Recommendation model for adaptive denoising and interaction enhancement based on comparative learning

Abstract

According to the adaptive denoising and interaction enhancement recommendation model based on comparative learning provided by the invention, adaptive denoising and interaction filling based on structural similarity are combined. During model training, interference to the recommendation system is reduced by dynamically pruning interactions identified as noise. In order to maintain the balance between the positive samples and the negative samples, the model selects high-reliability interaction and performs filling according to the structural similarity between the user and the project embedding, thereby enriching the training data set. Besides, the invention introduces a mutual information maximization-based comparative learning strategy, optimizes user preferences between trimmed and filled interactions, and improves the ability of the model to capture key features in a potential vector space. Under the condition of not depending on additional auxiliary information, preferences in historical behaviors of the user are effectively learned, and the recommended item list is generated. The information utilization rate and generalization ability of the model are effectively enhanced, and the recommendation quality of the model is improved.

Description

Technical Field

[0001] The present invention relates to the field of recommendation technology, and in particular to a recommendation model based on contrastive learning and adaptive denoising and interactive enhancement. Background Art

[0002] In the era of big data, information overload has become a serious problem. Recommendation system is the most widely used information filtering technology, which can alleviate information overload by filtering content related to users and realize personalized recommendations.

[0003] As one of the most commonly used recommendation techniques, collaborative filtering relies on users' historical behavior data to deeply explore the interaction between users and items, and then predict users' potential interest in unknown items. However, in rating-based collaborative filtering methods, user ratings are usually affected by various factors. With the rapid development of deep learning, collaborative filtering methods based on neural networks have emerged, which can effectively capture high-order user-item interactions. However, the multi-layer gradient propagation mechanism of neural networks makes these methods more susceptible to interaction noise. Existing strategies to reduce the impact of noise mainly focus on data preprocessing and model training, reducing the negative impact of noise on the model by directly discarding noisy interactions or assigning them lower weights. However, these strategies ignore the impact of the denoising process on the sparsity of training data, which may have an adverse effect on the quality of recommendation results due to the inherent sparsity of rating data. Summary of the invention

[0004] To solve the above problems, the present invention proposes a new recommendation model that combines adaptive denoising with interactive filling based on structural similarity, including the following:

[0005] The data collection and preparation module is used to acquire and clean the data collected in the database to obtain the original interaction matrix.

[0006] The adaptive denoising module is used to discard interactions whose predicted values ​​are significantly different from the true values ​​to reduce the impact of noise interactions on model training.

[0007] The reliability calculation module is used to calculate the structural similarity between users and items to evaluate the reliability of the interaction.

[0008] The filling reliability interaction module is used to fill high-similarity interactions back into the sample according to the interaction reliability to replace the discarded noise interactions.

[0009] The recommendation result acquisition module is used to obtain the prediction value of the main model and recommend the top k items with higher prediction values ​​to the target user to form a personalized recommendation list.

[0010] The adaptive denoising module of the present invention discards samples with high loss values ​​during training to reduce noise. Specifically, it includes the following steps:

[0011] 1) Truncate the training loss exceeding a predefined threshold to 0, defined as:

[0012]

[0013] where τ is a predefined threshold.

[0014] 2) The loss value decreases with the increase of training iterations. A dynamic threshold function τ(T) is set instead of a fixed threshold, and the threshold is changed during the training process.

[0015] 3) Since the loss value is different on different data sets, τ(T) is converted into a discard rate function, increasing the discard rate in a linear manner. The specific formula is as follows:

[0016] τ(T)=min(αT,τ max )

[0017] where τ max is the maximum dropout rate, and α is a hyperparameter adjusted to the maximum dropout rate.

[0018] 4) Use binary cross entropy (BCE) as the loss function for model training, as follows:

[0019]

[0020] Where m is the number of users, S + and S - are respectively the positive and negative sample sets, y ui and are the true value and the predicted value respectively.

[0021] The reliability calculation module of the present invention calculates the structural similarity between users and items to evaluate the reliability of the interaction. Specifically, it includes the following steps:

[0022] 1) According to the homophily theory in social networks, nodes with similar characteristics or structures are more likely to interact than nodes with different characteristics. Specifically, given a user-item interaction Construct the user's first-order interaction graph L U First-order interaction graph with item L I :

[0023]

[0024] Where D represents the degree matrix of the interaction matrix R.

[0025] 2) Given user and item embedding matrices and Perform matrix multiplication of the embedding matrices of users and items with the first-order interaction graph respectively:

[0026]

[0027] Among them, H U and H I are the latent feature representations of users and items, respectively. Represents a matrix multiplication operation.

[0028] 3) Calculate the cosine similarity C between the pruned interacting users and the first-order propagation embedding vectors of the non-interacting items ui , the similarity calculation process is as follows:

[0029]

[0030] where h u Indicates H U The corresponding embedding vector of the pruned user, h i Indicates H I The embedding vector of the corresponding uninteracted term in .

[0031] 4) The structural similarity C ui As the reliability of the interaction between user u and item i, higher structural similarity indicates a greater possibility of reliable interaction between the user and the item.

[0032] The filling reliability interaction module of the present invention replaces the discarded noise interactions with high similarity interactions according to the interaction reliability. The specific steps are as follows:

[0033] 1) Select the top M non-interaction items i with the highest reliability and use the structural feature vector h f Embedding vector used as padding.

[0034]

[0035] in Represents a binary judgment function that returns 1 when the condition is true. C u Represents the set of structural similarities between user u and all non-interacted items.

[0036] 2) Introduce contrastive learning to supervise the training of the filler item and embed the user of the pruned interaction into the vector h u Its padded item embedding vector h f As a positive sample pair (h u ,h f ), h u The embedding vector h of the item that interacts with the pruned r As a negative sample pair (h u ,hr ). This part of the loss is defined as L INCE ,as follows:

[0037]

[0038] The recommendation result acquisition module of the present invention is used to obtain the recommendation list of the target user u. The predicted values ​​of the items that the user has not interacted with are sorted, and the top k items are selected to form a personalized recommendation list for the target user u. In order to consider the accuracy and feature diversity of the proposed model at the same time, BCE loss and INCE loss are combined to jointly optimize the model. The loss function is as follows:

[0039]

[0040] where λ1 and λ2 are hyperparameters that control contrastive learning loss and regularization, respectively. Θ={E U ,E I} represents a set of model parameters. Adam optimizer is used to optimize model parameters.

[0041] The present invention proposes a recommendation model based on contrastive learning for adaptive denoising and interaction enhancement. Combining adaptive denoising with interaction filling based on structural similarity, the dynamic threshold function is used to identify and prune noise interactions. The model uses the cosine similarity between the users of discarded interactions and the embeddings of non-interacting items to select highly reliable interactions to replace pruned interactions, thereby reducing noise interference. In addition, contrastive learning is introduced to further optimize the model's ability to capture key features in the latent vector space, effectively improving recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the model process of the present invention.

[0043] Figure 2 It is a model framework diagram of the present invention.

[0044] Figure 3 The Recall@5 value, Recall@10 value, NDCG@5 value, and NDCG@10 value of the present invention are compared with other methods in the MovieLens dataset. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clear, the following is a detailed description of the recommendation model based on contrastive learning for adaptive denoising and interactive enhancement proposed by the present invention in conjunction with the accompanying drawings. It should be understood that the specific implementation method described herein is only used to explain the present invention and is not used to limit the present invention. Changes, modifications, additions, modifications or substitutions made by ordinary technicians in the technical field within the scope of the present invention should be included in the scope of the claims of the present invention.

[0046] Figure 1 Schematic diagram of the process of the recommendation model based on contrastive learning and interactive enhancement. Figure 1 It can be seen that the recommendation model based on adaptive denoising and interaction enhancement based on contrastive learning proposed in the present invention includes acquiring and cleaning data information from the database to obtain the required user-item original interaction matrix. Interactions with high loss values ​​are discarded during training to reduce noise. Specifically, the training loss exceeding a predefined threshold is truncated to zero. Then the user and item structural similarity is calculated, and the structural similarity is used as a reliability measure. The calculated similarities are then ranked, and the first few non-interacted items with the highest structural similarity to the user are selected, and their structural feature vectors are used as the embedding vectors of the filler items. Then contrastive learning is introduced to supervise the filler item training to enhance the feature representation of users and items after denoising. The user's global preferences are learned, and the user's historical interaction information on the items is used as input to generate prediction values ​​and sort them to generate a list of recommended items.

[0047] Figure 2 This is a framework diagram of the recommendation model based on contrastive learning for adaptive denoising and interactive enhancement. Figure 2 It can be seen that the recommendation model based on contrastive learning and adaptive denoising and interaction enhancement proposed in the present invention includes a data collection and preparation module, an adaptive denoising module, a reliability calculation module, a filling reliability interaction module, and a recommendation result acquisition module. The data collection and preparation module is used to extract data from the database, and perform operations such as cleaning and preprocessing on the data to obtain the required original interaction matrix. The adaptive denoising module aims to reduce the impact of noise in user-item interaction on neural network representation learning, and dynamically prune high loss value noise interactions that exceed a predefined threshold. The reliability calculation module calculates the cosine similarity between the first-order propagation embedding vectors of the pruned interacting users and non-interacting items, and selects user-item pairs with high structural similarity between users and non-interactive items as high-reliability interactions. The filling reliability interaction module uses the structural feature vector of the high-reliability interaction as the embedding vector of the filling item, and supervises the training through contrastive learning. The recommendation result acquisition module takes the user's historical behavior information as input, learns the user's preferences, sorts the predicted values ​​of the user's non-interacting items, and selects the top k items to form a personalized recommendation list for the target user u.

[0048] Further, the following situation is used as an example to illustrate:

[0049] There are M users U={u1,u2,...,u M} and N items I={i1,i2,...,i N}, the rating domain is {1,2,3,4,5}. represents the true preference matrix, r ui is the true preference of user u for item i. If there is interaction between user u and item i, then r ui =1, otherwise r ui =0.

[0050] First, the system data is acquired and cleaned through the data acquisition and preparation module to obtain the variables in the hypothesis. The specific implementation steps of the proposed adaptive denoising and interactive enhancement based on contrastive learning are as follows:

[0051] S1: The data collection and preparation module obtains the required data information from the database, including user ID, project ID and interaction matrix R;

[0052] S2: Adaptive denoising module, which trims noise interactions through a dynamic threshold function to reduce the negative impact of noise on model training

[0053] 1) Truncate the training loss above a predefined threshold to 0:

[0054]

[0055] where τ is a predefined threshold.

[0056] 2) The loss value decreases with the increase of training iterations. A dynamic threshold function τ(T) is set instead of a fixed threshold, and the threshold is changed during the training process.

[0057] 3) Since the loss value is different on different data sets, the noise interaction is discarded in a linear manner. The specific formula is as follows:

[0058] τ(T)=min(αT,τ max )

[0059] where τ max is the maximum dropout rate, and α is a hyperparameter adjusted to the maximum dropout rate.

[0060] 4) Use binary cross entropy (BCE) as the loss function for model training, as follows:

[0061]

[0062] Where m is the number of users, S + and S - are respectively the positive and negative sample sets, y ui and are the true value and the predicted value respectively.

[0063] S3: Reliability calculation module, which calculates the structural similarity between users and items to evaluate the reliability of the interaction. It specifically includes the following steps:

[0064] 1) Given a user-item interaction Construct the user's first-order interaction graph L U First-order interaction graph with item L I :

[0065]

[0066] Where D represents the degree matrix of the interaction matrix R.

[0067] 2) Given user and item embedding matrices and Perform matrix multiplication of the embedding matrices of users and items with the first-order interaction graph respectively:

[0068]

[0069] Among them, H U and H I are the latent feature representations of users and items, respectively. Represents a matrix multiplication operation.

[0070] 3) Calculate the cosine similarity C between the first-order propagation embedding vectors of the pruned interacting users and the non-interacting items ui , the similarity calculation process is as follows:

[0071]

[0072] where h u Indicates H U The corresponding embedding vector of the pruned interacting user, h i Indicates H I The embedding vector of the corresponding uninteracted term in .

[0073] 4) The structural similarity C ui As the reliability of the interaction between user u and item i, higher structural similarity indicates a greater possibility of reliable interaction between the user and the item.

[0074] S4: Filling reliability interaction module, according to the interaction reliability, the discarded noise interactions are replaced with high similarity interactions.

[0075] The specific steps are as follows:

[0076] 1) Select the top M non-interaction items i with the highest reliability and use the structural feature vector h f Embedding vector used as padding.

[0077]

[0078] in Represents a binary judgment function that returns 1 when the condition is true. C u represents the set of structural similarities between user u and all non-interacted items.

[0079] 2) The user embedding vector h of the pruned interaction u Its padded item embedding vector h f As a positive sample pair (h u ,h f ), h u The embedding vector h of the item that interacts with the pruned r As a negative sample pair (h u ,h r ). This part of the loss is defined as L INCE ,as follows:

[0080]

[0081] S5: Recommendation result acquisition module, used to obtain the recommendation list of target user u. In order to consider the accuracy and feature diversity of the proposed model at the same time, BCE loss and INCE loss are combined to jointly optimize the model. The loss function is as follows:

[0082]

[0083] where λ1 and λ2 are hyperparameters that control contrastive learning loss and regularization, respectively. Θ={E U ,E I} represents a set of model parameters. Adam optimizer is used to optimize model parameters.

[0084] Figure 3 The proposed method is measured in two data and two aspects of the performance indicators, and compared with other methods. The two metrics are: Recall and Normalized Discounted Cumulative Gain (NDCG). Their calculation methods and measurement contents are as follows:

[0085] Recall: measures the model's ability to cover the items that users actually interact with. The calculation method is as follows:

[0086]

[0087] Where k is the number of recommended items, is the item recommended to user u, R u are the items that user u actually interacts with.

[0088] Normalized Discounted Cumulative Gain (NDCG): This shows the quality of the model’s project ranking recommendations by assigning higher scores to top-ranked projects. The calculation formula is as follows:

[0089]

[0090] Among them, DCG@k represents the discounted cumulative benefits of the first k recommended items, IDCG u @k is an ideal DCG. p represents the relevance of the recommended item at position p. If the item is one of the actual recommended items, then rel p Defined as 1, otherwise defined as 0.

Claims

1. Adaptive denoising and interaction enhancement recommendation model based on contrastive learning is used to reduce the impact of noise on model training. At the same time, it can enrich the interaction between users and items and improve the quality of recommendations while maintaining data sparsity, including the following: The data collection and preparation module is used to acquire and clean the data collected in the database to obtain the required original interaction matrix. The adaptive denoising module is used to discard interactions whose predicted values ​​are significantly different from the true values ​​to reduce the impact of noise interactions on model training. The reliability calculation module is used to calculate the structural similarity between users and items to evaluate the reliability of the interaction. The filling reliability interaction module is used to fill high-similarity interactions back into the sample according to the interaction reliability to replace the discarded noise interactions. The recommendation result acquisition module is used to obtain the prediction value of the main model and recommend the top k items with higher prediction values ​​to the target user to form a personalized recommendation list.

2. The recommendation model based on contrastive learning for adaptive denoising and interactive enhancement as claimed in claim 1, characterized in that: The adaptive denoising module discards samples with high loss values ​​during training to reduce noise. Specifically, it includes the following steps: S1: Truncate the training loss exceeding a predefined threshold to 0, defined as: where τ is a predefined threshold. S2: The loss value decreases with the increase of training iterations. A dynamic threshold function τ(T) is set instead of a fixed threshold, and the threshold is changed during training. S3: Since the loss value is different on different data sets, τ(T) is converted into a discard rate function, increasing the discard rate in a linear manner. The specific formula is as follows: τ(T)=min(αT,τ max ) where τ max is the maximum dropout rate, and α is a hyperparameter adjusted to the maximum dropout rate. S4: Binary cross entropy (BCE) is used as the loss function for model training, as follows: Where m is the number of users, S + and S - are respectively the positive and negative sample sets, y ui and are the true value and the predicted value respectively.

3. The recommendation model based on contrastive learning for adaptive denoising and interactive enhancement as claimed in claim 1, characterized in that: The reliability calculation module calculates the structural similarity between users and items to evaluate the reliability of the interaction. Specifically, it includes the following steps: S1: According to the homophily theory in social networks, nodes with similar characteristics or structures are more likely to interact than nodes with different characteristics. Specifically, given a user-item interaction Construct the user's first-order interaction graph L U First-order interaction graph with item L I : Where D represents the degree matrix of the interaction matrix R. S2: Given user and item embedding matrices and Perform matrix multiplication of the embedding matrices of users and items with the first-order interaction graph respectively: Among them, H U and H I are the latent feature representations of users and items, respectively. Represents a matrix multiplication operation. S3: Calculate the cosine similarity C between the pruned interacting users and the first-order propagated embedding vectors of the non-interacting items ui , the similarity calculation process is as follows: where h u Indicates H U The corresponding embedding vector of the pruned user, h i Indicates H I The embedding vector of the corresponding uninteracted term in . S4: Taking structural similarity as the reliability of the interaction between user u and item i, higher structural similarity indicates a greater possibility of reliable interaction between the user and the item.

4. The recommendation model based on contrastive learning for adaptive denoising and interactive enhancement as claimed in claim 1, characterized in that: The filling reliability interaction module replaces the discarded noise interactions with high similarity interactions according to the interaction reliability. The specific steps are as follows: S1: Select the top M non-interacting items i with the highest reliability and use the structural feature vector h f Embedding vector used as padding. in Represents a binary judgment function that returns 1 when the condition is true. C u represents the set of structural similarities between user u and all non-interacted items. S2: Introduce contrastive learning to supervise the training of the filler term and embed the user of the pruned interaction into the vector h u Its padded item embedding vector h f As a positive sample pair (h u ,h f ), h u The embedding vector h of the item that interacts with the pruned r As a negative sample pair (h u ,h r ). This part of the loss is defined as L INCE ,as follows:

5. The recommendation model based on contrastive learning for adaptive denoising and interactive enhancement as claimed in claim 1, characterized in that: The recommendation result acquisition module is used to obtain the recommendation list of the target user u. The predicted values ​​of the items that the user has not interacted with are sorted, and the top k items are selected to form a personalized recommendation list for the target user u. In order to consider the accuracy and feature diversity of the proposed model at the same time, BCE loss and INCE loss are combined to jointly optimize the model. The loss function is as follows: where λ1 and λ2 are hyperparameters that control contrastive learning loss and regularization, respectively. Θ={E U ,E I } represents a set of model parameters. Adam optimizer is used to optimize model parameters.

Images