A Personalized Recommendation Method for Fair Representation Learning of Exposure

Abstract

The present invention relates to a personalized recommendation method for exposure-fair representation learning, belonging to the technical fields of recommendation systems and deep learning. It includes the following steps: obtaining e-commerce data, constructing an e-commerce data set, and dividing it into a training set and a test set; constructing a personalized recommendation model for exposure-fair representation learning, the model including an initial embedding module, an encoder learning module, and an exposure fairness module; training the personalized recommendation model for exposure-fair representation learning with the e-commerce data in the training set; optimizing the model during the training process by using a loss function and the Adam optimizer to obtain a trained model; inputting the e-commerce data in the test set into the trained model to obtain personalized recommendation results. The present invention can not only provide personalized recommendations, but also significantly reduce the bias in the recommendation system, improve the user experience and the exposure fairness of the system.

Description

Technical Field

[0001] The present invention belongs to the technical fields of recommendation systems and deep learning, and particularly relates to a personalized recommendation method for exposure fairness representation learning. Background Art

[0002] With the rapid development of Internet technology, the explosion of data volume has made information overload an inevitable challenge. To help users efficiently find the content they are interested in among the vast amount of information, recommendation systems have emerged. These systems extract the most relevant information from the huge data through various algorithms and recommend it to users to improve the user experience and satisfaction. In recent years, many recommendation algorithms have been continuously innovated to meet the needs of different scenarios, including item-based collaborative filtering, neural network-based, graph model-based, variational autoencoder-based, and contrastive learning-based methods. The continuous progress of these technologies has significantly improved the recommendation effect, making personalized recommendation systems widely used.

[0003] However, with the development and application of recommendation systems, fairness issues have gradually emerged. In traditional recommendation systems, algorithms often focus on optimizing the accuracy of recommendations while ignoring the fairness of recommendation results among user groups and items. This unfairness may lead to the information cocoon phenomenon, restricting users' access to diverse information and thus inhibiting the discovery of their potential interests. In addition, the imbalance in exposure rates among items may affect the revenues of underlying merchants, which in turn has a negative impact on the platform's ecosystem. Therefore, how to ensure the exposure fairness of the system while improving the recommendation effect has become an important topic in the research of recommendation systems.

[0004] To address this challenge, exposure fairness representation learning has gradually become a research hotspot. This method reduces bias in the recommendation process by learning the exposure fairness representations of users and items. However, most existing recommendation methods do not pay attention to the exposure fairness of items and lack consideration of the exposure fairness requirements. In fact, more fair exposure is needed among different items. Therefore, in practical applications, recommendation systems need to strike a balance between personalization and exposure fairness.

[0005] In addition, when recommendation systems achieve the goal of exposure fairness, they also face the problem of data sparsity. This problem is particularly evident in the recommendation for cold items or users in minority groups, and traditional methods are difficult to ensure the fair exposure of these groups. In recent years, some studies have tried to study the exposure fairness of recommendation systems by using exposure fairness metrics. However, these metrics are usually non-differentiable, which limits their application in model training. Summary of the Invention

[0006] To solve the above problems, the present invention provides a personalized recommendation method for exposure fairness representation learning.

[0007] To achieve the above object, the present invention is implemented through the following technical solutions:

[0008] The present invention provides a personalized recommendation method for exposure fairness representation learning, including the following steps:

[0009] S1. Obtain e-commerce data, construct an e-commerce data set, and divide it into a training set and a test set;

[0010] S2. Construct a personalized recommendation model for exposure fairness representation learning, the model includes an initial embedding module, an encoder learning module, and an exposure fairness module; train the personalized recommendation model for exposure fairness representation learning with the e-commerce data in the training set;

[0011] S3. Use a loss function and optimize the model during the training process through an Adam optimizer to obtain a trained model;

[0012] S4. Input the e-commerce data in the test set into the trained model to obtain personalized recommendation results.

[0013] Further, step S1 specifically includes:

[0014] The items interacted by users in the e-commerce data set are divided according to the leave-one-out method, and the last interaction of each user is retained as the test set, and the remaining data is used as the training set;

[0015] The e-commerce data set includes a user set U composed of M users, an item set I composed of N items, and their interaction behaviors form an interaction matrix Y.

[0016] Further, the initial embedding module in step S2 specifically includes:

[0017] In the training set, the position labels of users in set U and items in set I are respectively set as user IDs and item IDs, and the user IDs and item IDs are respectively mapped into one-hot vector matrices and , each user and item are represented by d-dimensional vectors to obtain user vectors and item vectors, and the user vectors and item vectors are vertically concatenated in sequence to obtain a user embedding matrix and an item embedding matrix , and the Xavier method is used to initialize the parameters of the user embedding matrix and the item embedding matrix ;

[0018] According to the one-hot vector matrix mapped by the user ID and the user embedding matrix , the initial user embedding is obtained, and the calculation formula is as follows:

[0019] ,

[0020] where, represents the one-hot vector of the \(u\)-th user in the matrix , represents the dot product operation between the matrix and the vector, , represents the embedding dimension;

[0021] According to the one-hot vector matrix mapped by the item ID and the item embedding matrix , the initial item embedding is obtained, and the calculation formula is as follows:

[0022] ,

[0023] where, represents the one-hot vector of the \(i\)-th item in the matrix , represents the dot product operation between the matrix and the vector, .

[0024] Furthermore, the encoder learning module in step S2 specifically includes:

[0025] The embedding of the user at each layer is obtained through the graph encoder, and the formula is expressed as follows:

[0026] ,

[0027] where, represents the graph convolution result of the user embedding at the -th layer, , is the current graph convolution layer number, \(L\) is the total number of graph convolution layers, represents the graph convolution result of the item embedding at the -th layer, represents the set of items interacting with the user , represents the set of users interacting with item \(i\), represents the size of the set;

[0028] Aggregate the user embeddings of each layer to obtain the final user embedding , and the formula is expressed as follows:

[0029] ,

[0030] where, is the total number of graph convolution layers, Denote the graph convolution result of the user embedding at the th layer;

[0031] The embedding of the item at each layer is obtained through the graph encoder, and the formula is as follows:

[0032] ,

[0033] where, The graph convolution result of the item embedding at the th layer, is the current graph convolution layer number, L is the total number of graph convolution layers, Denote the graph convolution result of the user embedding at the th layer, denotes the set of items interacting with the user , denotes the set of users interacting with item i, denotes the size of the set;

[0034] Aggregate the item embeddings at each layer to obtain the final item embedding , and the formula is as follows:

[0035] ,

[0036] where, is the graph convolution layer number, Denote the graph convolution result of the user embedding at the th layer.

[0037] Furthermore, the exposure fairness module in step S2 specifically includes:

[0038] Perform "pre-recommendation" in the exposure fairness module. Through the final user embedding and the item embedding After the vector dot product operation, obtain the preference degree score of user u for item i. The calculation formula is as follows:

[0039] ,

[0040] where, denotes the vector dot product operation, denotes the preference degree score of user u for item i;

[0041] When performing "pre-recommendation", calculate the (k + 1)-th largest score among all the scores given by user u to all items through the formula . The items with scores greater than the (k + 1)-th largest score are the recommended items. Among them, denotes the preference degree of the user for all items, denotes finding the k-th largest value from the set;

[0042] Map the user's rating to the exposure of the item to obtain the exposure of item \(i\) for user \(u\). The formula is as follows:

[0043] ,

[0044] where, represents the rectified linear unit function, represents the sigmoid function, \(k\) represents that the top \(k\) items play a role when calculating the metrics, represents the temperature coefficient;

[0045] Sum up the exposure of item \(i\) for user \(u\) to obtain the total exposure of the item. The formula is as follows;

[0046] ,

[0047] where, represents the total exposure of the item;

[0048] Calculate the exposure fairness metric based on the total exposure of the item. The formula is as follows:

[0049] ,

[0050] where, represents the set of the total exposures of all items, represents any differentiable exposure fairness metric calculation function.

[0051] Furthermore, step S3 specifically includes:

[0052] Adopt the total loss to optimize the model to obtain the trained model. The formula is as follows:

[0053] ,

[0054] ,

[0055] ,

[0056] where, represents the BPR loss, represents the training set, , represents observing the interaction set existing between user and item , represents user sampling an unobserved item in the interaction set, represents the Sigmoid activation function; Represents the fairness loss, and respectively represent the set of differentiable exposure fairness metric calculation functions where the larger the function value, the more exposure fair the recommendation system is, and the smaller the function value, the fairer the recommendation system is. Represents the function value calculated by the exposure fairness metric function fair of the recommendation system; Represents the total loss, Represents the hyperparameter.

[0057] The present invention also provides a system applying the above personalized recommendation method for exposure fairness representation learning, including:

[0058] Data acquisition module: used to acquire e-commerce data, construct an e-commerce data set, and divide it into a training set and a test set;

[0059] Model construction module: used to construct a personalized recommendation model for exposure fairness representation learning, the model includes an initial embedding module, an encoder learning module, and an exposure fairness module; and train the personalized recommendation model for exposure fairness representation learning with the e-commerce data in the training set.

[0060] Model optimization module: used to optimize the model during the training process using the loss function and through the Adam optimizer to obtain a trained model;

[0061] Recommendation module: input the e-commerce data in the test set into the trained model to obtain personalized recommendation results.

[0062] The advantages of the present invention are:

[0063] The present invention proposes a personalized recommendation method for exposure-fair representation learning, which consists of an encoder learning module and a fairness module. First, the encoder is used to learn user embeddings and item embeddings in each behavior. On this basis, the preferences of users for items are calculated, and a differentiable fairness loss is calculated according to these preferences and used as part of the final loss to train the model. Finally, after optimization, the user-item embeddings become more fair, and such embeddings are more exposure-fair for subsequent recommendation tasks. Therefore, the present invention can not only provide personalized recommendations, but also significantly reduce the bias in the recommendation system, improve the user experience and the exposure fairness of the system. By introducing exposure-fair constraints in the process of user and item embeddings, the over-recommendation of items is restricted, ensuring that the recommendation system reduces the unfair impact on the exposure of different items without sacrificing the recommendation effect. This method uses a deep neural network to learn the embedding representations of users and items, and introduces exposure-fairness constraints in the embedding process to ensure that the model has the ability to represent each item fairly in terms of exposure when processing user behavior data. Using the embedding representations adjusted by exposure fairness for recommendation prediction effectively improves the overall performance and exposure fairness of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

[0065] Figure 1 is a flowchart of the method of the present invention;

[0066] Figure 2 is a graph of the balance between exposure fairness and performance of the present invention;

[0067] Figure 3 is a comparison of different exposure fairness metrics of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] Embodiment 1

[0070] In this embodiment, as Figure 1 shown, the present invention provides a personalized recommendation method for exposure-fair representation learning, and the specific steps include:

[0071] S1. Obtain e-commerce data, construct an e-commerce dataset, and divide it into a training set and a test set;

[0072] Specifically, the items interacted by users in the e-commerce dataset are divided according to the leave-one-out method. The last interaction of each user is retained as the test set, and the remaining data is used as the training set. The e-commerce dataset includes a user set U composed of M users and an item set I composed of N items, and their interaction behaviors form an interaction matrix Y.

[0073] S2. Construct a personalized recommendation model for exposure fairness representation learning. The model includes an initial embedding module, an encoder learning module, and an exposure fairness module; train the personalized recommendation model for exposure fairness representation learning with the e-commerce data in the training set;

[0074] Specifically, the initial embedding module: in the training set, set the position labels of users in the set U and items in the set I as user IDs and item IDs respectively, and map the user IDs and item IDs into one-hot vector matrices and , represent each user and item with a d-dimensional vector to obtain a user vector and an item vector, vertically splice the user vector and the item vector in order to obtain a user embedding matrix and an item embedding matrix , and use the Xavier method to initialize the parameters of the user embedding matrix and the item embedding matrix ;

[0075] According to the one-hot vector matrix mapped by the user ID and the user embedding matrix , obtain the user initial embedding, and the calculation formula is as follows:

[0076] ,

[0077] where represents the u-th user's one-hot vector of the matrix , represents the dot product operation between the matrix and the vector, , represents the embedding dimension;

[0078] According to the one-hot vector matrix mapped by the item ID and the item embedding matrix , obtain the item initial embedding, and the calculation formula is as follows:

[0079] ,

[0080] Among them, represents the one-hot vector of the i-th item of the matrix , represents the dot product operation between the matrix and the vector, .

[0081] Specifically, the encoder learning module: obtains the embedding of the user at each layer through the graph encoder, and the formula is expressed as follows:

[0082] ,

[0083] Among them, represents the graph convolution result of the user embedding at the -th layer, , is the current graph convolution layer number, L is the total number of graph convolution layers, represents the graph convolution result of the item embedding at the -th layer, represents the set of items interacting with the user , represents the set of users interacting with item i, represents the size of the set;

[0084] Aggregates the user embeddings of each layer to obtain the final user embedding , and the formula is expressed as follows:

[0085] ,

[0086] Among them, is the total number of graph convolution layers, represents the graph convolution result of the user embedding at the -th layer;

[0087] Obtains the embedding of the item at each layer through the graph encoder, and the formula is expressed as follows:

[0088] ,

[0089] Among them, the graph convolution result of the item embedding at the -th layer, is the current graph convolution layer number, L is the total number of graph convolution layers, represents the graph convolution result of the user embedding at the -th layer, represents the set of items interacting with the user , represents the set of users interacting with item i, represents the size of the set;

[0090] Aggregate the item embeddings for each layer to obtain the final item embedding , which is expressed by the following formula:

[0091] ,

[0092] where is the number of layers of graph convolution, represents the result of graph convolution of the user embedding at the -th layer.

[0093] Specifically, the exposure fairness module: perform "pre-recommendation" in the exposure fairness module, and through the final user embedding and item embedding perform a vector dot product operation to obtain the preference score of user u for item i. The calculation formula is expressed as follows:

[0094] ,

[0095] where represents the vector dot product operation, represents the preference score of user u for item i;

[0096] When performing "pre-recommendation", calculate the (k + 1)-th largest score among all the scores given by user u to all items through the formula . Items with scores greater than the (k + 1)-th largest score are recommended items. Among them, represents the preference of the user for all items, represents finding the k-th largest value from the set;

[0097] Map the user's scores to the exposure situation of the items to obtain the exposure volume of item i for user u. The formula is expressed as follows:

[0098] ,

[0099] where represents the rectified linear unit function, represents the sigmoid function, k represents that the top k items play a role when calculating the metrics, represents the temperature coefficient;

[0100] Perform a summation operation on the exposure volume of item i for user u to obtain the total exposure volume of the item. The formula is expressed as follows;

[0101] ,

[0102] where represents the total exposure volume of the item;

[0103] Calculate the exposure fairness index based on the total exposure of items, and the formula is as follows:

[0104] ,

[0105] where, represents the set of the total exposures of all items, represents any differentiable exposure fairness index calculation function.

[0106] S3. Use the loss function and optimize the model during the training process through the Adam optimizer to obtain the trained model;

[0107] Specifically, use the total loss to optimize the model to obtain the trained model, and the formula is as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] where, represents the BPR loss, represents the training set, , represents observing the set of interactions existing between user and item , represents the item sampled by user that is not observed in the interaction set, represents the Sigmoid activation function; represents the fairness loss, , respectively represent the sets of differentiable exposure fairness index calculation functions where the larger the function value, the more exposure fair the recommendation system is, and the smaller the function value, the fairer the recommendation system is, represents the function value calculated by the exposure fairness index function fair of the recommendation system; represents the total loss, represents the hyperparameter.

[0112] S4. Input the e-commerce data in the test set into the trained model to obtain the personalized recommendation results.

[0113] Example 2

[0114] This example provides a system applying the above personalized recommendation method for exposure fairness representation learning, including:

[0115] Data acquisition module: used to acquire e-commerce data, construct an e-commerce data set, and divide it into a training set and a test set;

[0116] Model construction module: used to construct a personalized recommendation model for exposure fairness representation learning, the model includes an initial embedding module, an encoder learning module, and an exposure fairness module; and train the personalized recommendation model for exposure fairness representation learning with the e-commerce data in the training set;

[0117] Model optimization module: used to optimize the model during the training process using a loss function and the Adam optimizer to obtain a trained model;

[0118] Recommendation module: input the e-commerce data in the test set into the trained model to obtain personalized recommendation results.

[0119] Embodiment 3

[0120] In this embodiment, we further introduce the Ml-1m data set, which is one of the most classic data sets in the field of recommendation systems. First, divide the Ml-1m data set into a training set and a test set, and use the data in the training set to construct an interaction matrix and initial user-item embeddings. Secondly, apply the encoder of the interaction matrix to perform graph convolution on the initial user-item embeddings, and use the results of the graph convolution to obtain the embeddings directly used for recommending user items. Then, use the embeddings to calculate the recommendation loss and the exposure fairness loss respectively, and use the gradient descent method to optimize the model to obtain the final embeddings. Finally, use the final embeddings for recommendation to generate a recommendation list.

[0121] As Figure 2 and Figure 3, taking the LightGCN model as an example, we conduct extensive parameter experiments on the LightGCN after introducing the module and compare it with the original LightGCN model. In the two figures, REL and FAIR represent the accuracy evaluation index and the exposure fairness index respectively. Among them, {↑ / ↓}{type}@{k} represents the final result of the type index among the top k recommended items given to the user. The arrow's direction indicates how the index changes better. The type of type can be hit representing the hit rate, precision representing the precision rate, recall representing the recall rate, ndcg representing the normalized discounted cumulative gain, Jain representing the Jain fairness index, Ent representing the entropy, Gini representing the Gini coefficient, VoCD representing the customer voice data. Among them, hit, precision, recall, and ndcg are accuracy indicators, and Jain, Ent, Gini, and VoCD are exposure fairness indicators. For example, ↑hit@10 represents the calculation result of the hit index among the top 10 recommended items given to the user, and the larger the index, the better. There are three colors in the figure, namely red, green, and gray. Red represents a performance decline, green represents a performance improvement, and gray represents the loss of improvement in the exposure fairness index when using the module.

[0122] Figure 2 shows four representative experimental results after extensive parameter experiments by introducing our module (taking Ent as an example) into LightGCN. They are the best accuracy effect without considering the fairness of the model (i.e., the highest accuracy), the best fairness effect when ensuring 100% accuracy (ensuring 100% accuracy), the best fairness effect when ensuring 90% accuracy (ensuring 90% accuracy), and the best fairness effect when ensuring 80% accuracy (ensuring 80% accuracy). We conclude that by changing the training parameters, the balance between exposure fairness and accuracy can be changed, thus meeting different actual needs.

[0123] Figure 3 uses different indicators (VoCD, Gini, Ent, Jain) to improve the performance change of the LightGCN model. When using the module, there will also be an improvement in accuracy.

[0124] In summary, compared with the original LightGCN model, the present invention not only has a 2% improvement in accuracy and different degrees of improvement in exposure fairness in the optimal case, but also the present invention can meet the exposure fairness requirements in different situations, such as ensuring a large improvement in exposure fairness with 90% of the recommended performance. In addition, the present invention has good portability and can be transplanted into other models to improve exposure fairness, including but not limited to various models such as BPR, NeuMF, SGL, MultiVAE, SLIM, etc.

[0125] Example 4

[0126] In this embodiment, assume that we have a movie recommendation platform where users watch movies on the platform and interact with movies. We first collect the users' viewing history and preprocess the data.

[0127] Then, create user-item embeddings and design an encoder to capture the interaction information between users and movies, and use an exposure fairness module to ensure that the user-item embeddings learn exposure information. For example, if Movie 1, Movie 2, and Movie 3 are recommended 5, 3, and 1 times respectively, and Movie 3 is suffering from unfair exposure, through our module, we can guide the model to give more exposure to Movie 3.

[0128] By optimizing the model, minimizing the exposure fairness loss, ensuring that the recommendation results are more exposure-fair, and at the same time maximizing the information between the user's embedding and their movie preferences, ensuring the accuracy of the recommendation results. Use the training dataset to train the model, optimize the parameters of the sensitive attribute encoder and the interest encoder, and further improve the performance of the model through cross-validation and hyperparameter tuning.

[0129] Finally, based on the user's exposure fairness embedding, calculate the similarity between him and all movies on the platform, and generate a personalized movie recommendation list according to the similarity. For example, recommend "Movie 1", "Movie 2", and "Movie 3". Through the above steps, we can achieve a personalized movie recommendation with exposure fairness representation learning, ensuring the fairness and accuracy of the recommendation results.

[0130] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A personalized recommendation method based on exposure fair representation learning, characterized in that: The following steps are involved: S1. Obtain e-commerce data, build an e-commerce dataset, and divide it into training set and test set; S2. Constructing a personalized recommendation model for exposure fairness representation learning, the model includes an initial embedding module, an encoder learning module and an exposure fairness module; The personalized recommendation model is trained by learning fair representations through e-commerce data in the training set; The initial embedding modules include: In the training set, the position labels of users in set U and items in set I are set as user ID and item ID respectively, and the user ID and item ID are mapped into one-hot vector matrices respectively. and , each user and item is represented by a d-dimensional vector, and the user vector and item vector are obtained. The user vector and item vector are spliced ​​vertically in sequence to obtain the user embedding matrix and the item embedding matrix , and use the Xavier method to embed the user matrix and the item embedding matrix Initialize parameters; The specific steps in the encoder learning module are as follows: The user's embedding at each layer is obtained through the graph encoder, and the formula is as follows: ， in, Indicates that the user embeds The graph convolution result of the layer, , is the number of current graph convolution layers, L is the total number of graph convolution layers, Indicates that the item is embedded in The graph convolution result of the layer, Represents user A collection of interactive items. represents the set of users who interact with item i, Indicates the size of the collection; Aggregate the user embeddings of each layer to get the final user embedding , the formula is as follows: ， in, is the total number of graph convolution layers, Indicates that the user embeds Graph convolution result of the layer; The embedding of the object at each layer is obtained through the graph encoder, and the formula is as follows: ， in, Items embedded in The graph convolution result of the layer, is the number of current graph convolution layers, L is the total number of graph convolution layers, Indicates that the user embeds The graph convolution result of the layer, Represents user A collection of interactive items. represents the set of users who interact with item i, Indicates the size of the collection; Aggregate the item embeddings of each layer to get the final item embedding , the formula is as follows: ， in, is the number of graph convolution layers, Indicates that the user embeds Graph convolution result of the layer; S3. Use the loss function and the Adam optimizer to optimize the model during the training process to obtain a trained model; S4. Input the e-commerce data in the test set into the trained model to obtain personalized recommendation results.

2. The personalized recommendation method for exposure fair representation learning according to claim 1, characterized in that: Step S1 specifically includes: The items interacted by users in the e-commerce dataset are divided according to the leave-one-out method, and the last interaction of each user is retained as a test set, and the rest of the data is used as a training set; The e-commerce data set includes M users constituting a user set U, N items constituting an item set I, and the interaction behaviors between the two constitute an interaction matrix Y.

3. The personalized recommendation method for exposure fair representation learning according to claim 2, characterized in that: The initial embedding module in step S2 further includes: One-hot vector matrix mapped according to the user ID and the user embedding matrix , get the user's initial embedding, the calculation formula is as follows: ， in, Representation Matrix The one-hot vector of the u-th user, Represents the dot product operation between a matrix and a vector, , represents the embedding dimension; One-hot vector matrix mapped according to the item ID and the item embedding matrix , get the initial embedding of the item, the calculation formula is as follows: ， in, Representation Matrix The one-hot vector of the ith item, Represents the dot product operation between a matrix and a vector, .

4. The personalized recommendation method for exposure fair representation learning according to claim 3, characterized in that: The exposure fairness module in step S2 specifically includes: In the exposure fairness module, "pre-recommendation" is performed, through the final user embedding and item embedding After the vector dot multiplication operation, we get the preference score of user u for item i. The calculation formula is as follows: ， in, Represents a vector dot product operation, Represents the preference rating of user u for item i; When making a "pre-recommendation", the formula Calculate the k+1th largest score given by user u among all items. The item with a score greater than the k+1th largest score is the recommended item, where: Indicates the user's preference for all items. It means to find the kth largest value from the set; Map the user's rating to the item's exposure to get the exposure of item i to user u. The formula is as follows: ， in, represents the rectified linear unit function, represents the sigmoid function, k represents the selection of the first k items to be used in calculating the index, represents the temperature coefficient; The exposure of the item i to the user u is summed up to obtain the total exposure of the item, which is expressed as follows: ， in, Indicates the total exposure of the item; The exposure fairness index is calculated based on the total exposure of the item. The formula is as follows: ， in, Represents the total exposure of all items. represents any differentiable exposure fairness indicator calculation function.

5. The personalized recommendation method for exposure fair representation learning according to claim 4, characterized in that: Step S3 specifically includes: The total loss optimization model is used to obtain the trained model. The formula is as follows: ， ， ， in, represents the BPR loss, represents the training set, , Indicates that the user is observed and items The set of interactions that exist between Indicates user Sampling unobserved items in the interaction set , Represents the Sigmoid activation function; represents fairness loss, , The set of derivable exposure fairness index calculation functions that represent the larger the function value, the more fair the recommendation system is, and the smaller the function value, the fairer the recommendation system is. Represents the calculated function value of the exposure fairness indicator function fair of the recommendation system; represents the total loss, represents a hyperparameter.

6. A system using the personalized recommendation method for learning exposure fair representation according to claim 1, characterized in that: include: Data acquisition module: used to acquire e-commerce data, build e-commerce data sets, and divide them into training sets and test sets; Model building module: used to build a personalized recommendation model for learning exposure fairness representation, the model includes an initial embedding module, an encoder learning module and an exposure fairness module; and train the personalized recommendation model for learning exposure fairness representation using e-commerce data in the training set; Model optimization module: used to optimize the model in the training process using the loss function and the Adam optimizer to obtain a trained model; Recommendation module: Input the e-commerce data in the test set into the trained model to obtain personalized recommendation results.

Images