A personalized label recommendation method based on contrastive learning

Abstract

The application belongs to the field of artificial intelligence recommendation systems, and discloses a personalized label recommendation method based on contrast learning, which comprises the following steps: initializing an implicit feature matrix through Gaussian prior, obtaining high-order embedding representations of users, items, user-oriented labels and item-oriented labels on a user and label interaction graph and an item and label interaction graph by using a graph convolutional neural network, injecting noise into the high-order embedding representations to obtain embedding representations of the users and the user-oriented labels with noise, calculating a contrast loss by using Info_NEC, constructing an objective function, learning model parameters by using an Adam optimizer, predicting the scores of the labels by using the sum of the inner products of the embedding representations of the users and the user-oriented labels and the inner products of the embedding representations of the items and the item-oriented labels, and providing personalized label recommendations according to the predicted scores. The application extracts valuable information from the input data itself by using contrast learning, thereby reducing the data sparsity problem.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence recommendation systems, specifically relating to a personalized tag recommendation method based on contrastive learning. Background Technology

[0002] With the explosive growth of internet information, tags have become increasingly important in the field of internet information retrieval. Tag recommendation greatly improves the quality of tags by assisting users in tagging, and has therefore attracted widespread attention from researchers.

[0003] Contrastive learning, a form of self-supervised learning, extracts useful information from raw data to compensate for insufficient supervisory signals, thereby improving model performance. It has received widespread attention from industry and academia and has achieved great success in fields such as computer vision and natural language processing. It augments the data itself, comparing the representations of nodes under different augmentations. By measuring the similarity of the same and different nodes under various augmentations, it extracts useful supervisory signals, enhancing the learning ability of machine learning models.

[0004] Personalized tag recommendation is an important branch of tag recommendation. It recommends personalized tags to each user by considering their tag preferences. Traditional personalized tag recommendation algorithms, such as PITF, NITF, and GNN-PTR, all employ supervised learning methods to build their models. Due to data sparsity issues, traditional supervised learning-based personalized tag recommendation models cannot fully learn the embedded representations of users, items, and tags, resulting in suboptimal recommendation performance. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention integrates a contrastive learning module into the traditional personalized tag recommendation algorithm. By utilizing contrastive learning to extract valuable information from the input data itself, the invention alleviates the data sparsity problem in traditional personalized tag recommendation models and improves the performance of existing recommendation algorithms.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] This invention is a personalized tag recommendation method based on contrastive learning, which includes the following steps:

[0008] Step 1): First, initialize the implicit user feature matrix, implicit item feature matrix, implicit user-oriented label feature matrix, and implicit item-oriented label feature matrix using Gaussian priors.

[0009] Step 2) Use graph convolutional neural networks to obtain high-order embedding representations of users, items, user-oriented tags, and item-oriented tags on the user-tag interaction graph and the item-tag interaction graph, respectively.

[0010] Step 3): For the user-tag interaction graph, noise is injected into the higher-order embedding representations of users and user-tags after each convolution operation, resulting in noisy user and user-tag embedding representations. The contrastive loss is calculated using Info_NEC based on these noisy embedding representations. For the item-tag interaction graph, noise is injected into the higher-order embedding representations of items and item-tags after each convolution operation, resulting in noisy item and item-tag embedding representations. The contrastive loss is calculated using Info_NEC based on these noisy item and item-tag embedding representations.

[0011] Step 4) Integrate contrastive loss into the Bayesian personalized ranking loss to construct the objective function, and learn the model parameters through the Adam optimizer.

[0012] Step 5) Predict the tag rating by summing the inner product of the embedding representations of user and user-oriented tags with the inner product of the embedding representations of item and item-oriented tags, and provide personalized tag recommendations based on the predicted ratings.

[0013] A further improvement of the present invention is that, in step 1), firstly, the implicit user feature matrix is ​​initialized using a Gaussian prior. Implicit item feature matrix Implicit user-oriented tag feature matrix and implicit item-oriented label feature matrix , Indicates the embedding dimension. The quantities of users, items, and tags are represented as follows:

[0014]

[0015]

[0016]

[0017]

[0018] Given user, item, and tag ID information, through the feature matrix , , and To acquire users Embedded representation of item i and tag t facing the user and tag t facing the item:

[0019]

[0020]

[0021]

[0022]

[0023] in, . The operation means extracting the corresponding embedding vector from the corresponding embedding representation matrix based on the entity's index information.

[0024] A further improvement of the present invention is that, in step 2), a graph convolutional neural network is used to obtain high-order embedding representations of users, items, user-oriented tags, and item-oriented tags on the user-tag interaction graph and the item-tag interaction graph, respectively.

[0025] First, based on the interaction information of users, items, and tags, an interaction matrix between users and tags is constructed. Interaction matrix of items and tags If user u and label t have interacted, then The value in the u-th row and t-th column of the matrix is ​​1, otherwise it is 0. If item i and tag t have interacted, then... The value in the i-th row and t-th column of the matrix is ​​1, otherwise it is 0. Then, using... and Construct the corresponding adjacency matrix and ,as follows:

[0026]

[0027]

[0028] in, and They are respectively and The transpose of .

[0029] Next, based on the adjacency matrix , Construct a Laplace matrix of users and tags Laplace matrix of items and tags ,as follows:

[0030]

[0031]

[0032] in, It is a diagonal matrix of users and tags. It is a diagonal matrix of items and labels, as follows:

[0033]

[0034]

[0035] in, Indicates user The number of first-order neighbors, Represents items The number of first-order neighbors, Indicates label The number of first-order neighbors is calculated, and the number of first-order neighbors for other users, items, and tags is calculated similarly.

[0036] Based on the Laplace matrix of users and tags Laplace matrix of items and tags We construct higher-order embedding representations of users, items, user-oriented tags, and item-oriented tags, as follows:

[0037]

[0038]

[0039] in, = , = This is the initial embedding representation. Indicates user The initial embedding representation, Represents items The initial embedding representation, Indicates label The initial embedding representation.

[0040] After L layers of convolution, the final embedding representation is obtained, as follows:

[0041]

[0042]

[0043] Will The final embedded representation of all users is obtained by segmenting based on the number of users and the number of tags. and all user-facing tag embedding representations .Will The final embedded representation of all items is obtained by segmenting based on the number of items and the number of tags. and the embedded representation of each item tag .

[0044] A further improvement of this invention is that, in step 3), for the user-tag interaction graph, noise is injected into the higher-order embedding representations of the user and user-tag after each convolution operation, finally obtaining noisy embedding representations of the user and user-tag. Based on the noisy embedding representations of the user and user-tag, the contrastive loss is calculated using Info_NEC. For the item-tag interaction graph, noise is injected into the higher-order embedding representations of the item and item-tag after each convolution operation, finally obtaining noisy embedding representations of the item and item-tag. Based on the noisy embedding representations of the item and item-tag, the contrastive loss is calculated using Info_NEC.

[0045] First, for the user-label interaction graph, noise is injected into the higher-order embedding representations of the user and user-label after each convolution operation, resulting in noisy embedding representations of the user and user-label, as follows:

[0046]

[0047]

[0048]

[0049]

[0050] in, and Let represent the noise vector, and . Indicates the injection of the first type of noise An embedded representation of all users and tags. Indicates the injection of a second type of noise An embedded representation of all users and tags. Indicates the injection of the first type of noise An embedded representation of all items and tags. Indicates the injection of a second type of noise An embedded representation of all items and tags.

[0051] After L layers of convolution, the final embedded noise representation is obtained, as follows:

[0052]

[0053]

[0054]

[0055]

[0056] Will( and The final embedded noise representation is obtained by segmenting based on the number of users and the number of tags. .Will and The final embedded noise representation is obtained by segmenting based on the number of items and tags. and .

[0057] exist Based on this, the contrast loss of the user is calculated using Info_NEC. ,as follows:

[0058]

[0059] in, Indicates the sampling batch. This represents the regularized embedding representation of user u under the injection of the first type of noise, i.e. . This represents the regularized embedding representation of user u under the injection of the second type of noise, i.e. = . This represents the temperature coefficient.

[0060] exist Based on this, the contrast loss of user-oriented labels is calculated using Info_NEC. ,as follows:

[0061]

[0062] in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. . This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = .

[0063] exist Based on this, the comparative loss of the item is calculated using Info_NEC. ,as follows:

[0064]

[0065] in, The embedding representation of item i after regularization under the injection of the first type of noise is, i.e. . This represents the regularized embedding representation of item i under the injection of the second type of noise, i.e. = .

[0066] exist Based on this, the contrastive loss for item-oriented tags is calculated using Info_NEC. ,as follows:

[0067]

[0068] in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. . This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = .

[0069] Finally, according to Calculate contrast loss :

[0070]

[0071] A further improvement of the present invention is that, in step 4), a contrastive loss is integrated on the basis of the Bayesian personalized ranking loss to construct an objective function, and the model parameters are learned through the Adam optimizer.

[0072] First, according to and The final embedding representations of user u, item i, and tag t (user-oriented and item-oriented) are obtained as follows:

[0073]

[0074]

[0075]

[0076]

[0077] Then, according to Obtain the probability score of user u tagging item i with t. ,as follows:

[0078]

[0079] in, This represents the dot product of vectors.

[0080] Finally, the objective formula for personalized tag recommendation based on contrastive learning is as follows:

[0081] in, It is the sigmoid function. , This represents the score for the positive sample label. The score represents the label of the negative sample. This represents the coefficient controlling for the contrastive learning loss. Indicates comparative loss, Indicates control The regularization coefficient, These are the parameters of the model. Minimize using the Adam optimizer. And optimize model parameters .

[0082] A further improvement of the present invention is that, in step 5), the rating of the tag is predicted by the sum of the inner product of the embedded representations of the user and the user-oriented tag and the inner product of the embedded representations of the item and the item-oriented tag.

[0083] Using the learned user embedding representation User-oriented tag embedding representation Embedded representation of items and item-oriented tag embedding representation Calculate the predicted rating for user u to tag item i with t. ,as follows:

[0084]

[0085] After calculating predicted scores for all tags, the tags with the highest scores are recommended to users.

[0086] The beneficial effects of this invention are: This method integrates a contrastive learning module into the traditional personalized label recommendation algorithm, and uses contrastive learning to extract valuable information from the input data itself, thereby alleviating the data sparsity problem in traditional personalized label recommendation and improving the performance of existing personalized label recommendation models. Attached Figure Description

[0087] Figure 1 This is a flowchart illustrating the personalized tag recommendation method of the present invention. Detailed Implementation

[0088] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential.

[0089] like Figure 1As shown, this invention is a personalized tag recommendation method based on contrastive learning, which includes the following steps:

[0090] Step 1: Initialize the implicit user feature matrix, implicit item feature matrix, implicit user-oriented label feature matrix, and implicit item-oriented label feature matrix using Gaussian priors. This includes the following steps:

[0091] Step 1-1: Initialize the implicit user feature matrix using Gaussian priors. Implicit item feature matrix Implicit user-oriented tag feature matrix and implicit item-oriented label feature matrix ,as follows:

[0092]

[0093]

[0094]

[0095]

[0096] In the formula, Indicates the embedding dimension. These represent the number of users, items, and tags, respectively.

[0097] Steps 1-2: Given user, item, and tag ID information, use the user feature matrix... Item feature matrix User-oriented tag feature matrix and item-oriented label feature matrix To acquire users Embedded representation of item i and tag t facing the user and tag t facing the item:

[0098]

[0099]

[0100]

[0101]

[0102] in, , The operation means extracting the corresponding embedding vector from the corresponding embedding representation matrix based on the entity's index information.

[0103] Step 2: On the user-tag interaction graph and the item-tag interaction graph, respectively, use a graph convolutional neural network to obtain high-order embedding representations of users, items, user-oriented tags, and item-oriented tags. This includes the following steps:

[0104] Step 2-1: Construct an interaction matrix between users and tags based on the interaction information of users, items, and tags. Interaction matrix of items and tags If user u and label t have interacted, then the user-label interaction matrix is... The value in the u-th row and t-th column of the matrix is ​​1, otherwise it is 0. If item i and tag t have interacted, then the interaction matrix between the item and tag is... The value in the i-th row and t-th column is 1, otherwise it is 0;

[0105] Step 2-2: Utilize the user-tag interaction matrix Interaction matrix of items and tags Construct the corresponding adjacency matrix and :

[0106]

[0107]

[0108] in, and They are respectively and The transpose of the matrix;

[0109] Steps 2-3: Based on the adjacency matrix , Construct a Laplace matrix of users and tags Laplace matrix of items and tags :

[0110]

[0111]

[0112] in, It is a diagonal matrix of users and tags. This is a diagonal matrix of items and labels, represented as:

[0113]

[0114]

[0115] in, Indicates user The number of first-order neighbors, Represents items The number of first-order neighbors, Indicates label The number of first-order neighbors is calculated, and the number of first-order neighbors for other users, items, and tags is calculated similarly.

[0116] Steps 2-4: Based on the Laplace matrix of users and tags Laplace matrix of items and tags We construct higher-order embedding representations of users, items, user-oriented tags, and item-oriented tags, as follows:

[0117]

[0118]

[0119] in, = = This is the initial embedding representation. Indicates user The initial embedding representation, Represents items The initial embedding representation, Indicates label The initial embedding representation;

[0120] Steps 2-5: After L layers of convolution, the final embedding representation is obtained, as follows:

[0121]

[0122]

[0123] Will The final embedded representation of all users is obtained by segmenting based on the number of users and the number of tags. and all user-facing tag embedding representations ,Will The final embedded representation of all items is obtained by segmenting based on the number of items and the number of tags. and the embedded representation of each item tag .

[0124] Step 3: For the user-tag interaction graph, after each convolution operation, noise is injected into the higher-order embedding representations of the user and user-tag, resulting in noisy user and user-tag embedding representations. Based on these noisy embedding representations, the contrastive loss is calculated using Info_NEC. For the item-tag interaction graph, after each convolution operation, noise is injected into the higher-order embedding representations of the item and item-tag, resulting in noisy item and item-tag embedding representations. Based on these noisy item and item-tag embedding representations, the contrastive loss is calculated using Info_NEC. The specific steps include:

[0125] Step 3-1: For the user-label interaction graph, inject noise into the higher-order embedding representations of the user and user-label after each convolution operation, and finally obtain the noisy embedding representations of the user and user-label, as follows:

[0126]

[0127]

[0128]

[0129]

[0130] in, and Describes the noise vector and , Indicates the injection of the first type of noise An embedded representation of all users and tags. Indicates the injection of a second type of noise An embedded representation of all users and tags. Indicates the injection of the first type of noise An embedded representation of all items and tags. Indicates the injection of a second type of noise An embedded representation of all items and tags.

[0131] Step 3-2: After L layers of convolution, the final embedded noise representation is obtained, as follows:

[0132]

[0133]

[0134]

[0135]

[0136] Will( and The final embedded noise representation is obtained by segmenting based on the number of users and the number of tags. ,Will and The final embedded noise representation is obtained by segmenting based on the number of items and tags. and ;

[0137] Step 3-3, in Based on this, the contrast loss of the user is calculated using Info_NEC. ,as follows:

[0138]

[0139] in, Indicates the sampling batch. This represents the regularized embedding representation of user u under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of user u under the injection of the second type of noise, i.e. = , Indicates the temperature coefficient;

[0140] Steps 3-4, in Based on this, the contrast loss of user-oriented labels is calculated using Info_NEC. ,as follows:

[0141]

[0142] in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = ;

[0143] Steps 3-5, in Based on this, the comparative loss of the item is calculated using Info_NEC. ,as follows:

[0144]

[0145] in, The embedding representation of item i after regularization under the injection of the first type of noise is, i.e. , This represents the regularized embedding representation of item i under the injection of the second type of noise, i.e. = ;

[0146] Steps 3-6, in Based on this, the contrastive loss for item-oriented tags is calculated using Info_NEC. ,as follows:

[0147]

[0148] in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = ;

[0149] Steps 3-7, according to Calculate contrast loss :

[0150] .

[0151] Step 4: Construct an objective function by integrating contrastive loss on top of the Bayesian personalized ranking loss, and learn the model parameters using the Adam optimizer. This includes the following steps:

[0152] Step 4-1, according to and The final embedding representations of user u, item i, and tag t (user-oriented and item-oriented) are obtained as follows:

[0153]

[0154]

[0155]

[0156]

[0157] Step 4-2, according to Obtain the probability score of user u tagging item i with t. ,as follows:

[0158]

[0159] in, Represents the dot product of vectors;

[0160] Step 4-3: The objective formula for personalized tag recommendation based on contrastive learning is as follows:

[0161] in, It is the sigmoid function. , This represents the score for the positive sample label. The score represents the label of the negative sample. This represents the coefficient controlling for the contrastive learning loss. Represents the regularization coefficient. These are the model's parameters, minimized by the Adam optimizer. and learn model parameters .

[0162] Step 5: Predict the tag rating by summing the inner product of the embedding representations of user and user-oriented tags with the inner product of the embedding representations of item and item-oriented tags, and provide personalized tag recommendations based on the predicted ratings, specifically:

[0163] Using the learned user embedding representation User-oriented tag embedding representation Embedded representation of items and item-oriented tag embedding representation Calculate the predicted rating for user u to tag item i with t. ,as follows:

[0164]

[0165] After calculating predicted scores for all tags, the tags with the highest scores are recommended to users.

[0166] This method integrates a contrastive learning module into traditional personalized label recommendation algorithms. By utilizing contrastive learning to extract valuable information from the input data itself, it alleviates the data sparsity problem in traditional personalized label recommendation and improves the performance of existing personalized label recommendation models.

[0167] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A personalized tag recommendation method based on contrastive learning, characterized in that: The personalized tag recommendation method includes the following steps: Step 1: Initialize the implicit user feature matrix, implicit item feature matrix, implicit user-oriented label feature matrix, and implicit item-oriented label feature matrix using Gaussian priors; Step 2: On the user-tag interaction graph and the item-tag interaction graph, respectively, use a graph convolutional neural network to obtain high-order embedding representations of users, items, user-oriented tags, and item-oriented tags; Step 3: For the user-tag interaction graph, after each convolution operation, noise is injected into the higher-order embedding representations of the user and the user-label, resulting in noisy embedding representations of the user and the user-label. Based on the noisy embedding representations of the user and the user-label, the contrastive loss is calculated using Info_NEC. For the item-tag interaction graph, after each convolution operation, noise is injected into the higher-order embedding representations of the item and the item-label, resulting in noisy embedding representations of the item and the item-label. Based on the noisy embedding representations of the item and the item-label, the contrastive loss is calculated using Info_NEC. Step 4: Construct an objective function by integrating contrastive loss on top of the Bayesian personalized ranking loss, and learn the model parameters through the Adam optimizer; Step 5: Predict tag ratings by summing the inner product of the embeddings of user and user-oriented tags with the inner product of the embeddings of item and item-oriented tags, and provide personalized tag recommendations based on the predicted ratings. Step 2 specifically includes the following steps: Step 2-1: Construct an interaction matrix between users and tags based on the interaction information of users, items, and tags. Interaction matrix of items and tags If user u and label t have interacted, then the user-label interaction matrix is... The value in the u-th row and t-th column of the matrix is ​​1, otherwise it is 0. If item i and tag t have interacted, then the interaction matrix between the item and tag is... The value in the i-th row and t-th column is 1, otherwise it is 0; Step 2-2: Utilize the user-tag interaction matrix Interaction matrix of items and tags Construct the corresponding adjacency matrix and : ， ， in, and They are respectively and The transpose of the matrix; Steps 2-3: Based on the adjacency matrix , Construct a Laplace matrix of users and tags Laplace matrix of items and tags : ， ， in, It is a diagonal matrix of users and tags. This is a diagonal matrix of items and labels, represented as: ， ， in, Indicates user The number of first-order neighbors, Represents items The number of first-order neighbors, Indicates label The number of first-order neighbors; Steps 2-4: Based on the Laplace matrix of users and tags Laplace matrix of items and tags We construct higher-order embedding representations of users, items, user-oriented tags, and item-oriented tags, as follows: ， ， in, = = This is the initial embedding representation. Indicates user The initial embedding representation, Represents items The initial embedding representation, Indicates label The initial embedding representation; Steps 2-5: After L layers of convolution, the final embedding representation is obtained, as follows: ， ， Will The final embedded representation of all users is obtained by segmenting based on the number of users and the number of tags. and all user-facing tag embedding representations ,Will The final embedded representation of all items is obtained by segmenting based on the number of items and the number of tags. and the embedded representation of each item tag ; Step 3 specifically includes the following steps: Step 3-1: For the user-label interaction graph, inject noise into the higher-order embedding representations of the user and user-label after each convolution operation, and finally obtain the noisy embedding representations of the user and user-label, as follows: ， ， ， ， in, and Describes the noise vector and , Indicates the injection of the first type of noise An embedded representation of all users and tags. Indicates the injection of a second type of noise An embedded representation of all users and tags. Indicates the injection of the first type of noise An embedded representation of all items and tags. Indicates the injection of a second type of noise An embedded representation of all items and tags. Step 3-2: After L layers of convolution, the final embedded noise representation is obtained, as follows: ， ， ， ， Will( and The final embedded noise representation is obtained by segmenting based on the number of users and the number of tags. ,Will and The final embedded noise representation is obtained by segmenting based on the number of items and tags. and ; Step 3-3, in Based on this, the contrast loss of the user is calculated using Info_NEC. ,as follows: ， in, Indicates the sampling batch. This represents the regularized embedding representation of user u under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of user u under the injection of the second type of noise, i.e. = , Indicates the temperature coefficient; Steps 3-4, in Based on this, the contrast loss of user-oriented labels is calculated using Info_NEC. ,as follows: ， in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = ; Steps 3-5, in Based on this, the comparative loss of the item is calculated using Info_NEC. ,as follows: ， in, The embedding representation of item i after regularization under the injection of the first type of noise is, i.e. , This represents the regularized embedding representation of item i under the injection of the second type of noise, i.e. = ; Steps 3-6, in Based on this, the contrastive loss for item-oriented tags is calculated using Info_NEC. ,as follows: ， in, This represents the regularized embedding representation of label t under the injection of the first type of noise, i.e. , This represents the regularized embedding representation of label t under the injection of the second type of noise, i.e. = ; Steps 3-7, according to Calculate contrast loss : ； Step 4 specifically includes the following steps: Step 4-1, according to and The final embedding representations of user u, item i, and tag t (user-oriented and item-oriented) are obtained as follows: ， ， ， ， Step 4-2, according to Obtain the probability score of user u tagging item i with t. ,as follows: ， in, Represents the dot product of vectors; Step 4-3: The objective formula for personalized tag recommendation based on contrastive learning is as follows: ， in, It is the sigmoid function. , The score represents the label of the negative sample. This represents the coefficient controlling for the contrastive learning loss. Represents the regularization coefficient. These are the model's parameters, minimized by the Adam optimizer. and learn model parameters The implicit user feature matrix is ​​initialized using a Gaussian prior. Implicit item feature matrix Implicit user-oriented tag feature matrix and implicit item-oriented label feature matrix , Indicates the embedding dimension. These represent the number of users, items, and tags, respectively.

2. The personalized tag recommendation method based on contrastive learning according to claim 1, characterized in that: Step 5, which uses the inner product of the embeddings of user and user-oriented tags with the inner product of the embeddings of item and item-oriented tags to predict the tag rating, specifically involves: Using the learned user embedding representation User-oriented tag embedding representation Embedded representation of items and item-oriented tag embedding representation Calculate the predicted rating for user u to tag item i with t. ,as follows: ， After calculating predicted scores for all tags, the tags with the highest scores are recommended to users.

3. A personalized tag recommendation method based on contrastive learning according to claim 1 or 2, characterized in that: Step 1 specifically includes the following steps: Step 1-1: Initialize the implicit user feature matrix using Gaussian priors. Implicit item feature matrix Implicit user-oriented tag feature matrix and implicit item-oriented label feature matrix ,as follows: ， ， ， ， In the formula, Indicates the embedding dimension. These represent the number of users, items, and tags, respectively. Steps 1-2: Given user, item, and tag ID information, use the user feature matrix... Item feature matrix User-oriented tag feature matrix and item-oriented label feature matrix To acquire users Embedded representation of item i and tag t facing the user and tag t facing the item: ， ， ， ， in, , The operation means extracting the corresponding embedding vector from the corresponding embedding representation matrix based on the entity's index information.

Images