrastive learning module, and a prediction module connected in sequence; the variational graph generation module is constructed based on a variational autoencoder; the twin-graph encoder contrastive learning module includes a main encoder and a twin encoder running parallel to the main encoder; the parameters of the twin encoder are determined by the current parameters of the main encoder based on a momentum update mechanism. Obtain the interaction data between the users to be recommended and the projects, and construct the interaction graph between the users to be recommended and the projects based on the interaction data; The user-item interaction graph to be recommended is processed by the GGDCL model to obtain recommendation scores for multiple items relative to the user to be recommended. The variational graph generation module reparameterizes the user-item interaction graph to be recommended according to preset node distribution parameters to obtain a latent semantic representation. The latent semantic representation is then reconstructed to obtain an augmented graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distributions of user nodes and item nodes in the user-item interaction graph to be recommended. The main encoder extracts the embedding features of the user to be recommended and the embedding features of multiple items from the augmented graph structure. The prediction module determines the recommendation scores of multiple items relative to the user to be recommended based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; the multiple items are ranked according to the recommendation scores of the multiple items relative to the user to be recommended, and the item recommendation results for the user to be recommended are determined based on the ranking results.

2. The user-item recommendation method according to claim 1, characterized in that, The variational graph generation module reparameterizes the user-item interaction graph to be recommended according to preset node distribution parameters to obtain a latent semantic representation. The latent semantic representation is then reconstructed to obtain an enhanced graph structure, including: The variational graph generation module performs reparameterized sampling on each node in the user-item interaction graph to be recommended, based on its node distribution parameters, to obtain the latent semantic representation of the node. The semantic similarity between multiple latent semantic representations is calculated, and the enhanced graph structure is reconstructed from the semantic similarity between each node.

3. The user-item recommendation method according to claim 1, characterized in that, The user-project interaction graph constructed from the interaction data between the users to be recommended and the projects includes: A node set is constructed based on the user set and item set in the interaction data between the users and items to be recommended; Construct a user-project interaction matrix based on the node set and the interaction records in the user-project interaction data; The adjacency matrix of the bipartite graph is constructed based on the user-item interaction matrix, and the user-item interaction graph to be recommended is determined by the node set and the adjacency matrix of the bipartite graph.

4. The user-item recommendation method according to claim 2, characterized in that, The main view node embedding features include user node embedding features and item node embedding features. The calculation of alignment loss and uniformity loss based on the main view node embedding features includes: The alignment loss is calculated based on the Euclidean distance between the user node embedding features and the item node embedding features. The uniformity loss is calculated based on the negative log-expected distance of the user node embedding feature distribution and the negative log-expected distance of the project node embedding feature distribution.

5. The user-item recommendation method according to claim 2, characterized in that, Parameter optimization of the GGDCL model based on the total loss includes: The current parameters of the twin encoder are obtained by updating the momentum of the current parameters of the main encoder. The current parameters of the master encoder are updated by backpropagation based on the total loss.

6. A user-item recommendation system, characterized in that, include: The model training module is used to acquire interaction data between training users and projects, construct a training user-project interaction graph from the interaction data, perform semantic reconstruction on the training user-project interaction graph to obtain a training augmented graph structure, extract main view node embedding features from the training augmented graph structure through the main encoder of the GGDCL model, and extract twin view node embedding features from the training augmented graph structure through the twin encoder of the GGDCL model. Alignment loss and uniformity loss are calculated based on the embedding features of the main view nodes; contrastive learning loss is calculated based on the embedding features of the main view nodes and the embedding features of the twin view nodes; variational graph generation loss is calculated based on the training user-item interaction graph and the training augmentation graph structure; the total loss of the GGDCL model is determined by the variational graph generation loss, alignment loss, uniformity loss, and contrastive learning loss; the parameters of the GGDCL model are optimized based on the total loss to obtain the trained GGDCL model, wherein the GGDCL model includes a variational graph generation module, a twin-graph encoder contrastive learning module, and a prediction module connected in sequence; the variational graph generation module is constructed based on a variational autoencoder; the twin-graph encoder contrastive learning module includes a main encoder and a twin encoder running parallel to the main encoder; the parameters of the twin encoder are determined by the current parameters of the main encoder based on a momentum update mechanism. The interaction graph construction module is used to obtain the interaction data between the user to be recommended and the project, and to construct the user-project interaction graph based on the interaction data between the user to be recommended and the project. The graph structure enhancement module processes the user-item interaction graph to be recommended using the GGDCL model to obtain recommendation scores for multiple items relative to the user to be recommended. The variational graph generation module reparameterizes the user-item interaction graph to be recommended according to preset node distribution parameters to obtain a latent semantic representation. The latent semantic representation is then reconstructed to obtain an enhanced graph structure. The node distribution parameters include the mean and standard deviation of the latent Gaussian distributions of user nodes and item nodes in the user-item interaction graph to be recommended. The main encoder extracts embedding features of the user to be recommended and embedding features of multiple items from the enhanced graph structure. The recommendation result determination module is used to determine the recommendation scores of multiple items relative to the user to be recommended by the prediction module based on the inner product of the embedded features of the user to be recommended and the embedded features of multiple items; to rank the multiple items according to the recommendation scores of the multiple items relative to the user to be recommended, and to determine the item recommendation result of the user to be recommended based on the ranking result.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the user-item recommendation method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the user-item recommendation method according to any one of claims 1 to 5.

Images