Fair recommendation method for dissociating sensitive information from double view angles

Abstract

The invention provides a double-view fair recommendation method. The fairness and accuracy of recommendation are improved by dissociating sensitive information. The method starts from two perspectives: one perspective is based on a real sensitive attribute of a user, and the other perspective is an anti-fact hypothesis perspective based on the change of the sensitive attribute of the user. The method comprises the following steps: firstly, constructing four representation vectors which are respectively used for performance orientation, true sensitivity, anti-fact sensitivity and fairness; these representations are generated and optimized by different algorithms to ensure that sensitive information is removed and user preferences are preserved. In particular, the fair representation enables dissociation of information by minimizing mutual information and cosine similarity differences between the fair representation and the sensitive representation. And meanwhile, mutual information between the fair representation and the performance oriented representation is maximized, so that the preference of the user is effectively captured. Finally, according to the method, the interaction score of the user and the project is calculated through fair representation, a recommendation result is generated, and the fairness of recommendation is evaluated with the assistance of fairness measurement.

Description

Technical Field

[0001] The present invention relates to technical fields such as recommendation systems, recommendation fairness and fair representation learning, and in particular to a fair recommendation method for dissociating sensitive information from dual perspectives. Background Art

[0002] As an important application of artificial intelligence technology, the recommendation system is widely used in e-commerce, audio and video, and social media. In the era of information explosion, the recommendation system plays a vital role. On the one hand, it helps users find valuable information for themselves, and on the other hand, it allows information to be displayed to users who are interested in it. From the user's perspective, it solves the problem of how to efficiently obtain information of interest in the case of information overload. From the perspective of the enterprise, the recommendation system solves the problem of maximizing the attraction and retention of users, thereby achieving the company's business goals. Although automatic recommendation provides convenience for people, it also has negative effects. On e-commerce platforms, the products recommended by the platform will affect consumers' choices, and consumers' choices are crucial to the interests of the platform and product providers. On media websites, the information that users can receive mainly depends on the recommended content, which not only affects individual decisions and actions, but also affects the interests of the entire social group.

[0003] However, recommendation systems often show potential bias and unfairness in practical applications. Since recommendation systems are data-driven, they inherit data biases related to specific sensitive factors, such as gender, age, and activity, and lead to fairness issues in recommendation systems. Moreover, this bias and unfairness stemming from data bias is enhanced in the design of recommendation algorithms. For example, the information recommendation algorithm of a news media platform causes users to only see political news that matches their views, exacerbating the information cocoon. The product recommendation algorithm of an e-commerce platform tends to recommend the platform's best-selling products, resulting in uneven market popularity and the formation of a Matthew effect.

[0004] Fair data representation learning is a common method to improve the fairness of recommendation systems. Its core idea is to blur the information related to sensitive attributes in the data representation while effectively encoding user preference information. Existing fair recommendation methods usually use methods that minimize the correlation between data representation and user sensitive information, such as adversarial learning methods and orthogonal regularization methods, to promote the fairness of recommendation results. Fair representation learning methods based on adversarial learning generate data representations that blur user sensitive attributes by constructing attribute discriminators and alternately training representation generators. Orthogonal regularization-based methods achieve the purpose of improving fairness by making data representation orthogonal to sensitive information representation. These methods still have limitations. On the one hand, due to the lack of consideration of recommendation accuracy, these methods will lead to a decrease in recommendation performance. On the other hand, these methods ignore the various possibilities of bias and unfairness in the recommendation system, so the improvement of recommendation fairness is limited. Summary of the invention

[0005] Existing fair recommendation methods dissociate users’ sensitive information from a single perspective for data representation. For example, these methods minimize the correlation between a male user’s data representation and male sensitive information, but cause the user’s data representation to be associated with female sensitive information. Similarly, in multi-category sensitive attributes, such as occupation, these methods cause a teacher’s data representation to be associated with any other occupation. This means that these methods not only eliminate existing unfairness problems, but also introduce multiple other potential biases, resulting in the recommendation unfairness problem not being effectively solved.

[0006] In view of the shortcomings of the existing fair recommendation methods, the present invention proposes a fair recommendation method for dissociating sensitive information from dual perspectives, and the specific steps are as follows:

[0007] Step 1. Download a public recommendation system dataset containing user attribute information (age, occupation, gender). Preprocess the dataset, remove users and items with less than 20 interactions, and use consecutive numbers to identify users and items. Randomly select 70% of the data as the training set for model training, and the remaining 30% as the validation set to verify the model performance.

[0008] Step 2: Construct four types of representation vectors for users and items, namely performance-oriented representation vector, fairness representation vector, sensitive representation vector, and counterfactual sensitive representation vector. All vectors are randomly initialized according to Gaussian distribution.

[0009] Step 3: Import the training set and construct the user and item triples (user, interacted item, non-interacted item). Use the graph convolution algorithm or matrix decomposition method to generate the representation vectors of users and items, and pre-train a performance-oriented recommendation model in combination with the Bayesian personalized ranking loss function (BPR Loss). Generate the performance-oriented representation vectors of users and items through this performance-oriented recommendation model.

[0010] Step 4. Construct sensitive representation vectors for users and items based on the graph convolution algorithm. Input these vectors into the attribute classifier constructed by a one-layer fully connected neural network to obtain the predicted values ​​of sensitive attributes for users and items. Calculate the cross entropy loss between the predicted values ​​of sensitive attributes and the label values, which is recorded as the true attribute prediction loss. Measure the similarity of users' sensitive attributes and preferences. Users with similar sensitive attributes but dissimilar preferences to the target user are recorded as positive users, and users with dissimilar sensitive attributes but similar preferences to the target user are recorded as negative users, and form a triplet with the target user (target user, positive user, negative user). Promote the true sensitive representation of the target user and the positive user to be close, while promoting the true sensitive representation of the target user and the negative user to be far away. Use the true sensitive representation vector to measure the BPR Loss of the triplet (target user, positive user, negative user), which is recorded as the true sensitive representation loss. Update the parameters of the sensitive representation vectors of users and items by using backpropagation and gradient descent algorithms to minimize the true attribute prediction loss and the true sensitive representation loss.

[0011] Step 5. Construct the counterfactual sensitive representation vectors of users and items based on the graph convolution algorithm. Here, counterfactual means intervening in the sensitive attributes of users, assuming that these sensitive attributes change, and calculating the output of the recommendation model based on this assumption. This method helps to identify potential biases and unfairness in the recommendation system and provides a dual perspective to measure and improve the fairness of recommendation results. Assuming that the user's sensitive attributes change, calculate the counterfactual sensitive attribute label values ​​of users and items. Input the counterfactual sensitive representation vectors of users and items into the attribute classifier whose parameters have been trained and frozen in step 4 to obtain the counterfactual sensitive attribute prediction values ​​of users and items. Calculate the cross entropy loss function value of the counterfactual sensitive attribute prediction value and the counterfactual sensitive attribute label value of users and items, denoted as the counterfactual attribute prediction loss. Import the triple (target user, positive user, negative user) calculated in step 4. Promote the counterfactual sensitive representations of the target user and the positive user to be close, and promote the counterfactual sensitive representations of the target user and the negative user to be far away. Use the counterfactual sensitive representation vector to calculate the BPR Loss of the triple (target user, positive user, negative user), denoted as the counterfactual sensitive representation loss. The parameters of the sensitive representation vectors of users and items are updated by using back-propagation and gradient descent algorithms to minimize the counterfactual attribute prediction loss and the counterfactual sensitive representation loss.

[0012] Step 6. Use the same algorithm as step 3 to construct fair representation vectors for users and items. Use the mutual information estimation upper bound method to calculate the mutual information between the fair representation and the sensitive representation, and the counterfactual sensitive representation, respectively, and record it as the fairness loss. Calculate the absolute difference of the cosine similarity between the fair representation and the sensitive representation, and the counterfactual sensitive representation, and record it as the fair regularization loss. Among them, the mutual information estimation upper bound method uses two neural networks to calculate the mean and variance of the fair representation vector respectively. The mutual information estimation lower bound method is used to calculate the mutual information between the fair representation and the performance-oriented representation, and record it as the preference loss. Calculate the BPR Loss for all triples (users, interacted items, and non-interacted items).

[0013] Step 7: Set appropriate weights for the fairness loss, fairness regularization loss, preference loss, and BPR Loss calculated in step 6 and sum them up. The result is recorded as the joint loss. Update the parameters of the user and item fairness representation vectors through back propagation and gradient descent algorithms to minimize the joint loss.

[0014] Step 8: Import the validation set, use the optimized fair representation vectors of users and items to calculate the interaction prediction probability, and sort them to obtain the TOP-K recommendation list for each user. Evaluate the recommendation effect of the designed method based on the recommendation accuracy and fairness measurement methods.

[0015] The recommendation model designed by the present invention takes into account the various possibilities of causing bias in data representation and comprehensively encodes the user's sensitive information, which enables the recommendation system to more accurately understand and eliminate potential bias and discrimination, and solve the unfairness problem of the recommendation system from the real perspective and the counterfactual perspective respectively. In addition, the present invention applies the mutual information estimation method to the study of fair recommendation. Compared with the existing methods, the mutual information estimation upper bound method can more accurately measure the correlation between data representation and sensitive information. By minimizing the correlation, the recommendation model is prompted to dissociate the sensitive information in the data representation, providing a new method for removing sensitive factors in the data representation in the recommendation system. At the same time, the mutual information estimation lower bound method can accurately measure the correlation between data representation and preference information. By maximizing the correlation, the recommendation model can effectively encode user preference information to ensure recommendation accuracy while improving fairness. DETAILED DESCRIPTION BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of the method of the present invention

[0017] Figure 2 The model framework diagram of the present invention

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in combination with specific examples and with reference to the accompanying drawings.

[0019] See also Figure 1 and Figure 2 The fair recommendation method for separating user sensitive information proposed in the present invention specifically comprises the following steps:

[0020] Step 1: Download the recommendation system datasets MovieLens and Last.FM that contain user attribute information publicly available on the Internet. The datasets mainly contain the interaction information between users and items, as well as the gender, age, and occupation information of users. Filter out users and items with too few interactions from the dataset, and remap the IDs of all users and items to consecutive integers starting from 0 and increasing by 1. Construct a user-item two-dimensional interaction matrix based on the interaction information between users and items, where a row of the matrix represents the interaction feedback between a user and all items, and a column of the matrix represents the interaction feedback between a project and all users.

[0021] Step 2: Construct four different representation vectors for users and projects, namely performance-oriented representation vector and Fair Representation Vector and Sensitive Representation Vector and Counterfactual Sensitive Representation Vector and All representation vectors are randomly initialized according to Gaussian distribution and enhanced by graph convolution method. The process is expressed as:

[0022]

[0023] in, The mean is 0 and the variance is σ 2 Gaussian distribution.

[0024] Step 3: Import the user-item interaction information in the training set of the dataset, and randomly extract an item that the user has not interacted with for each interaction information to form a triple (user, interacted item, non-interacted item). Generate the performance-oriented representation vector of the user and item based on the matrix decomposition method or graph convolution algorithm. The matrix decomposition method selects the 0th layer input as the performance-oriented representation vector of the user and item, that is, and

[0025] The performance-oriented representation vector is pre-trained in combination with BPR Loss, and the performance-oriented representation vector at the best performance moment is stored. The BPR Loss calculation process is as follows:

[0026]

[0027] in, represents the training data set, The performance-oriented representation vector representing user u and item i and The inner product of , S(·) is the sigmoid function, that is

[0028] Step 4: Generate the true sensitive representation vectors of user u and item v using the graph convolution-based method and Input the true sensitive representation vector of user u into the classifier θ constructed by the neural network c Predict the sensitive attribute value of user u in As shown below:

[0029]

[0030] Among them, K represents the dimension of the vector. After that, the cross entropy loss between the predicted value of the sensitive attribute and the true value of the user's sensitive attribute is calculated:

[0031]

[0032] Among them, U represents the user set, C represents the number of user sensitive attribute types, and a u Indicates the real sensitive attribute of user u.

[0033] Calculate the sensitive attribute similarity and preference similarity between users. Specifically, the sensitive attribute similarity is measured by comparing the sensitive attributes of users, and the preference similarity is measured by the performance-oriented representation vector of users. Users with similar sensitive attributes but dissimilar preferences to the target user are sampled as positive users, and users with dissimilar sensitive attributes but similar preferences to the target user are sampled as negative users. The positive user sampling process is expressed as: The negative user sampling process is expressed as: Among them, t1 and t2 are hyperparameters. Construct user similarity triples (target user, positive user, negative user), that is, (u,u i ,u j ), and use the true sensitive representation vector to calculate its BPR Loss, recorded as the true sensitive representation loss. The calculation process is as follows:

[0034]

[0035] in, represents the set of positive users sampled from user u, Represented as the set of negative user samples of user u, the loss function value is minimized by back propagation and gradient descent optimization method and To optimize the true sensitive representation vectors of users and items.

[0036] Step 5: Generate counterfactual sensitive representation vectors of user u and item v using graph convolution-based methods and The neural network classifier θ after training in step 4 c Fixed parameters, expressed as Afterwards, the counterfactual sensitive representation vectors of user u and item v are and Input into the classifier to obtain the predicted value of the user's counterfactual sensitive attribute, as shown below:

[0037]

[0038] Calculate the counterfactual sensitive attribute label value of user u, the formula is as follows:

[0039]

[0040] 1 of them C Represents a vector of all 1s with dimension C. Calculate the counterfactual sensitive attribute label value of item v. The calculation process is as follows:

[0041]

[0042] Calculate the cross entropy loss between the predicted value of the counterfactual sensitive attribute and the user's counterfactual sensitive attribute label value:

[0043]

[0044] Where V represents the set of items. The counterfactual sensitive representation vector parameters of users and items are updated through back propagation and gradient descent optimization methods to minimize the loss function value. and

[0045] Import the user similarity triplet (target user, positive user, negative user) calculated in step 4, and use the counterfactual sensitive representation vector to calculate its BPR Loss, recorded as the counterfactual sensitive representation loss. The calculation process is as follows:

[0046]

[0047] Step 6: Use the same algorithmic paradigm as step 3 for data modeling and reasoning. Specifically, if step 3 uses the graph convolution algorithm, the current step also uses the algorithm. If step 3 uses a matrix decomposition-based method, the current step also uses the matrix decomposition method. The mutual information estimation upper bound method vCLUB is used to calculate the mutual information between the user's fair representation vector and the true sensitive representation vector and the counterfactual sensitive representation vector, that is, the fairness loss. The calculation process is as follows:

[0048]

[0049] Among them, q θ (·) represents the conditional distribution q(·) parameterized by the neural network, and θ represents the parameters of the neural network. This calculation measures the correlation between the fair representation vector and the sensitive information, and minimizing this loss is conducive to dissociating the sensitive information from the fair representation.

[0050] The mutual information between the fair representation vector and the performance-oriented representation vector is calculated through the mutual information estimation lower bound method infoNCE, that is, the preference loss. The calculation process on the user side is shown below:

[0051]

[0052] Among them, α is a hyperparameter. The calculation process on the project side is similar to this:

[0053]

[0054] Finally, the preference loss can be expressed as:

[0055]

[0056] The mutual information between the fair representation vector and the performance-oriented representation vector measures the correlation between the fair representation and the user preference information. Maximizing the mutual information value helps the fair representation encode user preferences, thereby improving the recommendation accuracy.

[0057] Calculate the absolute difference of cosine similarity between the fair representation vector and the true sensitive representation vector and the counterfactual sensitive representation vector, that is, the fair regularization loss. The calculation process is as follows:

[0058]

[0059] Different dissociation effects of sensitive information will be produced between the real and counterfactual perspectives. This loss balances the correlation between fair representation and real sensitive representation and counterfactual sensitive representation to further eliminate potential bias. The BPR Loss of all training data triplets (user, interacted item, non-interacted item) is calculated as follows:

[0060]

[0061] This loss is minimized to improve the accuracy of the fair representation for recommendation.

[0062] Step 7. Combine the losses calculated in step 6 and set appropriate weights for each loss:

[0063]

[0064] where λ v , N , M The parameters of the fair representation vectors of users and items are updated through back-propagation and gradient descent algorithms to minimize the value of the joint loss function.

[0065] Step 8. Import the dataset validation set and use the fair representation vectors of users and items for interaction prediction. The prediction formula is as follows:

[0066]

[0067] For each user, items are sorted according to the interaction probability, and K items with the highest interaction probability are selected to form the TOPK recommendation list of the user. Recall and Normalized Discounted Cumulative Gain (NDCG) are used to evaluate the recommendation accuracy of the user's TOPK recommendation list. The recall rate is calculated as follows:

[0068]

[0069] Among them, TP@K (True Positives) represents the number of true positives in the TOPK recommendation list, and FN (False Negatives) represents the number of false negatives in the TOPK recommendation list. The normalized loss cumulative gain is calculated as follows:

[0070]

[0071] Among them, rel i Represents the true value of the interaction between the user and the i-th item in the TOPK recommendation list, which is 1 if the user has interacted, otherwise 0. * It represents the result of sorting the TOPK recommendation list in descending order according to user preference, that is, in rel * In the , the interacted items are ranked before the uninteracted items, and then similar calculations are performed. In order to evaluate the fairness of the model, two widely used fairness evaluation indicators are selected: DP@K and EO@K. The calculation method of DP@K is as follows:

[0072]

[0073] Among them, D js Represents the JS divergence (Jensen-Shannon Divergence) metric function. G a

[0074] Represents the set of users whose sensitive attribute is a. EO@K is calculated as follows:

[0075]

[0076] in Represents the set of items that user u has interacted with in the test set.

[0077] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A fair recommendation method for dissociating sensitive information from two perspectives, characterized in that The following steps are included: Step 1. Download a public recommendation system dataset containing user attribute information (age, occupation, gender). Preprocess the dataset, remove users and items with less than 20 interactions, and use consecutive numbers to identify users and items. Randomly select 70% of the data as the training set for model training, and the remaining 30% as the validation set to verify the model performance. Step 2: Construct four types of representation vectors for users and items, namely performance-oriented representation vector, fairness representation vector, sensitive representation vector, and counterfactual sensitive representation vector. All vectors are randomly initialized according to Gaussian distribution. Step 3: Import the training set and construct the user and item triples (user, interacted item, non-interacted item). Use the graph convolution algorithm or matrix decomposition method to generate the representation vectors of users and items, and pre-train a performance-oriented recommendation model in combination with the Bayesian personalized ranking loss function (BPR Loss). Generate the performance-oriented representation vectors of users and items through this performance-oriented recommendation model. Step 4: Construct sensitive representation vectors for users and items based on the graph convolution algorithm. Input these vectors into the attribute classifier constructed by a layer of fully connected neural network to obtain the predicted values ​​of sensitive attributes of users and items. Calculate the cross entropy loss between the predicted value of sensitive attributes and the label value, which is recorded as the true attribute prediction loss. Measure the similarity of users' sensitive attributes and preferences. Users with similar sensitive attributes but dissimilar preferences to the target user are recorded as positive users, and users with dissimilar sensitive attributes but similar preferences to the target user are recorded as negative users. Together with the target user, they form a triplet (target user, positive user, negative user). Promote the target user and the true sensitive representation of the positive user to be closer, while promoting the target user and the true sensitive representation of the negative user to be farther away. Use the true sensitive representation vector to measure the BPR Loss of the triplet (target user, positive user, negative user), recorded as the true sensitive representation loss. Update the sensitive representation vector parameters of users and items by using backpropagation and gradient descent algorithms to minimize the true attribute prediction loss and the true sensitive representation loss. Step 5. Construct counterfactual sensitive representation vectors for users and items based on the graph convolution algorithm. Here, counterfactual means intervening in the sensitive attributes of users, assuming that these sensitive attributes have changed, and calculating the output of the recommendation model based on this assumption. This method helps to identify potential biases and unfairness in the recommendation system and provides a dual perspective to measure and improve the fairness of recommendation results. Assuming that the user's sensitive attributes have changed, calculate the counterfactual sensitive attribute label values ​​of the user and item. Input the counterfactual sensitive representation vectors of the user and item into the attribute classifier whose parameters have been trained and frozen in step 4 to obtain the counterfactual sensitive attribute prediction values ​​of the user and item. Calculate the cross entropy loss function value of the counterfactual sensitive attribute prediction value and the counterfactual sensitive attribute label value of the user and item, denoted as the counterfactual attribute prediction loss. Import the triplet (target user, positive user, negative user) calculated in step 4 to promote the counterfactual sensitive representations of the target user and the positive user to be close, and at the same time promote the counterfactual sensitive representations of the target user and the negative user to be far apart. The counterfactual sensitive representation vector is used to calculate the BPR Loss of the triple (target user, positive user, negative user), which is recorded as the counterfactual sensitive representation loss. The sensitive representation vector parameters of users and items are updated by using back propagation and gradient descent algorithms to minimize the counterfactual attribute prediction loss and the counterfactual sensitive representation loss. Step 6. Use the same algorithm as step 3 to construct fair representation vectors for users and items. Use the mutual information estimation upper bound method to calculate the mutual information between the fair representation and the sensitive representation, and the counterfactual sensitive representation, respectively, and record it as the fairness loss. Calculate the absolute difference of the cosine similarity between the fair representation and the sensitive representation, and the counterfactual sensitive representation, and record it as the fair regularization loss. Among them, the mutual information estimation upper bound method uses two neural networks to calculate the mean and variance of the fair representation vector respectively. The mutual information estimation lower bound method is used to calculate the mutual information between the fair representation and the performance-oriented representation, and record it as the preference loss. Calculate the BPR Loss for all triples (users, interacted items, and non-interacted items). Step 7: Set appropriate weights for the fairness loss, fairness regularization loss, preference loss, and BPR Loss calculated in step 6 and sum them up. The result is recorded as the joint loss. Update the parameters of the user and item fair representation vectors through back propagation and gradient descent algorithms to minimize the joint loss. Step 8: Import the validation set, use the optimized fair representation vectors of users and items to calculate the interaction prediction probability, and sort them to obtain the TOP-K recommendation list for each user. Evaluate the recommendation effect of the designed method based on the recommendation accuracy and fairness measurement methods.

2. According to claim 1, a fair recommendation method for dissociating sensitive information from two perspectives is characterized in that: In step 2, four different functional representation vectors are constructed for users and projects respectively - performance-oriented representation vectors and Fair Representation Vector and True sensitive representation vector and Counterfactual Sensitive Representation Vector and The build process is as follows: Among them, * is a wildcard, which can represent any of the four representation vectors. Indicates the initialized user representation, The mean is 0 and the variance is σ 2 Gaussian distribution.

3. According to claim 1, a fair recommendation method for dissociating sensitive information from two perspectives is characterized in that: In step 4, the user's true sensitive representation vector is used to predict the user's sensitive attributes. The prediction process formula is expressed as: Among them, U represents the user set, C represents the number of user sensitive attribute types, and a u represents the real sensitive attribute of user u, θ c It is a classifier constructed by a neural network.

4. The fair recommendation method for dissociating sensitive information from two perspectives according to claim 1 is characterized in that: The positive and negative user sampling method of the triplet (target user, positive user, negative user) in step 4 is expressed as and Among them, t1 and t2 are hyperparameters.

5. The fair recommendation method for dissociating sensitive information from dual perspectives according to claim 1 is characterized in that: In step 4, the BPR Loss of the triplet is calculated based on the user's true sensitive representation and is recorded as the true sensitive representation loss: in, represents the set of positive users sampled from user u, Represents the set of negative users sampled from user u.

6. The fair recommendation method for dissociating sensitive information from two perspectives according to claim 1 is characterized in that: In step 5, the counterfactual sensitive representation vectors of users and items are used to predict counterfactual sensitive attributes and calculate the prediction loss of counterfactual sensitive attributes. The process formula is expressed as: in represents a trained neural network classifier with fixed parameters θ c , 1 C represents a vector of all 1s with dimension C, and Represent the predicted values ​​of counterfactual sensitive attributes of users and items, and Represent the counterfactual sensitive attribute values ​​of users and items respectively.

7. The fair recommendation method for dissociating sensitive information from dual perspectives according to claim 1, characterized in that: In step 5, the BPRLoss of the triple (target user, positive user, negative user) is calculated using the counterfactual sensitive representations of users and items:

8. The fair recommendation method for dissociating sensitive information from dual perspectives according to claim 1, characterized in that: In step 6, the mutual information between the fair representation and the sensitive representation and the counterfactual sensitive representation is calculated by the mutual information estimation upper bound method:

9. The fair recommendation method for dissociating sensitive information from two perspectives according to claim 1, characterized in that: In step 6, the absolute difference of cosine similarity between the fair representation and the sensitive representation and counterfactual sensitive representation is calculated:

10. The fair recommendation method for dissociating sensitive information from two perspectives according to claim 1, characterized in that: In step 6, the mutual information between the fair representation and the performance-oriented representation is calculated by the mutual information estimation lower bound method:

11. The fair recommendation method for dissociating sensitive information from dual perspectives according to claim 1, characterized in that: In step 6, the BPR Loss of all training data triplets (user, interacted item, non-interacted item) is calculated:

12. The fair recommendation method for dissociating sensitive information from dual perspectives according to claim 1, characterized in that: In step 7, the joint loss is calculated:

Images