A localized differential privacy recommendation method and system that balances algorithmic fairness
By dividing users into groups based on their activity levels in the recommendation system and performing local pre-training and dynamic gradient pruning, combined with adaptive noise adjustment, the problem of recommendation performance differences caused by privacy protection technologies is solved, and the algorithm's fairness and recommendation accuracy are synergistically optimized while ensuring privacy and security.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-02
AI Technical Summary
Existing recommendation systems struggle to simultaneously optimize algorithmic fairness and recommendation accuracy while ensuring privacy and security. This is especially evident in low-activity groups, where significant performance differences and model utility losses due to privacy protection are observed.
By grouping users according to their activity levels, performing local pre-training and matrix factorization, and introducing dynamic gradient pruning and adaptive noise adjustment during training, the performance differences in recommendations between different activity groups are mitigated. Combined with differential privacy protection technology, privacy leaks caused by untrusted servers are prevented.
While effectively preventing privacy leaks, it alleviates the imbalance in recommendation performance among groups, achieving a balance between privacy protection, algorithmic fairness, and recommendation performance, and enhancing the credibility and inclusiveness of the system.
Smart Images
Figure CN122132630A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of computer recommendation systems and information security, and relates to a localized differential privacy recommendation method and system that takes into account both algorithmic fairness. Background Technology
[0002] With the rapid development of the internet, big data, and artificial intelligence technologies, information overload has become an increasingly prominent problem, making personalized recommendation systems an indispensable key technology for various online platforms. Recommendation systems, through in-depth mining and modeling of user historical behavior, interests, interaction characteristics, and contextual information, can automatically filter and push content, goods, or services that meet user needs from massive amounts of data, effectively improving information acquisition efficiency, enhancing user experience, and simultaneously increasing user stickiness, activity, and commercial benefits for the platform. Among numerous recommendation methods, matrix factorization-based collaborative filtering technology is widely used in explicit rating data-driven recommendation scenarios due to its strong feature representation ability and generalization performance. This method decomposes a high-dimensional sparse user-item rating matrix into low-dimensional user latent factor vectors and item latent factor vectors, and uses the vector inner product to approximately reconstruct user ratings. This effectively captures the potential associations between users and items and minimizes prediction errors through optimization methods such as stochastic gradient descent, thereby achieving accurate and stable personalized recommendations.
[0003] In recent years, algorithmic fairness in recommender systems has gradually become a focus of research and application. Group fairness focuses on the unbiasedness and balance of the recommender model across different user groups, emphasizing that user groups with different behavioral characteristics, activity levels, and data sparsity should all receive recommender services of similar quality. Traditional recommender algorithms primarily optimize global recommendation accuracy, which can easily lead to bias during training towards highly active groups with abundant data and dense interactions. This results in a significant decrease in recommendation performance for less active groups with sparse data and fewer interactions, thus creating recommendation discrepancies between groups. This unfairness not only affects user experience but also reduces the system's credibility and inclusiveness. Therefore, achieving a fair, reliable, and non-discriminatory recommender mechanism has significant theoretical and practical implications.
[0004] Meanwhile, recommender systems face increasingly severe privacy and security challenges in collecting and using user data. Differential privacy, as a privacy protection technique with rigorous security proofs, effectively prevents privacy leaks and resists inference attacks by adding controllable noise during model training, and has become a mainstream solution for privacy protection in recommender systems. However, while achieving security protection, differential privacy inevitably leads to a loss of model utility, resulting in a decrease in recommendation accuracy. More significantly, the performance degradation caused by differential privacy varies significantly across different user groups, with a more severe impact on low-activity groups with sparse data and limited samples, further exacerbating the recommendation gap between groups and worsening the existing fairness problem. This contradiction is particularly pronounced in practical application scenarios such as untrusted servers. How to mitigate the group performance differences caused by differential privacy while ensuring privacy and security, and achieve synergistic optimization of privacy protection, recommendation accuracy, and algorithmic fairness, has become a key problem that current technologies struggle to solve. Summary of the Invention
[0005] In view of this, the purpose of this invention is to design a localized differential privacy recommendation method and system that takes into account the fairness of the algorithm, so as to solve the problem of privacy leakage and alleviate the problem of increased recommendation performance differences between groups with different activity levels caused by privacy protection technologies.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A localized differential privacy recommendation method and system that balances algorithmic fairness includes the following steps:
[0008] The scenarios considered in this invention include Each user group and an untrusted recommendation server. Holding your own interaction data and user latent factor matrix Project latent factor matrix It is stored by an untrusted server. We consider a semi-honest scenario, where all user groups and the server correctly execute the protocol and only observe the information they receive.
[0009] S1: Before the pre-training iteration begins, the dataset is divided into three groups: active, moderate, and inactive, based on user activity.
[0010] S2: Each user group Holding your own interaction data and user latent factor matrix Initialize the user latent factor matrix and project latent factor matrix To mitigate the performance degradation caused by differential privacy, the user group In training Previously Perform local pre-training;
[0011] S3: At the beginning of each iteration, the recommendation server will update the item latent factor matrix. Distribute to various groups;
[0012] S4: Each group stores and trains its user latent factor matrix locally. Calculate the gradient To prevent pruning bias in different activity groups from causing differences in recommendation performance, [the following measures are taken]. Dynamic gradient clipping is performed based on different activity groups; after aggregating the clipping gradients for each group, adaptive noise addition is achieved by combining a dynamic clipping threshold. To achieve privacy protection and alleviate the problem of exacerbated recommendation performance differences between groups with different activity levels caused by privacy protection technologies, the threshold for gradient pruning in the next round is recalculated. and noise scale ;
[0013] S5: Different User Groups First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix ;
[0014] S6: After the iteration is complete, the final project latent factor matrix will be generated. Delivered to each group, and to the users in each group. Calculate the predicted value on its device , Unobserved interactions are sorted in descending order of predicted values, and the top-K items are then recommended to the user. .
[0015] Furthermore, in step S2, to mitigate the performance degradation caused by differential privacy, the user latent factor matrix is first initialized by the local user group. and project latent factor matrix Then perform matrix factorization training on the whole to obtain and Finally save This will be used for subsequent training. The algorithm will use the user group's rating matrix for the project. User-project pairs dataset The learning rate of the pre-trained user latent factor matrix Number of iterations The user latent factor matrix is output as input after training. Specifically, it includes the following steps:
[0016] S21: First, randomize the user latent factor matrix. Project latent factor matrix ;
[0017] S22: During each training session, for each user in the user group... Every interactive project Used to minimize the loss function:
[0018]
[0019] in, Represented as user For the project Real ratings and Representing users respectively and projects latent factor vectors, and They are collections of users and projects, respectively. For user-project interaction with datasets, and The regularization coefficient is used to prevent overfitting. It is updated iteratively through stochastic gradient descent to obtain a more accurate representation of user preferences and item features.
[0020] S23: Calculate the user latent factor vectors sequentially. Project latent factor vector gradient:
[0021]
[0022]
[0023] S24: For each user's latent factor vector Project latent factor vector To perform the update, use stochastic gradient descent and iteratively update according to the following rules:
[0024]
[0025]
[0026] in, It is the learning rate;
[0027] Repeat steps S22-S24. After each iteration, the trained user latent factor matrix is output. .
[0028] Furthermore, in step S4, each group locally stores and trains the user latent factor matrix. Calculate the gradient To prevent pruning bias in different activity groups from causing differences in recommendation performance, [the following measures are taken]. Dynamic gradient clipping is performed based on different activity groups; after aggregating the clipping gradients for each group, adaptive noise addition is achieved by combining a dynamic clipping threshold. To achieve privacy protection and alleviate the problem of exacerbated recommendation performance differences between groups with different activity levels caused by privacy protection technologies, the threshold for gradient pruning in the next round is recalculated. and noise scale Specifically, it includes the following steps:
[0029] S41: Train for each local group, minimizing the information about... Loss function:
[0030]
[0031] S42: Calculate the item latent factor vector for all user-item pairs in this group. gradient:
[0032]
[0033] S43: Count the number of gradients exceeding the clipping threshold in this group:
[0034]
[0035] in, It is the gradient pruning threshold from the previous round. It is the L2 norm of the gradient of the project's latent factor vector;
[0036] S44: Gradient of latent factor vectors for items exceeding the clipping threshold Cutting:
[0037]
[0038] S45: Aggregate gradients after each training round:
[0039]
[0040] S46: After local training is complete, each group uses a mean of 0 and a variance of... Gaussian noise The noise gradient after perturbation is obtained. :
[0041]
[0042] in, It is the noise multiplier;
[0043] S47: Different groups dynamically adjust the gradient clipping threshold for the next round based on the average gradient norm and clipping ratio of the current round. :
[0044]
[0045] in, This is the average gradient norm ratio value for this round. This is the proportion of the gradient that needs to be clipped in this round. It is the average gradient norm of this round. ,and , Defined as the average gradient norm calculated after local population pre-training;
[0046] S48: The noise multiplier will be modified when the average gradient norm ratio decreases steadily for three consecutive rounds. :
[0047]
[0048] in, For the clipping factor, .
[0049] Furthermore, in step S5, each user group First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix :
[0050]
[0051]
[0052] The beneficial effects of this invention are as follows: For scenarios involving untrusted recommendation servers in recommendation systems, this invention proposes a localized differential privacy recommendation method and system that balances algorithmic fairness. This invention first divides users into groups based on their activity levels and performs matrix factorization on each group. Then, it introduces noise to prevent user privacy leaks caused by untrusted servers. To address the issue of privacy protection measures exacerbating performance differences between groups, it reduces gradient bias through dynamic gradient pruning thresholds and combines this with adaptive noise adjustment to control noise introduction. Ultimately, while effectively preventing privacy leaks, it alleviates performance imbalances between groups, achieving a balance between privacy protection, algorithmic fairness, and recommendation performance, demonstrating significant technical advantages.
[0053] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0055] Figure 1 This is a schematic diagram of the main workflow of the present invention.
[0056] Figure 2 The chart shows a comparison of the RMSE and MAE of this invention and various privacy-preserving recommendation algorithms on the Movielens100K dataset.
[0057] Figure 3 This is a comparison chart of RMSE and MAE of the present invention and various privacy-preserving recommendation algorithms on the Movielens1M dataset.
[0058] Figure 4 The chart shows a comparison of the RMSE and MAE of this invention and various privacy-preserving recommendation algorithms on the Yahoo dataset.
[0059] Figure 5 This is a comparison chart of the fairness of this invention with different algorithms on the Movielens100K dataset.
[0060] Figure 6 This is a comparison chart of the fairness of this invention with different algorithms on the Movielens1M dataset.
[0061] Figure 7 This is a comparison chart of the fairness of this invention with different algorithms on the Yahoo dataset. Detailed Implementation
[0062] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0063] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0064] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0065] Please see Figure 1 Specifically, it includes the following steps:
[0066] Step 1: Before starting pre-training in the iteration, divide the dataset into three groups: active, moderate, and inactive, based on user activity.
[0067] Step 2: Initialize the user latent factor matrix and project latent factor matrix To mitigate the performance degradation caused by differential privacy, the user group In training Previously Perform local pre-training;
[0068] Step 3: At the beginning of each iteration, the recommendation server will update the item latent factor matrix. Distribute to various groups;
[0069] Step 4: Calculate the item latent factor vector for all user-item pairs in the group. gradient:
[0070]
[0071] Step 5: Count the number of gradients exceeding the clipping threshold in this group:
[0072]
[0073] in, It is the gradient pruning threshold from the previous round. It is the L2 norm of the gradient of the project's latent factor vector;
[0074] Step 6: Gradient of latent factor vectors for items exceeding the clipping threshold Cutting:
[0075]
[0076] Step 7: Aggregate the gradients after each round of training:
[0077]
[0078] Step 8: After local training is complete, each group uses a mean of 0 and a variance of... Gaussian noise The noise gradient after perturbation is obtained. :
[0079]
[0080] in, It is the noise multiplier;
[0081] Step 9: Different groups dynamically adjust the gradient clipping threshold for the next round based on the average gradient norm and clipping ratio of the current round. :
[0082]
[0083] in, This is the average gradient norm ratio value for this round. This is the proportion of the gradient that needs to be clipped in this round. It is the average gradient norm of this round. ,and , Defined as the average gradient norm calculated after local population pre-training;
[0084] Step 10: The noise multiplier will be modified when the average gradient norm ratio decreases steadily for three consecutive rounds. :
[0085]
[0086] in, For the clipping factor, ;
[0087] Step 11, Each User Group First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix :
[0088]
[0089]
[0090] Repeat steps 3-11 until the iteration is complete, and obtain the final result. ;
[0091] Step 12: After the iteration is complete, generate the final project latent factor matrix. Delivered to each group, and to the users in each group. Calculate the predicted value on its device , Unobserved interactions are sorted in descending order of predicted values, and the top-K items are then recommended to the user. .
[0092] Example
[0093] The privacy-preserving distributed implicit feedback matrix factorization recommendation algorithm of this invention specifically includes the following steps:
[0094] Step 1: Before starting pre-training in the iteration, divide the dataset into three groups: active, moderate, and inactive, based on user activity.
[0095] Step Two: Each User Group Holding your own interaction data and user latent factor matrix Initialize the user latent factor matrix and project latent factor matrix To mitigate the performance degradation caused by differential privacy, the user group In training Previously Perform local pre-training;
[0096] Step 3: At the beginning of each iteration, the recommendation server will update the item latent factor matrix. Distribute to various groups;
[0097] Step 4: Save and train the user latent factor matrix locally for each group. Calculate the gradient To prevent pruning bias in different activity groups from causing differences in recommendation performance, [the following measures are taken]. Dynamic gradient clipping is performed based on different activity groups; after aggregating the clipping gradients for each group, adaptive noise addition is achieved by combining a dynamic clipping threshold. To achieve privacy protection and alleviate the problem of exacerbated recommendation performance differences between groups with different activity levels caused by privacy protection technologies, the threshold for gradient pruning in the next round is recalculated. and noise scale ;
[0098] Step 5: Different User Groups First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix ;
[0099] Step Six: After the iteration is complete, generate the final project latent factor matrix. Delivered to each group, and to the users in each group. Calculate the predicted value on its device , Unobserved interactions are sorted in descending order of predicted values, and the top-K items are then recommended to the user. .
[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A localized differential privacy recommendation method and system that balances algorithmic fairness, characterized in that, Includes the following steps: S1: Before the pre-training iteration begins, the dataset is divided into three groups: active, moderate, and inactive, based on user activity. S2: Each user group Holding your own interaction data and user latent factor matrix Initialize the user latent factor matrix and project latent factor matrix To mitigate the performance degradation caused by differential privacy, the user group In training Previously Perform local pre-training; S3: At the beginning of each iteration, the recommendation server will update the item latent factor matrix. Distribute to various groups; S4: Each group stores and trains the user latent factor matrix locally. Calculate the gradient To prevent pruning bias in different activity groups from causing differences in recommendation performance, [the following measures are taken]. Dynamic gradient clipping is performed based on different activity groups; after aggregating the clipping gradients for each group, adaptive noise addition is achieved by combining a dynamic clipping threshold. To achieve privacy protection, while mitigating the problem of exacerbated recommendation performance differences between groups with different activity levels caused by privacy protection technologies, the threshold for gradient pruning in the next round is recalculated. and noise scale ; S5: Different User Groups First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix ; S6: After the iteration is complete, the final project latent factor matrix will be generated. Delivered to each group, and to the users in each group. Calculate the predicted value on its device , Unobserved interactions are sorted in descending order of predicted values, and the top-K items are then recommended to the user. .
2. The localized differential privacy recommendation method and system that balances algorithmic fairness according to claim 1, characterized in that: In step S2, each user group Holding your own interaction data and user latent factor matrix Initialize the user latent factor matrix and project latent factor matrix To mitigate the performance degradation caused by differential privacy, the user group In training Previously Perform local pre-training; the algorithm will generate a user group rating matrix for the project. User-project pairs dataset The learning rate of the pre-trained user latent factor matrix Number of iterations As input, the user latent factor matrix is output after training. ; Specifically, the following steps are included: S21: First, randomize the user latent factor matrix. Project latent factor matrix ; S22: During each training session, for each user in the user group... Every interactive project Used to minimize the loss function: ， in, Represented as user For the project Real ratings and Representing users respectively and projects latent factor vectors, and They are collections of users and projects, respectively. For user-project interaction with datasets, and The regularization coefficient is used to prevent overfitting. It is updated iteratively through stochastic gradient descent to obtain a more accurate representation of user preferences and item features. S23: Calculate the user latent factor vectors sequentially. Project latent factor vector gradient: ， ； S24: For each user's latent factor vector Project latent factor vector To perform the update, use stochastic gradient descent and iteratively update according to the following rules: ， ， in, It is the learning rate; Repeat steps S22-S24. After each iteration, the trained user latent factor matrix is output. .
3. The localized differential privacy recommendation method and system that balances algorithmic fairness according to claim 1, characterized in that: In step S4, each group locally saves and trains the user latent factor matrix. Calculate the gradient To prevent pruning bias in different activity groups from causing differences in recommendation performance, [the following measures are taken]. Dynamic gradient clipping is performed based on different activity groups; after aggregating the clipping gradients for each group, adaptive noise addition is achieved by combining a dynamic clipping threshold. To achieve privacy protection, while mitigating the problem of exacerbated recommendation performance differences between groups with different activity levels caused by privacy protection technologies, the threshold for gradient pruning in the next round is recalculated. and noise scale ; Specifically, the following steps are included: S41: Train for each local group, minimizing the information about... Loss function: ； S42: Calculate the item latent factor vector for all user-item pairs in this group. gradient: ； S43: Count the number of gradients exceeding the clipping threshold in this group: ， in, It is the gradient clipping threshold from the previous round. It is the L2 norm of the gradient of the project's latent factor vector; S44: Gradient of latent factor vectors for items exceeding the clipping threshold Cutting: ； S45: Aggregate gradients after each training round: ； S46: After local training is complete, each group uses a mean of 0 and a variance of... Gaussian noise The noise gradient after perturbation is obtained. : ， in, It is the noise multiplier; S47: Different groups dynamically adjust the gradient clipping threshold for the next round based on the average gradient norm and clipping ratio of the current round. : ， in, This is the average gradient norm ratio value for this round. This is the proportion of the gradient that needs to be clipped in this round. It is the average gradient norm of this round. ,and , Defined as the average gradient norm calculated after local population pre-training; S48: The noise multiplier will be modified when the average gradient norm ratio decreases steadily for three consecutive rounds. : ， in, For the clipping factor, .
4. The localized differential privacy recommendation method and system that balances algorithmic fairness according to claim 1, characterized in that: In step S5, each user group First Uploaded to the recommendation server, which then collects all the data. Summation yields Finally, we recommend using the server. Update the project latent factor matrix : ， 。