A recommendation system and method based on category distribution perception and noise enhanced contrast learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUZHOU UNIVERSITY
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-26
AI Technical Summary
Existing graph collaborative filtering recommendation methods face problems such as sparse interactions, interaction noise, and uneven interaction distribution in real-world applications. Existing contrastive learning reinforcement methods are difficult to cover multi-source noise and are disconnected from uneven distribution suppression, resulting in insufficient representation learning quality and recommendation robustness.
A recommendation system based on category distribution perception and noise-enhanced contrastive learning is adopted. The category distribution perception module generates initial representations of users and items with category priors, and the graph propagation and view generation module constructs noise-free views and views with various noise perturbations. The joint optimization module jointly optimizes the main recommendation task and the contrastive learning task in the same training process to generate more stable and discriminative representations.
It alleviates the popularity bias caused by uneven interaction distribution, improves the representation learning quality and recommendation robustness under sparse and noisy conditions, enhances the ability to adapt to noise, and improves recommendation ranking performance.
Smart Images
Figure CN122285990A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of personalized recommendation, specifically relating to a recommendation system and method based on category distribution perception and noise enhancement contrastive learning. Background Technology
[0002] Recommendation systems based on category distribution awareness and noise-enhanced contrastive learning belong to the graph collaborative filtering recommendation method within the field of personalized recommendation. Graph collaborative filtering methods typically model user-item interactions as a bipartite graph and learn user and item representations through neighborhood propagation and aggregation on the graph. Based on this, they calculate the user's predicted preference score for candidate items and generate a recommendation list. The recommendation performance of this type of method largely depends on the accuracy of representation learning and its stability under sparse and noisy conditions.
[0003] A typical graph collaborative filtering representation learning process generally includes: (1) constructing a bipartite graph based on interaction data and capturing high-order collaborative signals; (2) learning user and item representations based on collaborative signals and performing preference prediction and ranking output. With the development of graph neural networks, graph propagation-based collaborative filtering can make fuller use of neighborhood information to characterize complex interaction structures. However, in real applications, this type of method still faces the following challenges: First, the interaction is sparse, that is, a large number of users only generate a small number of behaviors, resulting in insufficient effective training signals that can be propagated; second, there is interaction noise, that is, random clicks, accidental touches and exposure-driven passive browsing make it difficult for the interaction to stably reflect the real preferences; third, the interaction distribution is uneven, that is, popularity and exposure strategies cause the interaction to be concentrated on a few popular items, making the training prone to bias towards high-frequency patterns, thus making it difficult to characterize the diverse preferences of users.
[0004] Traditional collaborative filtering methods learn latent representations through user-item interaction matrices, primarily relying on the assumption that interaction data provides stable and sufficient supervisory signals to characterize collaborative relationships. However, in real-world environments characterized by sparsity and noise, these supervisory signals are often insufficient and unreliable, leading to model biases towards high-frequency interaction structures. In recent years, graph contrastive learning has been introduced into recommendation systems to provide additional self-supervised constraints. Its basic idea is to construct multiple augmented views and constrain the consistency of representations of the same node across different views, mitigating the degradation in representation learning quality caused by the sparsity of supervisory signals. Existing augmentation methods typically include structural perturbations and embedding perturbations: structural perturbations may disrupt key collaborative semantics; embedding perturbations are often simple random offsets with insufficient correlation to noise sources. Therefore, existing technologies still have some unresolved issues, including: (1) insufficient enhancement of view diversity, making it difficult to cover the multiple sources of noise and different noise intensities that coexist in real interactions, resulting in insufficient constraints for contrastive learning and limited improvement in representation stability; (2) coarse-grained noise modeling and decoupling from noise sources, making it difficult to distinguish between weak noise such as random clicks and stronger disturbances such as exposure / popularity induction, resulting in insufficient view complementarity; (3) the impact of uneven interaction distribution is difficult to suppress during the representation learning stage, and training is prone to biasing the high-frequency patterns corresponding to popular items, and this bias is easily amplified during graph propagation. Therefore, a recommendation system and method that can take into account both the suppression of uneven interaction distribution and robust representation learning of multi-source noise without destroying the user-item collaborative structure is needed to improve the recommendation accuracy and stability under sparse and noisy conditions. Summary of the Invention
[0005] This invention addresses the problems of interaction sparsity, interaction noise, and uneven interaction distribution in existing graph collaborative filtering recommendation methods in real-world applications. In particular, it addresses the issue that existing contrastive learning enhancement methods are unable to cover multi-source noise and are disconnected from uneven distribution suppression. This invention proposes a recommendation system and method based on category distribution perception and noise-enhanced contrastive learning to improve the quality of representation learning and recommendation robustness under sparse and noisy conditions.
[0006] The recommendation system based on category distribution perception and noise-enhanced contrastive learning provided by this invention includes: a category distribution perception module, a graph propagation and view generation module, a joint optimization module, and a recommendation module. The category distribution perception module constructs user category distribution vectors and item category attribute vectors based on user and item category-related data, and generates initial user representation vectors and initial item representation vectors with category priors accordingly. The graph propagation and view generation module performs graph propagation based on a bipartite graph constructed from user-item interaction relationships, generating user and item representation vectors corresponding to noise-free views and at least one set of user and item representation vectors corresponding to noisy perturbation views during the graph propagation process. The joint optimization module constructs a main task optimization objective based on the user and item representation vectors corresponding to the noise-free views, and constructs a contrastive learning optimization objective based on the user and item representation vectors corresponding to the noisy perturbation views, thereby performing joint optimization to update user and item representation parameters. The recommendation module calculates user preference scores for items based on the jointly optimized user and item representation vectors and outputs the corresponding item preference ranking or Top-K recommendation results.
[0007] Furthermore, the category distribution perception module constructs a user's category distribution vector based on user-category interaction data and item-category information. Category attribute vector of items The category set is , The total number of categories, Indicates user Category Preference weights, Represents items With category Association weights; learning category embedding matrix ,in As the embedding dimension; and the user's category distribution vector Category attribute vector of items Mapping to the embedding space to generate user initial representation vectors and the initial representation vector of the item ,in and All A dimensional vector, where "·" represents the product of a vector and a matrix.
[0008] Furthermore, the graph propagation and view generation module includes a graph propagation submodule and a view generation submodule; the graph propagation submodule performs neighborhood aggregation on the user-item bipartite graph to obtain richer user and item vector representations; including: given a user-item bipartite graph ,in A collection of nodes representing users and items. To represent the set of edges representing the interaction relationships between users and items, a normalized adjacency matrix is constructed. ,in The bipartite diagram The adjacency matrix, The degree matrix; the user's initial representation matrix and the initial representation matrix of the item splicing to form the initial representation matrix ,in For the number of users, For the quantity of items; then, based on Propagation occurs at different levels, and the propagation rules are expressed as follows: After propagation, the information from all layers is aggregated and averaged layer by layer to obtain the updated representation. ,in, Indicates the number of floors. The view generation submodule is used to add noise during graph propagation to construct a noisy view; including: representing each layer during layer propagation. Add Gaussian noise Disturbance characterization ,in For noise terms, The noise intensity coefficient is used; after propagation ends, the final propagation characterization is calculated. As a Gaussian noise view Indicate; construct noise terms , Represents element-wise symbolic functions. This represents element-wise multiplication. The noise intensity coefficient represents the propagation characteristics. Add fixed norm noise The disturbance characterization is obtained. After the propagation ends, the final propagation characterization is calculated. As a fixed norm noise view express.
[0009] Furthermore, the joint optimization module is used to jointly optimize the main recommendation task loss and the multi-noisy view contrastive learning loss during the same training process to update the model parameters; wherein, the joint optimization module is based on the node representation vector of any view output by the graph propagation and view generation module. ,conduct Normalization process yields a normalized vector. ; and by using independent sampling, user profiles were constructed under the same perturbation mode. The two-sample view characterization As positive sample pairs, and As negative sample pairs, among them To and Different users, Indicates the noise view type. A Gaussian noise view. For fixed norm noise views, and This represents two independent sampled views under the same noise view type; given a batch of users. With batch of items Build the view Contrast loss The user-side loss is The loss on the item side is ,in It is an exponential function. For similarity function, The temperature coefficient is used; and the total contrast loss is obtained by weighting the contrast losses of the Gaussian noise view and the fixed norm noise view. ,in and Views With View The weighting coefficients; the joint optimization module is also used to optimize the training triples. Construct the BPR loss for the main recommendation task, where, For users Positive sample items, For users Negative sample items, To train a set of triples and predict preference scores The loss of the main task is obtained from the inner product of the user representation and the item representation. ,in The objective function is the Sigmoid function; and the contrastive loss, the main task loss, and the regularization term are combined to construct a joint objective function. in The regularization coefficient is... This is the set of trainable parameters for the model.
[0010] Furthermore, the recommendation module is used to calculate the user's preference prediction score for items based on the user's final representation vector and the item's final representation vector during the prediction phase, and output the Top-K recommendation results; wherein for any user With any candidate item The preference prediction score is ,in This represents the user's final representation vector. Represents items The final representation vector, This represents the vector inner product operation; the recommendation module sorts the preference prediction scores of the candidate items and selects the top-scoring items. Items as users The recommendation results, among which This is a preset integer value.
[0011] This invention also provides a recommendation method based on category distribution perception and noise enhancement contrastive learning, characterized in that it includes:
[0012] Step 1: Construct initial representations with category priors: Using user-category interaction data and item-category information as input, construct the user's category distribution vector and the item's category attribute vector, learn the category embedding matrix, and map the user's category distribution vector and the item's category attribute vector to the embedding space to generate user initial representation vectors and item initial representation vectors with category priors.
[0013] Step 2: Perform graph propagation and generate multi-view representations: Using the initial user representation vector and the initial item representation vector as input, perform neighborhood aggregation and layer propagation on the user-item bipartite graph constructed based on the user-item interaction relationship to generate user and item representation vectors corresponding to the noise-free view; and construct two types of noise views with complementary statistical properties based on two types of random noise during the graph propagation process, including adding Gaussian noise to the representation during the propagation process to obtain a Gaussian noise view, and adding fixed-norm noise to the propagated representation to obtain a fixed-norm noise view;
[0014] Step 3: Construct a joint objective function and update model parameters: Construct the BPR loss for the main recommendation task based on the user and item representation vectors corresponding to the noise-free view; construct the InfoNCE contrastive learning loss based on the Gaussian noise view and the fixed-norm noise view, wherein the node representation vectors are normalized and positive and negative sample pairs are constructed by independent sampling, the user-side loss and the item-side loss are calculated and weighted to obtain the total contrastive loss; the BPR loss, the contrastive learning loss and the regularization term are weighted and combined to form a joint objective function, and the model parameters are updated based on the joint objective function;
[0015] Step 4: Calculate preference prediction scores and output recommendation results: In the prediction stage, the final user representation vector and the final item representation vector are generated based on the updated model parameters. The user's preference prediction score for the candidate items is obtained by vector inner product operation. The preference prediction scores are sorted and the top K items with the highest scores are selected as the user's recommendation results, thereby outputting the user's corresponding item preference ranking or Top-K recommendation list.
[0016] The beneficial effects of this invention include: (1) alleviating the hot spot bias caused by uneven interaction distribution and improving the representation learning quality in sparse and long-tail scenarios. (2) constructing a noise view with complementary statistical properties to cover multi-source noise and improve robustness and stability in noisy environments. By injecting Gaussian noise and fixed norm noise into the graph propagation process to construct a noise view, it is possible to cover different noise forms such as weak perturbation and strong perturbation at the same time, and enhance the representation's adaptability to noise without destroying the user-item collaborative structure. (3) using noise to enhance the consistency of representations in different views through contrastive learning constraints, improving representation discriminability and improving recommendation ranking performance. By constraining the consistency of the representation of the same node under different noise views based on InfoNCE contrastive learning loss, it is possible to alleviate the representation instability and overfitting problem caused by relying solely on interaction supervision under conditions of sparseness and noise, thereby improving the effect of preference prediction and recommendation ranking. (4) requiring no additional annotation and easy to deploy and extend. This invention uses self-supervised contrastive learning to provide supplementary training signals, eliminating the need for additional manual annotation; it is also compatible with user-item bipartite graph collaborative filtering frameworks, and parameters such as noise intensity, temperature coefficient, and propagation layer number are configurable, facilitating deployment and expansion in different application scenarios. Attached Figure Description
[0017] Figure 1 Schematic diagram of the overall framework of the recommendation system of this invention
[0018] Figure 2 Schematic diagram of the graph propagation, view generation, and joint optimization process of this invention Detailed Implementation
[0019] As the background technology shows, existing graph collaborative filtering recommendation methods are easily affected by interaction sparsity, interaction noise, and uneven interaction distribution in real-world applications. This leads to unstable learning of user and item representations and insufficient characterization of long-tail interests, thus affecting recommendation accuracy and robustness. To address these issues, the applicant proposes a recommendation system and method based on category distribution awareness and noise-enhanced contrastive learning: a category prior is introduced during the representation initialization stage to mitigate the popularity bias caused by uneven distribution; during graph propagation, a noise-free view and multiple noise-perturbed views are constructed, and contrastive learning constrains the view... Figure 1 Consistency; jointly optimize the main recommendation task and the contrastive learning task during the same training process to obtain more stable and discriminative user and item representations.
[0020] The spirit and essence of the present invention will be further explained below with reference to the accompanying drawings and embodiments.
[0021] like Figure 1As shown, the recommendation system based on category distribution perception and noise-enhanced contrastive learning provided in this embodiment of the invention includes: a category distribution perception module 101, a graph propagation and view generation module 102, a joint optimization module 103, and a recommendation module 104. The category distribution perception module 101 generates initial representations of users and items with category priors; the graph propagation and view generation module 102 performs graph propagation based on user-item interaction relationships and generates noise-free and noise-perturbed view representations; the joint optimization module 103 constructs a main task optimization objective based on the noise-free view and a contrastive learning objective based on the noise-perturbed view to jointly optimize and update parameters; the recommendation module 104 calculates preference prediction scores based on the updated representations and outputs Top-K recommendation results.
[0022] like Figure 2 As shown, the graph propagation and view generation module 102 performs multi-layer propagation on the user-item bipartite graph to generate a noise-free view representation, and injects Gaussian noise and fixed norm noise respectively during the propagation process to construct a noise-perturbed view representation; the joint optimization module 103 constructs the main recommendation task loss based on the noise-free view representation, and constructs the contrastive learning loss based on the noise-perturbed view representation, thereby forming a joint objective function to update the model parameters.
[0023] In this embodiment, three public datasets are used: the Movie dataset, the Clothing dataset, and the Electronics dataset. For each user, approximately 30% of the user interaction data is retained as the test set, and the remaining interaction data is used as the training set.
[0024] In this embodiment, the category embedding dimension of the corresponding movie dataset, clothing dataset, and electronic dataset in the category distribution perception module is... The values are set to 19, 68, and 10 respectively. The category distribution perception module constructs a user's category distribution vector based on user-category interaction data and item-category information. Category attribute vector of items The category set is , The total number of categories, Indicates user Category Preference weights, Represents items With category Association weights; learning category embedding matrix ,in As the embedding dimension; and the user's category distribution vector Category attribute vector of items Mapping to the embedding space to generate user initial representation vectors and the initial representation vector of the item ,in and All A dimensional vector, where "·" represents the product of a vector and a matrix.
[0025] In this embodiment, the graph propagation and view generation module includes a graph propagation submodule and a view generation submodule; the graph propagation submodule performs neighborhood aggregation on the user-item bipartite graph to obtain richer user and item vector representations; including: given a user-item bipartite graph ,in A collection of nodes representing users and items. To represent the set of edges representing the interaction relationships between users and items, a normalized adjacency matrix is constructed. ,in The bipartite diagram The adjacency matrix, The degree matrix; the user's initial representation matrix and the initial representation matrix of the item splicing to form the initial representation matrix ,in For the number of users, For the quantity of items; then, based on Propagation occurs at different levels, and the propagation rules are expressed as follows: After propagation, the information from all layers is aggregated and averaged layer by layer to obtain the updated representation. ,in, Indicates the number of floors. In this embodiment Choose 2 and 3; the view generation submodule is used to add noise during graph propagation to construct a noisy view; including: representing each layer during layer propagation. Add Gaussian noise Disturbance characterization ,in For noise terms, The noise intensity coefficient is used; after propagation ends, the final propagation characterization is calculated. As a Gaussian noise view Indicate; construct noise terms , Represents element-wise symbolic functions. This represents element-wise multiplication. The noise intensity coefficient represents the propagation characteristics. Add fixed norm noise The disturbance characterization is obtained. After the propagation ends, the final propagation characterization is calculated. As a fixed norm noise view express;
[0026] In this embodiment, the joint optimization module is used to jointly optimize the main recommendation task loss and the multi-noisy view contrastive learning loss during the same training process to update the model parameters; wherein, the joint optimization module is based on the node representation vector of any view output by the graph propagation and view generation module. ,conduct Normalization process yields a normalized vector. ; and by using independent sampling, user profiles were constructed under the same perturbation mode. The two-sample view characterization As positive sample pairs, and As negative sample pairs, among them To and Different users, Indicates the noise view type. A Gaussian noise view. For fixed norm noise views, and This represents two independent sampled views under the same noise view type; given a batch of users. With batch of items In this example, and Set to 1024, build the view Contrast loss The user-side loss is The loss on the item side is ,in It is an exponential function. For similarity function, The temperature coefficient is used; and the total contrast loss is obtained by weighting the contrast losses of the Gaussian noise view and the fixed norm noise view. ,in and Views With View The weighting coefficients; the joint optimization module is also used to optimize the training triples. Construct the BPR loss for the main recommendation task, where, For users Positive sample items, For users Negative sample items, To train a set of triples and predict preference scores The loss of the main task is obtained from the inner product of the user representation and the item representation. ,in The objective function is the Sigmoid function; and the contrastive loss, the main task loss, and the regularization term are combined to construct a joint objective function. in The regularization coefficient is... This is the set of trainable parameters for the model.
[0027] In this embodiment, the recommendation module is used to calculate the user's preference prediction score for items based on the user's final representation vector and the item's final representation vector during the prediction phase, and output the Top-K recommendation results; wherein for any user With any candidate item The preference prediction score is ,in This represents the user's final representation vector. Represents items The final representation vector, This represents the vector inner product operation; the recommendation module sorts the preference prediction scores of the candidate items and selects the top-scoring items. Items as users The recommendation results, among which This is a preset integer value.
[0028] In this embodiment, to ensure the reproducibility of model training and the fairness of comparison, the Adam optimizer is used to iteratively update the model parameters, and the model parameters are initialized using the Xavier initialization method; the learning rate is set to 0.001, and the embedding dimension is uniformly set to 64. For any user... Positive sample interactive items Uniform negative sampling is used to randomly sample items from the candidate item set that do not match the user's preferences. Interacting items As negative samples, thus constructing training triples Used for BPR loss calculation. During training, parameters are iteratively updated using mini-batch stochastic gradient descent. Each iteration randomly samples a batch of users and a batch of items from the training set for comparison and loss calculation; in this embodiment, the batch size is set to 1024. In the prediction phase, the recommendation module performs BPR calculations for each user... Calculate the preference prediction score for all candidate items, and output the Top-K recommendation results in descending order of score, where K is 20, and Recall@20 and NDCG@20 are used as evaluation indicators.
[0029] To illustrate the scale and sparsity characteristics of the public datasets used in this embodiment, statistics were compiled on the number of users, items, categories, interactions, sparsity, and types of the Movie, Clothing, and Electronics datasets. The statistical results are shown in Table 1.
[0030] Table 1 shows the statistics of the three public benchmark datasets used in the experiment.
[0031]
[0032] To verify the effectiveness of the method in this embodiment, this embodiment selects the traditional collaborative filtering method BPR, the graph collaborative filtering method LightGCN, and several graph contrastive learning recommendation methods (SGL, SimGCL, NCL, XSimGCL, CECL) as the control baseline. They are compared under the same data partitioning and evaluation index conditions, and the results are shown in Table 2.
[0033] Table 2. Performance comparison of different baseline models and Da-NCL on Recall@20 and NDCG@20.
[0034]
[0035] To verify the contribution of the category distribution perception module and the multi-noise view contrastive learning mechanism to recommendation performance in this embodiment, several control variants were constructed based on the complete model Da-NCL and ablation experiments were conducted: First, the Gaussian noise generated during graph propagation was removed, and the contrastive learning loss of that view was also removed accordingly; Second, the fixed-norm noise generated during graph propagation was removed, and the contrastive learning loss of that view was also removed accordingly; Third, both types of noise views and their contrastive learning losses were removed simultaneously, retaining only the main task optimization under the noise-free view; Fourth, based on the third, the category distribution perception module was further removed, and representation learning was performed only based on the user-item interaction bipartite graph. The ablation experiments still used Recall@20 and NDCG@20 as evaluation metrics. The experiments are shown in Table 3: Removing any key module leads to varying degrees of decrease in Recall@20 and NDCG@20, with a more significant decrease when both types of noise view contrastive learning constraints were removed simultaneously, and the decrease was further amplified when the category distribution perception module was removed, indicating that category distribution perception and multi-noise view contrastive learning have a synergistic effect in improving representation stability and ranking quality.
[0036] As shown in Tables 2 and 3, the recommendation system and method based on category distribution awareness and noise-enhanced contrastive learning proposed in this embodiment achieve superior Recall@20 and NDCG@20 results on three public datasets. Meanwhile, ablation experiments show that the category distribution awareness module and the contrastive learning constraints corresponding to Gaussian noise views and fixed-norm noise views all contribute to the final performance improvement. Furthermore, the joint constraints of multiple noise views can further enhance the stability and robustness of representation learning, thereby improving the recommendation ranking quality.
[0037] Table 3 Comparison of ablation studies of Da-NCL and its variants
[0038]
[0039] In summary, this embodiment addresses the problem of unstable representation learning in graph collaborative filtering recommendation under conditions of sparse interaction, noisy interaction, and unbalanced interaction distribution. It proposes a recommendation system and method based on category distribution awareness and noise-enhanced contrastive learning: Category priors are introduced during the representation initialization phase to suppress popularity bias; during graph propagation, noise-free views and multiple noise-perturbed views are constructed, and contrastive learning constrains the consistency of representations of the same node across different views; the main recommendation task and the contrastive learning task are jointly optimized during the same training process to obtain more stable and discriminative user and item representations, thereby generating Top-K recommendation results.
[0040] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make possible variations and modifications to the solutions of the present invention using the disclosed methods and techniques without departing from the spirit and scope of the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the content of the technical solution of the present invention, shall fall within the protection scope of the technical solution of the present invention.
Claims
1. A recommendation system based on category distribution perception and noise enhancement contrastive learning, characterized in that, include: The module includes a category distribution perception module, a graph propagation and view generation module, a joint optimization module, and a recommendation module. The category distribution perception module constructs user category distribution vectors and item category attribute vectors based on user and item category-related data, and generates initial user representation vectors and initial item representation vectors with category priors accordingly. The graph propagation and view generation module performs graph propagation based on a bipartite graph constructed from user-item interaction relationships, and generates user and item representation vectors corresponding to noise-free views and at least one set of user and item representation vectors corresponding to noisy perturbation views during the graph propagation process. The joint optimization module constructs a main task optimization objective based on the user and item representation vectors corresponding to the noise-free views, and constructs a contrastive learning optimization objective based on the user and item representation vectors corresponding to the noisy perturbation views, and then performs joint optimization to update user and item representation parameters. The recommendation module calculates user preference scores for items based on the jointly optimized user and item representation vectors, and outputs the corresponding item preference ranking or Top-K recommendation results.
2. The recommendation system based on category distribution perception and noise enhancement contrastive learning according to claim 1, characterized in that, The category distribution perception module constructs a user's category distribution vector based on user-category interaction data and item-category information. Category attribute vector of items The category set is , The total number of categories, Indicates user Category Preference weights, Represents items With category The association weight; Learning category embedding matrix ,in For the embedded dimension; And the category distribution vector of the user Category attribute vector of items Mapping to the embedding space to generate user initial representation vectors and the initial representation vector of the item ,in and All A dimensional vector, where "·" represents the product of a vector and a matrix.
3. The recommendation system based on category distribution perception and noise enhancement contrastive learning according to claim 1, characterized in that, The graph propagation and view generation module includes a graph propagation submodule and a view generation submodule; The graph propagation submodule performs neighborhood aggregation on the user-item bipartite graph to obtain richer user and item vector representations; including: given a user-item bipartite graph ,in A collection of nodes representing users and items. To represent the set of edges representing the interaction relationships between users and items, a normalized adjacency matrix is constructed. ,in The bipartite diagram The adjacency matrix, The degree matrix; the user's initial representation matrix and the initial representation matrix of the item splicing to form the initial representation matrix ,in For the number of users, For the quantity of items; then, based on Propagation occurs at different levels, and the propagation rules are expressed as follows: After propagation, the information from all layers is aggregated and averaged layer by layer to obtain the updated representation. ,in, Indicates the number of floors. The view generation submodule is used to add noise during graph propagation to construct a noisy view; including: representing each layer during layer propagation. Add Gaussian noise term The disturbance characterization is obtained. ,in , The noise intensity coefficient is used; after propagation ends, the final propagation characterization is calculated. As a Gaussian noise view Representation; Construction of noise terms , Represents element-wise symbolic functions. This represents element-wise multiplication. The noise intensity coefficient represents the propagation characteristics. Add fixed norm noise The disturbance characterization is obtained. After the propagation ends, the final propagation characterization is calculated. As a fixed norm noise view express.
4. The recommendation system based on category distribution perception and noise enhancement contrastive learning according to claim 1, characterized in that, The joint optimization module is used to jointly optimize the main recommendation task loss and the multi-noisy view contrastive learning loss during the same training process to update the model parameters; wherein, the joint optimization module is based on the node representation vector of any view output by the graph propagation and view generation module. ,conduct Normalization process yields a normalized vector. ; and by using independent sampling, user profiles were constructed under the same perturbation mode. The two-sample view characterization As positive sample pairs, and As negative sample pairs, among them To and Different users, Indicates the noise view type. A Gaussian noise view. For fixed-norm noise views, and This represents two independent sampled views under the same noise view type; given a batch of users. With batch of items Build the view Contrast loss The user-side loss is The loss on the item side is ,in It is an exponential function. For similarity function, The temperature coefficient is used; and the total contrast loss is obtained by weighting the contrast losses of the Gaussian noise view and the fixed norm noise view. ,in and Views With View The weighting coefficients; the joint optimization module is also used to optimize the training triples. Construct the BPR loss for the main recommendation task, where, For users Positive sample items, For users Negative sample items, To train a set of triples and predict preference scores The loss of the main task is obtained from the inner product of the user representation and the item representation. ,in The objective function is the Sigmoid function; and the contrastive loss, the main task loss, and the regularization term are combined to construct a joint objective function. ,in The regularization coefficient is... This is the set of trainable parameters for the model.
5. The recommendation system based on category distribution perception and noise enhancement contrastive learning according to claim 1, characterized in that, The recommendation module is used to calculate the user's preference prediction score for items based on the user's final representation vector and the item's final representation vector during the prediction phase, and output the Top-K recommendation results; wherein for any user With any candidate item The preference prediction score is ,in This represents the user's final representation vector. Represents items The final representation vector, This represents the vector inner product operation; the recommendation module sorts the preference prediction scores of the candidate items and selects the top-scoring items. Items as users The recommendation results, among which This is a preset integer value.
6. The recommendation method based on category distribution perception and noise enhancement contrastive learning according to any one of claims 1 to 5, characterized in that, include: Step 1: Construct initial representations with category priors: Using user-category interaction data and item-category information as input, construct the user's category distribution vector and the item's category attribute vector, learn the category embedding matrix, and map the user's category distribution vector and the item's category attribute vector to the embedding space to generate user initial representation vectors and item initial representation vectors with category priors. Step 2: Perform graph propagation and generate multi-view representations: Using the initial user representation vector and the initial item representation vector as input, perform neighborhood aggregation and layer propagation on the user-item bipartite graph constructed based on the user-item interaction relationship to generate user and item representation vectors corresponding to the noise-free view. During the graph propagation process, two noise views with complementary statistical properties are constructed based on two types of random noise, including adding Gaussian noise to the representation during propagation to obtain a Gaussian noise view, and adding fixed norm noise to the propagation representation to obtain a fixed norm noise view. Step 3: Construct a joint objective function and update model parameters: Construct the BPR loss for the main recommendation task based on the user and item representation vectors corresponding to the noise-free view; construct the InfoNCE contrastive learning loss based on the Gaussian noise view and the fixed-norm noise view, wherein the node representation vectors are normalized and positive and negative sample pairs are constructed by independent sampling, the user-side loss and the item-side loss are calculated and weighted to obtain the total contrastive loss; the BPR loss, the contrastive learning loss and the regularization term are weighted and combined to form a joint objective function, and the model parameters are updated based on the joint objective function; Step 4: Calculate Preference Prediction Scores and Output Recommendation Results: During the prediction phase, based on the updated model parameters, generate the final user representation vector and the final item representation vector. Obtain the user's preference prediction score for candidate items through vector dot product operations. Sort the preference prediction scores and select the top-ranked items. Each item is used as a recommendation result for the user, thereby outputting a ranking of the user's item preferences or a Top-K recommendation list.