Fair recommendation method based on double-path feature modulation adversarial learning

Abstract

The invention discloses a fair recommendation method based on double-path feature modulation adversarial learning. The method comprises the following steps: S1, modeling according to interaction data of a user and an article; s2, generating a feature influence factor through a feature modulation module; s3, processing the characteristic influence factors for sensitive information of a user side and an article side through a two-way adversarial learning module; s4, performing weighted fusion on the processed embedding and the original embedding to generate final depolarization embedding representation; s5, model training; and S6, outputting a recommendation result. And the influence of sensitive information on the recommendation result is effectively removed, so that the cross bilateral fairness of the recommendation system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation systems, and in particular to a fair recommendation method based on double-path feature modulation adversarial learning. BACKGROUND

[0002] With the rapid development of information technology, recommendation systems have become an important part of the digital economy era and are widely used in e-commerce, streaming media, social networks and many other fields, providing better personalized experiences for users. However, in the process of pursuing personalized recommendations, the fairness problem of recommendation systems has gradually emerged. Traditional recommendation systems are mainly trained based on historical user data, which has many limitations, such as the unreliability of estimating user preferences based on sparse and noisy historical data, the neglect of online context environmental factors that affect user behavior, and the unreliable assumption that users are aware of their own preferences. These limitations may lead to biased recommendations for different user groups or item groups, thereby causing fairness problems.

[0003] Cross-bilateral fairness, as an important dimension of recommendation system fairness, requires that the recommendation system be fair on both the user side and the item side, i.e., not only to ensure that different user groups have fair recommendation opportunities, but also to ensure that different item groups have fair exposure and recommendations. For example, in the movie recommendation scenario, it is necessary to avoid recommendation bias for users of a particular gender or age, and to prevent certain types of movies or niche movies from suffering from unfair lack of exposure.

[0004] In recent years, the development of digital economy has promoted the widespread application of recommendation systems, but has also placed higher demands on the fairness of recommendation systems. As an important application field of artificial intelligence, the fairness of recommendation systems not only affects user experience, but also relates to the sustainable development of business and social fairness and justice. In order to solve the cross-bilateral unfairness problem in recommendation systems, existing research mainly focuses on the following aspects. One method is to eliminate the influence of sensitive information by preprocessing the recommendation data, such as data cleaning and resampling techniques, aiming to balance the data distribution of different groups; another method is to introduce fairness constraints or regularization terms in the recommendation algorithm to reduce dependence on sensitive attributes; and some research proposes to adjust the results in the post-processing stage of the recommendation results to meet the fairness requirements. However, these methods still have the following shortcomings in practical application:

[0005] 1. Data preprocessing methods may lose a lot of useful information, affecting the accuracy of the recommendation system; 2. Algorithms that introduce fairness constraints may increase model complexity, affecting recommendation efficiency; 3. Post-processing methods may not fundamentally solve the fairness problem in recommendation systems, but only modify the results to a certain extent. SUMMARY

[0006] The present invention aims to solve the technical problems existing in the prior art. In particular, it innovatively proposes a fair recommendation method based on dual-path feature modulation adversarial learning, which can effectively remove the influence of sensitive information on the recommendation results, thereby improving the cross-bilateral fairness of the recommendation system.

[0007] To achieve the above object, the present invention provides a fair recommendation method based on dual-path feature modulation adversarial learning, comprising the following steps:

[0008] S1: Modeling based on the interaction data between users and items;

[0009] S2: Generate feature impact factors through feature modulation module;

[0010] S3: The feature influencing factors are processed separately for sensitive information on the user side and the item side through a two-way adversarial learning module;

[0011] S4: Perform weighted fusion of the processed embedding and the original embedding to generate the final debiased embedding representation;

[0012] S5: Model training;

[0013] S6: Output recommendation results.

[0014] In the above scheme: step S2 includes the following steps:

[0015] S2-1: User sensitive information i Generate feature impact factors;

[0016] S2-2: Sensitive information about items j Generate feature impact factors.

[0017] In the above scheme, step S2-1 includes the following steps:

[0018] S2-1-1: Initialize feature influencing factors through multi-layer perceptron learning;

[0019] According to user sensitive information i Learning initial feature influence factors through MLP and The expression is as follows:

[0020]

[0021] Where, Indicates user sensitive information i The learning function, Indicates the initial feature impact factor of user sensitive information, Indicates that it contains user sensitive information i Embedding

[0022] S2-1-2: adaptively linearly modulate the initial feature influence factor;

[0023] According to the initial feature influence factor The scale parameter a and the shift parameter b are generated by the FiLM network layer, and the expression is as follows:

[0024]

[0025] In the formula, and are the scale and shift parameters of the user sensitive information s i FiLM network layer of the user sensitive information s i

[0026] S2-1-3: modulate the initial feature influence factor to the final feature influence factor The expression is as follows:

[0027]

[0028] In the formula, is the final user sensitive information feature influence factor, and are the scale and shift parameters of the user sensitive information s i

[0029] S2-1-4: standardize the feature influence factor; after the final feature influence factor is normalized,

[0030] connect the residual with the initial feature influence factor

[0031] The formula is as follows:

[0032]

[0033] In the formula, BatchNorm(·) represents the batch normalization operation.

[0034] In the above scheme: step S2-2 includes the following steps:

[0035] S2-2-1: learn the initial feature influence factor by a multi-layer perception;

[0036] According to the item sensitive information t j learn the initial feature influence factor The expression is as follows:

[0037] ​​​​

[0038] wherein, denotes the sensitive information t of the item j , and denotes the initial feature influence factor of the sensitive information of the item j ; denotes the embedding containing the sensitive information t of the item j ;

[0039] S2-2-2: Adaptive linear modulation is performed on the initial feature influence factor

[0040] According to the initial feature influence factor , the scale parameter a and the shift parameter b are generated through the FiLM network layer, and the expression is as follows:

[0041]

[0042] wherein, and are the scale and shift parameters of the sensitive information t of the item j ; denotes the FiLM network layer of the sensitive information t of the item j ;

[0043] S2-2-3: The initial feature influence factor is modulated into the final feature influence factor , and the expression is as follows:

[0044]

[0045] wherein, is the final feature influence factor of the sensitive information of the item, and are the scale and shift parameters of the sensitive information t of the item j ;

[0046] S2-2-4: Standardization processing is performed on the feature influence factor; after the final feature influence factor is normalized, the residual connection is performed with the initial feature influence factor ;

[0047] The formula is as follows:

[0048]

[0049] wherein, BatchNorm(·) denotes the batch normalization operation.

[0050] In the above scheme: step S3 comprises the following steps:

[0051] S3-1: Filter the bias of the embedding by the generator of the two-path adversarial learning module; the expression is as follows:

[0052]

[0053] where e u and e i are the original input embeddings of users and items respectively, and are different sensitive information bias filters for user embeddings and item embeddings respectively; e′ u and e′ i represent the debiased user embeddings and debiased item embeddings after single-path dynamic modulation filtering; |S| and |T| represent the number of categories of user and item sensitive information respectively;

[0054] S3-2: Filter the bias of the embedding again, and the formula is as follows:

[0055]

[0056] where e” u and e” i represent the unbiased user embeddings and unbiased item embeddings obtained after two-path filtering respectively;

[0057] S3-3: Calculate the predicted attribute;

[0058] The calculation is performed by the following formula:

[0059]

[0060] where, and represent the predicted probability set of the attribute, and represent the discriminators in the model, (·) T represents the transpose function; e” u and e” i are the unbiased user embeddings and unbiased item embeddings obtained after two-path filtering, e u and e i are the original input embeddings of users and items respectively; represents a sparse matrix, and m and n are the number of users and the number of items respectively; represents the transpose of the sparse matrix;

[0061] S3-4: Calculate the loss with respect to the user sensitive information s i and the item sensitive information t j respectively; if the user sensitive information s iIf there are only two categories, then S3-5 is performed, if the user sensitive information s i If there are at least three categories, then S3-6 is performed; if the item sensitive information t j If there are only two categories, then S3-5 is performed, if the item sensitive information t j If there are at least three categories, then S3-6 is performed;

[0062] S3-5: For binary user sensitive information s i and item sensitive information t j , the following formula is used:

[0063]

[0064] wherein, and respectively represent the user sensitive information discriminator and the item sensitive information discriminator, m and n are respectively the number of users and the number of items, and are respectively the real sample corresponding to the user and the real sample corresponding to the item; denotes the set of prediction probabilities of the sensitive information t j of the item k; denotes the set of prediction probabilities of the sensitive information s i of the user k;

[0065] S3-7 is performed;

[0066] S3-6: For multi-class user sensitive information s i and item sensitive information t j , the following formula is used:

[0067]

[0068] wherein, and respectively represent the user sensitive information discriminator and the item sensitive information discriminator, denotes the real sample of the c-th category of the sensitive information s i of the user k; denotes the real sample of the c-th category of the sensitive information t j of the item k. denotes the prediction probability of the c-th category of the sensitive information s i of the user k; denotes the prediction probability of the c-th category of the sensitive information t j of the item k; |s i | and |t j | respectively represent the number of categories of the user sensitive information and the item sensitive information;

[0069] S3-7 is executed;

[0070] S3-7: Calculate the comprehensive loss function; the loss of each discriminator is integrated into a comprehensive loss function through a weighted feature fusion method, and the expression is as follows:

[0071]

[0072] wherein, is the comprehensive loss, and respectively represent the user sensitive information discriminator and the item sensitive information discriminator.

[0073] In the above scheme, step S4 includes the following steps:

[0074] S4-1: Through the weighted feature fusion module, the single-path debiased original embedding e' u and e' i and the double-path debiased embedding e" u and e" i are fused; the formula is as follows:

[0075]

[0076] wherein, and are the fused user and item embeddings, |I u | is the set of historical interaction items corresponding to the user, is a weighting coefficient, used to balance the influence of the single-path debiased embedding and the double-path debiased embedding;

[0077] S4-2: Predict the rating of the user to the item; the inner product of the user embedding and the item embedding is calculated through the recommendation module to predict the rating of the user to the item; the formula is as follows:

[0078]

[0079] In the formula, represents the recommendation prediction score, and the top-k items in the prediction score can be recommended to the corresponding user; is the fused item embedding, is the transposed fused user embedding;

[0080] S4-3: The double-path filter loss function is constructed as follows:

[0081]

[0082] wherein, represents the mean square error loss function, r ui represents the actual rating, λ uand lambda i respectively represent regularization coefficients of user and item embedding; represents recommendation loss, represents recommendation prediction score; and represent L2 regularization on users and items.

[0083] In the above scheme, step S5 comprises the following steps:

[0084] The parameters of the two-way filter and the discriminator are iteratively optimized and updated in an adversarial training manner; the loss function expression of the model training target is as follows:

[0085]

[0086] In the formula, theta Rec and are respectively the learnable parameters of the two-way filter, i.e., the generator and the discriminator, and delta is a weight coefficient for controlling the recommendation performance and the de-biasing intensity; represents recommendation loss, represents minimizing the recommendation loss and maximizing the discriminator loss.

[0087] To sum up, the beneficial effects of the present application are: the influence of sensitive information of users and items on embedding representation can be effectively quantified and captured, the feature influence factor reflecting the influence of sensitive information is generated through the combination of the multi-layer perception and the FiLM network layer, and key data support is provided for subsequent adversarial learning; through adversarial training, the filtering intensity is dynamically adjusted, and the influence of sensitive information on the recommendation result is effectively removed, so as to improve the cross-bilateral fairness of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 is a framework structure diagram of the present application.

[0089] Figure 2 a is a comparison diagram of the U_ITGF index experimental results of the present application, NIFTY, FairAttack and ITFR based on the LastFM-2B data set, b is a comparison diagram of the I_ITGF index experimental results of the present application, NIFTY, FairAttack and ITFR based on the LastFM-2B data set, and c is a comparison diagram of the ITGF index experimental results of the present application, NIFTY, FairAttack and ITFR based on the LastFM-2B data set.

[0090] Figure 3a is a comparison chart of the experimental results of the U_ITGF index of the application and NIFTY, FairAttack and ITFR based on the Movielens-1M dataset, b is a comparison chart of the experimental results of the I_ITGF index of the application and NIFTY, FairAttack and ITFR based on the Movielens-1M dataset, and c is a comparison chart of the experimental results of the ITGF index of the application and NIFTY, FairAttack and ITFR based on the Movielens-1M dataset. DETAILED DESCRIPTION

[0091] The application will be further described below by way of examples and with reference to the accompanying drawings:

[0092] As shown in the figure, a fair recommendation method based on two-way feature modulation adversarial learning includes the following steps: Figure 1

[0093] S1: modeling according to user and item interaction data;

[0094] Let the user set be U = {u1, u2,..., u m}, the item set be I = {i1, i2,..., i n}, and the user's feedback to the item be formalized by the user-item interaction matrix , where the element r ui represents the interaction value between user u and item i, and 0 if there is no interaction; the sensitive information of the user and the item is represented by the set S = {s1, s2,..., s p} and T = {t1, t2,..., t q}, respectively, where p and q represent the number of categories of user and item sensitive information; each category of sensitive information includes various subcategories, such as male and female in the gender category; let the cross-group set be the user and item sensitive information subcategory and pairing set.

[0095] S2: generating feature influence factors through a feature modulation module;

[0096] S2-1: generating feature influence factors for user sensitive information s i ;

[0097] S2-1-1: learning initial feature influence factors through a multi-layer perception;

[0098] According to the user sensitive information s i (such as gender, age, etc.), the initial feature influence factor is learned through MLP, and the expression is as follows:

[0099] ​

[0100] Where, Indicates user sensitive information i The learning function, Indicates the initial feature impact factor of user sensitive information, Indicates that it contains user sensitive information i embedded; and and It can be obtained through matrix factorization, graph neural networks and self-supervised learning models.

[0101] S2-1-2: Adaptive linear modulation of the initial characteristic influencing factors;

[0102] According to the initial characteristic impact factor The scale parameter α and shift parameter β are generated by the FiLM network layer, and the expressions are as follows:

[0103]

[0104] Where, and Is user sensitive information i The scale and shift parameters, Indicates user sensitive information i FiLM network layer;

[0105] S2-1-3: Initial feature impact factor Modulation into the final characteristic influencing factor Make it more

[0106] To more accurately reflect the actual impact of sensitive information in the embedded representation, the expression is as follows:

[0107]

[0108] Where, It is the final user sensitive information feature influencing factor, and Is user sensitive information i The scale and shift parameters of ;

[0109] S2-1-4: Standardize the characteristic influencing factors; the final characteristic influencing factors Normalize

[0110] After processing, the initial characteristic influencing factor Perform residual connection;

[0111] To accelerate the training process and alleviate the problem of gradient vanishing or exploding, we introduce batch normalization and residual connection in the feature modulation module. Among them, batch normalization can standardize the feature influence factor, and residual connection can combine the initial feature influence factor with the modulated feature influence factor.

[0112] The formula is as follows:

[0113]

[0114] In the formula, BatchNorm(·) represents the batch normalization operation; the residual connection helps the model better learn the increment of feature modulation by adding the initial feature influence factor and to the modulated feature influence factor.

[0115] Through the above steps, the feature modulation module can generate feature influence factors reflecting the influence of sensitive information and These factors will be integrated into the subsequent dual-path adversarial learning module as the basis for generator and discriminator learning, helping the model effectively remove the influence of sensitive information and achieve fairness in recommendation results.

[0116] S2-2: Generate feature influence factors based on item sensitive information t j

[0117] S2-2-1: Learn the initial feature influence factor through a multi-layer perceptron;

[0118] According to the item sensitive information t j (e.g. theme, popularity, etc.), learn the initial feature influence factor through MLP. The expression is as follows:

[0119]

[0120] In the formula, represents the learning function of item sensitive information t j , represents the initial feature influence factor of item sensitive information, represents the embedding containing item sensitive information t j ; and and can be obtained through matrix decomposition, graph neural network and self-supervised learning model.

[0121] S2-2-2: Perform adaptive linear modulation on the initial feature influence factor;

[0122] According to the initial feature influence factor ​The scale parameter a and the shift parameter b are generated by the FiLM network layer, and the expression is as follows:

[0123]

[0124] In the formula, and are the scale and shift parameters of the item sensitive information t j . FiLM network layer of the item sensitive information t j .

[0125] S2-2-3: Modulate the initial feature influence factor to the final feature influence factor so that it can more accurately reflect the actual influence of sensitive information in the embedding representation, and the expression is as follows:

[0126]

[0127] In the formula, is the final item sensitive information feature influence factor, and are the scale and shift parameters of the item sensitive information t j .

[0128] S2-2-4: Standardize the feature influence factor; after normalizing the final feature influence factor , the residual connection is performed with the initial feature influence factor .

[0129] In order to accelerate the training process and alleviate the problem of gradient vanishing or explosion, batch normalization and residual connection are introduced in the feature modulation module. Among them, batch normalization can standardize the feature influence factor, and residual connection can connect the initial feature influence factor with the modulated feature influence factor.

[0130] The formula is as follows:

[0131]

[0132] In the formula, BatchNorm(·) represents the batch normalization operation; the residual connection helps the model better learn the increment of feature modulation by adding the initial feature influence factor and to the modulated feature influence factor.

[0133] Through the above steps, the feature modulation module can generate the feature influence factor reflecting the influence of sensitive information These factors will be integrated into the subsequent two-way adversarial learning module as the basis for generator and discriminator learning, helping the model effectively remove the influence of sensitive information and achieve fairness in recommendation results.

[0134] S3: The feature influence factor is processed by the two-way adversarial learning module for sensitive information on the user side and the item side respectively;

[0135] S3-1: The generator of the two-way adversarial learning module filters the embedded bias; the expression is as follows:

[0136]

[0137] In the formula, e u and e i are the original input embeddings of users and items respectively, and are different sensitive information bias filters for user embeddings and item embeddings respectively; e′ u and e′ i represent the debiased user embeddings and debiased item embeddings after single-path dynamic modulation filtering; |S| and |T| represent the number of categories of user and item sensitive information respectively;

[0138] S3-2: The embedded bias is filtered again, and the formula is as follows:

[0139]

[0140] In the formula, e" u and e" i represent the user unbiased embedding and item unbiased embedding obtained after two-way filtering respectively;

[0141] S3-3: Calculate the predicted attribute;

[0142] In order to optimize the minimax interaction with the two-way filter, the discriminator needs to accurately detect the complex distribution of sensitive information in the generated embedding. Therefore, we tailor the discriminator for each sensitive information category of users and items. It should be noted that the loss function of the discriminator is based on the difference between the predicted attribute and the real attribute. However, when the user (item) class sensitive information discriminator tries to identify the item (user) sensitive information, it will encounter the problem of mismatch between the length of the predicted attribute and the real attribute.

[0143] To solve this problem, we use a sparse matrix to convert the embedding before starting the dual target discrimination process.

[0144] The calculation is as follows:

[0145]

[0146] wherein, and respectively represent a set of predicted probabilities of attributes, and denote a discriminator in the model, (·) T denotes a transpose function; e" u and e" i respectively are the user and item unbiased embeddings after two-path filtering, e u and e i are the original input embeddings of users and items, respectively; denotes a sparse matrix, and m and n are the number of users and items, respectively; denotes a transpose of a sparse matrix;

[0147] S3-4: Calculate the loss with respect to the user sensitive information s i and the item sensitive information t j ; if the user sensitive information s i has only two categories, perform S3-5, if the user sensitive information s i has at least three categories, perform S3-6; if the item sensitive information t j has only two categories, perform S3-5, if the item sensitive information t j has at least three categories, perform S3-6;

[0148] S3-5: Calculate the binary user sensitive information s i and the item sensitive information t j by the following formula:

[0149]

[0150] wherein, and respectively represent the user sensitive information discriminator and the item sensitive information discriminator, m and n are the number of users and items, respectively, and are the real samples corresponding to the user and the real samples corresponding to the item, respectively; denotes a set of predicted probabilities of the sensitive information t j of the item k; denotes a set of predicted probabilities of the sensitive information s i of the user k;

[0151] perform S3-7;

[0152] S3-6: Calculate the multi-class user sensitive information s i and the item sensitive information t jThe calculation is as follows:

[0153]

[0154] In the formula, and respectively represent a user sensitive information discriminator and an item sensitive information discriminator, representing the true sample of the c-th category of the sensitive information s i of the user k; representing the true sample of the c-th category of the sensitive information t j of the item k. representing the prediction probability of the c-th category of the sensitive information s i of the user k; representing the prediction probability of the c-th category of the sensitive information t j of the item k. |s i |and|t j |respectively represent the number of categories of the user sensitive information and the item sensitive information;

[0155] S3-7 is executed;

[0156] S3-7: Calculate the comprehensive loss function; the loss of each discriminator is integrated into a comprehensive loss function through a weighted feature fusion method, and the expression is as follows:

[0157]

[0158] wherein, is the comprehensive loss, and respectively represent a user sensitive information discriminator and an item sensitive information discriminator; we integrate the feature influence factor into the discriminator loss to enhance the response ability of the discriminator to the filter.

[0159] S4: The processed embedding is weighted and fused with the original embedding to generate the final debiased embedding representation;

[0160] S4-1: Through a weighted feature fusion module, the single-path debiased original embedding e′ u and e′ i and the double-path debiased embedding e” u and e” i are fused; the formula is as follows:

[0161]

[0162] wherein, and are the fused user and item embeddings, respectively, |I u |is the set of historical interaction items corresponding to the user, is a weighting coefficient used to balance the influence of the single-path debiased embedding and the double-path debiased embedding. The coefficient can be tuned by a validation set to ensure that the fused embedding performs best in the recommendation task.

[0163] S4-2: Predict the rating of the user to the item; calculate the inner product of the user embedding and the item embedding through the recommendation module to predict the rating of the user to the item; the formula is as follows:

[0164]

[0165] In the formula: represents the recommendation prediction score, and the top-k items in the prediction score can be recommended to the corresponding user; is the fused item embedding, is the transposition of the fused user embedding;

[0166] S4-3: Construct the double-path filter loss function as follows:

[0167]

[0168] wherein, represents the mean square error loss function, r ui represents the actual rating, λ u and λ i respectively represent the regularization coefficients of the user and item embeddings; represents the recommendation loss, represents the recommendation prediction score; and represent L2 regularization for users and items;

[0169] S5: Model training;

[0170] The parameters of the double-path filter and the discriminator are iteratively optimized and updated in an adversarial training manner. In the training process of DFLAR, we first introduce a feature linear modulation bias capture mechanism to quantify the feature influence factor, and then apply it to the double-path adversarial learning mechanism to dynamically regulate the debiasing process, as well as the iterative optimization and update of the parameters. After reaching the preset training period, the training process ends.

[0171] The loss function expression of the model training target is as follows:

[0172]

[0173] In the formula, θ Rec and are the learnable parameters of the double-path filter, i.e., the generator and the discriminator, and δ is the weight coefficient controlling the recommendation performance and debiasing intensity; represents the recommendation loss, minimizing the recommendation loss and maximizing the discriminator loss, i.e., maximizing the min-max adversarial training between the recommendation and the discriminator.

[0174] S6: outputting the recommendation result.

[0175] To better illustrate the technical solutions, the following are verified through recommendation and fairness experiments and cross-bilateral fairness experiments:

[0176] Recommendation and fairness experiment

[0177] The DFLAR of the technical solutions is experimentally tested together with NIFTY, FairAttack, and ITFR on two commonly used data sets of the recommendation system, i.e., Movielens-1M and LastFM-2B, to debias different user and item sensitive information and their combinations contained in the data, so as to prove the applicability of the method DFLAR in different situations. RMSE represents the recommendation performance index, and the smaller the better; AUC represents the fairness index of the binary sensitive information, and F1 represents the fairness index of the multi-class sensitive information, both of which are the smaller the better. The experimental comprehensive results are shown in Table 1.

[0178] Table 1: Comprehensive results of recommendation and fairness experiment

[0179]

[0180] The input data of DFLAR contains pre-trained user and item embeddings. Analysis shows that: LR-GCCF has good recommendation performance (smaller RMSE) but more sensitive information exposure (larger AUC and F1); HIE performs well on Movielens-1M because it integrates social heterogeneous information and does not directly process sensitive information, but performs generally on LastFM-2B; ICML, FairGo, and FairAttack, etc., which are based on adversarial learning to eliminate sensitive information, have poor recommendation performance; ITFR guarantees group fairness through loss cooperation, but has medium recommendation performance; NIFTY has medium recommendation performance and does not completely overcome the inherent bias of graph data; DFLAR achieves bilateral fairness while guaranteeing group fairness and group recommendation performance, has the best ability to eliminate sensitive information, and has good recommendation performance. Compared with the baseline method, the debiasing ability of DFLAR on the Movielens-1M and LastFM-2B data sets is improved by 1.91% and 1.42% respectively, compared with the fairness-aware model, the recommendation performance is improved by 1.69% and 1.94% respectively while maintaining the debiasing effect, compared with ITFR, the debiasing ability is improved by 10.5% and 1.8% respectively, and the recommendation performance is improved by 4.9% and 8.2% respectively.

[0181] Cross-bilateral fairness experiment

[0182] Experiments are conducted on two commonly used datasets, Movielens-1M and LastFM-2B. The experiment is to verify whether the method DFLAR proposed by us has an advantage in cross-bilateral fairness. We conduct experiments on different cross groups (such as (gender, theme)). We combine the mean square error and the coefficient of variation to make it an index for verifying the fairness between groups, where U_ITGF and I_ITGF are the indexes of user and item cross one-way fairness respectively, and ITGF is the index of cross-bilateral fairness, all of which are better the smaller.

[0183] To further verify the effectiveness of our method in solving the problem of cross-bilateral unfairness, we compare the fairness performance of DFLAR with three baseline models, NIFTY, FairAttack and ITFR, on two datasets, as shown in Figure 2 and Figure 3 .

[0184] The main findings are as follows: the fairness of FairAttack is better than that of NIFTY as a whole, but there are exceptions, because it does not explicitly solve the fairness consistency between groups, and the uneven distribution of debiasing ability may lead to group-level unfairness; the fairness of ITFR is better than that of FairAttack as a whole, but there are exceptions, although it is designed for group fairness consistency, it lacks debiasing ability, and the influence of sensitive information on cross-bilateral fairness sometimes makes it worse than FairAttack; compared with NIFTY, although the overall fairness of ITFR is better, the fairness of cross groups under different sensitive information categories will affect each other, and in order to maintain the overall fairness, the fairness of some cross groups may be sacrificed, leading to occasional decline in fairness; compared with the perspective model, DFLAR performs well in the cross groups defined by different sensitive attribute categories on two datasets, but it occasionally underperforms compared with ITFR, because the aggregation embedding itself increases the unfairness, and the influence of sensitive information in different cross groups is different, and the overall consideration of the model may make the fairness of some cross groups lower than the average level. In summary, DFLAR exhibits excellent fairness consistency between groups within cross groups with the same user or item attributes, and is a better solution to the challenge of cross-bilateral unfairness.

[0185] 1. Balance recommendation performance and fairness, and reduce the risk of algorithm bias

[0186] DFLAR significantly reduces sensitive information leakage (low AUC / F1) while maintaining low recommendation error (low RMSE) through a bilateral fairness mechanism, helping enterprises avoid legal disputes or reputation loss caused by algorithm bias. Users will get more neutral recommendation results. For example, a music platform will not excessively recommend a specific genre to users of a certain gender, and a movie platform can reduce the neglect of niche themes, promoting content diversity.

[0187] 2. Solve the cross-bilateral fairness problem and adapt to complex real-world scenarios

[0188] User and item attributes in reality are often multi-dimensional and cross-cutting (e.g., "female users" and "action movies," "young users" and "independent music"). DFLAR verifies fairness in cross-cutting groups (e.g., gender + genre) and demonstrates its ability to handle complex scenarios with multi-attribute overlap, while traditional models (e.g., FairAttack, ITFR) may compromise on single-dimension fairness. For example, in job recommendation, the system needs to avoid cross-discrimination against "female + technical positions" or "minority + executive positions." DFLAR can provide technical support for such scenarios and promote social fairness.

[0189] 3. Improve user trust and experience, and break the information cocoon

[0190] DFLAR reduces the impact of sensitive information on recommendations and reduces the "information cocoon" effect. Users are no longer confined to the preferences set by algorithms (e.g., "male users only watch science fiction movies") and are exposed to a wider range of content. Users discover more content that meets their interests but is not actively searched (e.g., niche independent films, non-mainstream music), increasing platform stickiness. For example, LastFM users discover more cross-genre music combinations through DFLAR.

[0191] 4. Promote industry technology upgrade and meet regulatory requirements

[0192] With the global emphasis on AI ethics, businesses need to adopt more fair recommendation technologies. DFLAR's experimental data prove that it can meet such compliance requirements and become the preferred solution for technology upgrades. By deploying DFLAR, businesses can demonstrate algorithm fairness evidence in annual audits or user complaints, improve ESG (environment, society, governance) ratings, and attract socially responsible investments.

[0193] 5. Promote balanced resource allocation and narrow the digital divide

[0194] In e-commerce, education, and other scenarios, DFLAR can avoid recommendation suppression for disadvantaged groups (e.g., low-income users, niche creators). For example, e-commerce platforms will not reduce high-quality product recommendations for users with low historical consumption, helping niche brands gain fair exposure. Creators (e.g., independent filmmakers, niche musicians) gain audience through more fair recommendations, promoting cultural diversity; consumers can access more diverse products, reducing market monopolization caused by "head effect."

[0195] The research results of DFLAR not only improve the technical boundaries of recommendation systems, but also provide practical tools for solving algorithm discrimination and promoting social inclusion by balancing efficiency and fairness. The practical significance lies in making technology more "humanized" - meeting users' demand for accurate recommendations while respecting individual differences and group diversity, ultimately building a more fair and trustworthy digital ecosystem.

Claims

1. A fair recommendation method based on dual-path feature modulation adversarial learning, characterized by: The following steps are involved: S1: Modeling based on the interaction data between users and items; S2: Generate feature impact factors through feature modulation module; S3: The feature influencing factors are processed separately for sensitive information on the user side and the item side through a two-way adversarial learning module; S4: Perform weighted fusion of the processed embedding and the original embedding to generate the final debiased embedding representation; S5: Model training; S6: Output recommendation results.

2. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 1, characterized in that: Step S2 includes the following steps: S2-1: User sensitive information i Generate feature impact factors; S2-2: Sensitive information about items j Generate feature impact factors.

3. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 2, characterized in that: Step S2-1 includes the following steps: S2-1-1: Initialize feature influencing factors through multi-layer perceptron learning; According to user sensitive information i Learning initial feature influence factors through MLP and The expression is as follows: Where, Indicates user sensitive information i The learning function, Indicates the initial feature impact factor of user sensitive information, Indicates that it contains user sensitive information i Embedding; S2-1-2: Adaptive linear modulation of the initial characteristic influencing factors; According to the initial characteristic impact factor The scale parameter α and shift parameter β are generated by the FiLM network layer, and the expressions are as follows: Where, and Is user sensitive information i The scale and shift parameters, Indicates user sensitive information i FiLM network layer; S2-1-3: Initial feature impact factor Modulation into the final characteristic influencing factor The expression is as follows: Where, It is the final user sensitive information feature influencing factor, and Is user sensitive information i scale and shift parameters of ; S2-1-4: Standardize the characteristic influencing factors; the final characteristic influencing factors After normalization, the initial feature impact factor Perform residual connection; The formula is as follows: Where BatchNorm(·) represents the batch normalization operation.

4. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 2, characterized in that: Step S2-2 includes the following steps: S2-2-1: Initialize feature influencing factors through multi-layer perceptron learning; According to the sensitive information of the item j Learning initial feature influence factors through MLP The expression is as follows: Where, Indicates sensitive information of items j The learning function, Indicates the initial characteristic impact factor of the sensitive information of the item, Indicates that the item contains sensitive information j Embedding S2-2-2: Adaptive linear modulation of the initial characteristic influencing factors; According to the initial characteristic impact factor The scale parameter α and shift parameter β are generated by the FiLM network layer, and the expressions are as follows: Where, and Is the item sensitive information j scale and shift parameters of ; Indicates sensitive information of items j FiLM network layer; S2-2-3: Initial feature impact factor Modulation into the final characteristic influencing factor The expression is as follows: Where, is the final factor affecting the sensitive information characteristics of items. and Is the item sensitive information j The scale and shift parameters of ; S2-2-4: Standardize the characteristic impact factors; the final characteristic impact factors After normalization, the initial feature impact factor Perform residual connection; The formula is as follows: Where BatchNorm(·) represents the batch normalization operation.

5. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 1, characterized in that: Step S3 includes the following steps: S3-1: Filter the embedded bias through the generator of the two-way adversarial learning module; the expression is as follows: Where, e u and e i are the raw input embeddings for users and items respectively, and Different sensitive information bias filters for user embedding and item embedding respectively; e' u and e' i denote the debiased user embedding and debiased item embedding after single-pass dynamic modulation filtering; |S| and |T| denote the number of categories representing sensitive information of users and items, respectively; S3-2: The embedding is biased again, and the formula is as follows: Among them, e” u and e” i They represent the unbiased embedding of users and items after two-way filtering; and Different sensitive information bias filters for user embedding and item embedding respectively; S3-3: Calculate predicted attributes; Calculated using the following formula: in, and Represent the predicted probability set of attributes, and represents the discriminator in the model, (·) T represents the transpose function; e u ” and e i ” are respectively the unbiased embedding of users and items after two-way filtering, e u and e i The raw input embeddings for users and items respectively; represents a sparse matrix, and m and n are the number of users and items respectively; represents the transpose of the sparse matrix; |S| and |T| represent the number of categories representing sensitive information of users and items, respectively; S3-4: User sensitive information s i and sensitive information of items j Calculate the loss; if the user's sensitive information s i There are only two categories, then execute S3-5. If the user's sensitive information s i If there are at least three categories, execute S3-6; if the sensitive information of the item is t j If there are only two categories, execute S3-5. If the sensitive information of the item is t j If there are at least three categories, execute S3-6; S3-5: Use the following formula to classify the user sensitive information s into two categories i and sensitive information of items j Perform the calculation: Where, and They represent user sensitive information discriminators and item sensitive information discriminators, m and n are the number of users and the number of items respectively. and They are the real samples corresponding to users and the real samples corresponding to items; Represents the sensitive information t of item k j The predicted probability set of ; Represents the sensitive information s of user k i The predicted probability set of ; Execute S3-7; S3-6: Use the following formula to classify multi-class user sensitive information s i and sensitive information of items j Perform the calculation: Where, and Represent user sensitive information discriminator and item sensitive information discriminator respectively, Represents the sensitive information s of user k i The real sample of category c; Represents the sensitive information t of item k j The real sample of the c-th category. Represents the sensitive information s of user k i The predicted probability of the cth category; Represents the sensitive information t of project k j The predicted probability of the cth category; |s i | and |t j | respectively represent the number of types of user sensitive information and item sensitive information; Execute S3-7; S3-7: Calculate the comprehensive loss function; integrate the losses of each discriminator into a comprehensive loss function through the weighted feature fusion method, the expression is as follows: in, is the comprehensive loss, and They represent user sensitive information discriminator and item sensitive information discriminator respectively.

6. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 1, characterized in that: Step S4 includes the following steps: S4-1: Through the weighted feature fusion module, the original single-path debiased embedding e' u and e' i and the embedding e" after dual-path debiasing u and e” i Fusion is performed; the formula is as follows: in, and are the fused user and item embeddings, respectively. u | is the historical interaction item set of the corresponding user, is a weighting coefficient used to balance the effects of the embedding after single-path debiasing and the embedding after dual-path debiasing; S4-2: Predict the user's rating of the item; the recommendation module calculates the inner product of the user embedding and the item embedding to predict the user's rating of the item; the formula is as follows: Where: Represents the recommendation prediction score, and the top k items in the prediction score can be recommended to the corresponding user; To embed the fused items, It means transposing the fused user embedding; S4-3: Construct the two-way filter loss function as follows: in, represents the mean square error loss function, r ui represents the actual score, λ u and λ i denote the regularization coefficients of user and item embeddings respectively; represents the recommendation loss, represents the recommendation prediction score; and represents L2 regularization for users and items.

7. The fair recommendation method based on dual-path feature modulation adversarial learning according to claim 1, characterized in that: Step S5 includes the following steps: The parameters of the two-way filter and the discriminator are iteratively optimized and updated using adversarial training. The loss function of the model training objective is expressed as follows: Where θ Rec and are the learnable parameters of the two-way filter, namely the generator and the discriminator, and δ is the weight coefficient that controls the recommendation performance and debiasing strength; represents the recommendation loss, denotes minimizing the recommendation loss and maximizing the discriminator loss.

Images