Federated Recommendation Method and System Based on Optimal Transport between Groups and Pseudo-Sampling Difference

Abstract

The present invention belongs to the technical field of federated recommendation, and discloses a federated recommendation method and system based on optimal transport between groups and pseudo-sampling difference. The present invention uses the optimal transport mechanism to alleviate the data sparsity of disadvantaged users and narrow the training gap between advantaged users with large amounts of data and disadvantaged users with small amounts of data. Adopting a federated training method of local model training and joint global model update, it makes full use of the data of each user to improve the recommendation quality. It realizes double privacy protection, protecting both the original user data and the uploaded gradient information, thus meeting the requirements of privacy protection. Through these innovative designs, this method can effectively solve the problem of data sparsity of user data in the federated recommendation model while protecting privacy, and at the same time solve the problems of recommendation accuracy and fairness caused by sparsity.

Description

Technical Field

[0001] The present invention belongs to the technical field of federated recommendation, and particularly relates to a federated recommendation method and system based on optimal transport between groups and pseudo-sampling difference. Background Art

[0002] Today, with the rapid development of information technology, recommendation systems have become an indispensable tool in people's daily lives. From content recommendations on short video platforms, to product recommendations on e-commerce websites, to talent recommendations on job hunting platforms, recommendation systems help users quickly discover interesting content in various fields and improve the user experience.

[0003] Traditional recommendation systems usually adopt a centralized architecture, that is, the user's personal data, such as product browsing records and movie rating records, are centrally stored on a central server, and then a recommendation model is trained based on this data, and finally personalized recommendation results are generated. Although this centralized data processing method can make full use of the user's full amount of data and improve the accuracy of recommendations, it also brings a series of privacy protection problems. First, once the user's personal information is centrally stored on the central server, it is vulnerable to hacker attacks or abuse by service providers, resulting in the risk of privacy leakage. With the increasing attention of people to personal privacy rights, some countries and regions have formulated relevant laws and regulations. For example, the General Data Protection Regulation requires service providers to obtain the explicit consent of users before collecting and using personal data, and cannot let users waive their privacy rights in a deceptive or induced manner. The introduction of such regulations has undoubtedly brought great challenges to traditional centralized recommendation systems. Second, the centralized data processing mode may also lead to users' distrust of service providers. Users often worry that their privacy information may be leaked or abused by service providers, thereby reducing their enthusiasm for use.

[0004] To solve these problems, federated recommendation technology has emerged. The core idea of federated recommendation technology is based on the federated learning framework. Federated learning is a new distributed machine learning method that aims to solve the problem of data privacy protection. In federated learning, each participant involved in training, such as the user's mobile device or personal computer, can retain its own original data without the need to upload the data to the central server. At the same time, each device independently trains its own local model, and then transmits the model parameters to the central server for aggregation. In this way, the central server can integrate the model parameters of all parties to generate a global prediction model without accessing any original user data. This not only greatly reduces the risk of privacy leakage but also improves the overall performance of the model. The process of federated learning usually includes the following steps:

[0005] 1. Data Preparation: Instead of centralizing the dataset, each client independently retains its own data. It can be federated training based on geographical location, such as using data from different institutions like hospitals or banks; or it can be federated training based on user IDs, that is, the data of different users are saved on their respective devices for joint training.

[0006] 2. Model Training: Each client independently trains its own model without sharing the original data. Each client uses its own data to train a local model, which can be various models such as deep learning models, statistical models, and recommendation system models. During the training process, the client does not upload the original data to the central server but only uploads the trained model parameters. This can effectively protect the original data and further protect user privacy.

[0007] 3. Model Aggregation: The model parameters trained by each client are encrypted and uploaded to the central server, and the central server performs model aggregation to obtain a global model. The central server collects the model parameters uploaded by each client and uses an aggregation algorithm to combine them into a global model. This global model synthesizes the learning achievements of each client and represents the final model of the entire federated network.

[0008] 4. Model Update: After completing the model aggregation, the central server sends the updated global model parameters back to each client. In this way, each client can use this new global model to continue the next round of model training and update.

[0009] 5. Iterative Update: Repeat steps 2 - 4 until the model converges or reaches the preset number of training rounds. The entire training process of federated learning is an iterative process. Each client continuously trains the local model, uploads the parameters to the central server for aggregation, and then the updated global model is sent back to each client. This process repeats until the model performance reaches the expected goal or the number of training rounds is exhausted.

[0010] In the federated recommendation system, the idea of federated learning is well applied. The personal data of each user is stored on their local device, and each user device independently trains its own recommendation model, and then transmits the parameters of the recommendation model to the central server for aggregation to generate a comprehensive global recommendation model. This distributed training method can not only enable the recommendation model to better protect user privacy but also shows significant advantages over traditional recommendation systems in terms of improving recommendation accuracy and response speed, bringing more secure, fair, personalized, and timely recommendation services to people.

[0011] Based on the above advantages of federated learning, a large number of researchers have been attracted to explore its usability and potential. However, there are still many problems to be solved in the field of federated recommendation to promote the further development and application of federated recommendation technology. Among them, the problems that cannot be ignored include but are not limited to:

[0012] First, the federated recommendation model will always be dominated by the user group that contributes more to the model update process, and the recommendation results will be more biased towards this part of users. This leads to poor overall recommendation performance for users and shows significant unfairness. The root cause of this problem is that most users suffer from data sparsity, so they cannot get satisfactory recommendation results. Data sparsity not only affects the recommendation performance of the model, but also limits its generalization ability, making it difficult for the model to cope with new users and products. To solve this problem, the key is to enrich the user portrait using cross-device data integration methods. For example, the idea of ​​federated transfer learning can be used to migrate model parameters trained on other devices to the current device to enhance the model's learning ability for sparse data. It can also be combined with federated distillation technology to allow models between different devices to learn from each other, complement each other's knowledge, and further improve recommendation performance. It is also possible to use the method of expanding the federated graph neural network to use graph structure information to capture the implicit relationship between users and items. In this method, the optimal transmission matrix is ​​calculated by the optimal transmission method, clustering is used to help identify groups with similar user data distribution, and the content of users with non-sparse data is shared. Then, data-sparse users use these shared contents to supplement their own data deficiencies, thereby alleviating the data sparsity problem and improving the model's learning ability in data-sparse environments. Through data sharing and privacy protection between clients, as well as the above innovative solutions, the federated recommendation model can truly improve the overall quality of recommendation services and bring a better experience to users.

[0013] Second, as an emerging distributed machine learning framework, federated learning does provide better guarantees in terms of privacy protection. However, during the model update process, there are still some risks of privacy leakage. First, during the gradient transmission stage, a malicious central server may use the embedded gradient information of non-zero terms to attempt to recover the interaction history or certain sensitive attributes of users and items, thus causing privacy leakage. Second, in a federated recommendation system, users need to upload the embedded representations of the entire item set to protect the user interaction history, and there are also privacy issues in this process. To further protect the privacy of data and models, researchers have proposed some new methods. One method is homomorphic encryption technology, which can be applied to gradient transmission to effectively protect user privacy. In this method, the user device encrypts the gradient locally and then uploads it to the central server. The central server cannot directly obtain the original gradient information, thus avoiding privacy leakage. Another method is to use differential privacy technology, which can inject appropriate noise when training the model, thereby reducing the risk of privacy information leakage from the model. This method can protect users' privacy to a certain extent without having too much impact on the model performance. This is the method selected in this paper.

[0014] Third, most existing recommendation systems perform personalized recommendations based on factors such as users' historical behavior data, social relationships, and geographical locations. Although this individual feature-based recommendation method can improve the user experience, it may also produce some unfair results. For example, for some users with sparse data, due to their limited historical behavior data or differences in data characteristics in their regions from the mainstream group, it is difficult for the recommendation system to accurately depict their preferences and needs, which may lead to poor-quality content recommended for them, exacerbating social unfairness. Therefore, how to effectively model disadvantaged groups using limited data is a key issue to be solved. Summary of the Invention

[0015] Aiming at the problems existing in the prior art, the present invention provides a federated recommendation method based on optimal transport between groups and pseudo-sampling differential. Combining the above technical solutions and the technical problems to be solved, the advantages and positive effects of the technical solution to be protected by the present invention are analyzed from the following aspects:

[0016] First, regarding the technical problems existing in the above prior art, some creative technical effects brought about after solving the problems are as follows. The specific description is as follows:

[0017] The present invention proposes a federated recommendation method based on optimal transport between groups and double privacy protection. First, aiming at the common data sparsity problem in the federated recommendation system, an innovative optimization design is adopted, that is, an optimal transport mechanism between groups is introduced. First, the optimal transport matrix and similarity are calculated between dominant users and disadvantaged users, then clustering is performed with dominant users as the clustering centers, and then the recommendation results of dominant users in the same clustering result are shared with disadvantaged users, effectively alleviating the impact brought by data sparsity. Through the above steps, this method ensures that the federated recommendation system can maintain a high recommendation accuracy even in the case of sparse user data.

[0018] In addition, the present invention also conducts in-depth design and implementation for the privacy protection problem in the federated recommendation process. This method adopts a double privacy protection mechanism. First, this method uses the pseudo-interaction item sampling technique to generate pseudo-data with a similar distribution for the training and update of the federated model without revealing the original data of the participants. Second, in the process of model training and update, this method adopts differential privacy technology. First, gradient clipping can ensure that the maximum change of each gradient element does not exceed a certain threshold to prevent privacy leakage caused by individual elements being too large. Then, noise is added to the original data to make the connection between the query result and personal data blurred, thus effectively preventing the client from leaking each other's user privacy data. This double privacy protection mechanism not only protects user privacy but also ensures the security of the entire federated recommendation process.

[0019] In addition to accuracy and privacy protection, the present invention also makes an important contribution to the fairness of recommendations. In the design of the recommendation model, this method incorporates fairness considerations to ensure that different user groups can obtain a fair recommendation experience. Specifically, the recommendation model is often dominated by dominant users with a larger amount of data, and there is a deviation from disadvantaged users with a smaller number. The present invention aims to enable the federated recommendation system to balance accuracy and fairness and provide fair and reasonable recommendation services for all participating users.

[0020] Specifically, first, users are divided into dominant users and weak users according to the number of interactions between users and items. Dominant users refer to those who have rich interactions with a large number of items on the platform, while weak users are the group of users with fewer interactions with items. This difference in data distribution will lead to obvious fairness problems in the training and application of the recommendation model. Next, the data sparsity optimization model uses a novel optimal transport mechanism to find the most similar dominant user for each weak user by calculating the similarity between dominant users and weak users on shared interaction items. Then, the embeddings of these dominant users are sent to the corresponding weak users. And the embedding distance between the weak users and the similar dominant users is minimized to narrow the training gap between the two types of user groups. At the same time, each user device locally learns a recommendation model. This model learns the embeddings of users and items based on the local user and item interaction data on the device. During the model training process, the user device calculates the model gradients and the user and item embeddings, and then uploads these parameters to the central server. The central server receives and aggregates the parameters of each user to generate a combined global model, and then distributes the aggregated global model to each user for subsequent iterations. To protect user privacy, a dual privacy protection mechanism is adopted during the model update process: on the one hand, when uploading the gradients of item embeddings, a certain number of pseudo-interaction item embedding gradients are randomly generated and combined with the real gradients. This can prevent the server from identifying the real gradient information, thus protecting user privacy. On the other hand, differential privacy technology is used to protect user privacy when calculating local gradients on the user client side. Thus, it is ensured that personal data will not be leaked during the statistical analysis process, thereby meeting the privacy protection requirements.

[0021] In summary, the design of the present invention includes:

[0022] 1. Using the optimal transport mechanism to alleviate the data sparsity of weak users and narrow the training gap between dominant users with large amounts of data and weak users with small amounts of data.

[0023] 2. Adopting a federated training method of local model training and joint global model update to make full use of the data of each user and improve the recommendation quality.

[0024] 3. Implementing dual privacy protection to protect both the original user data and the uploaded gradient information, thus meeting the privacy protection requirements.

[0025] Through these innovative designs above, this method can effectively solve the problem of data sparsity of user data in the federated recommendation model while protecting privacy, and at the same time solve the recommendation accuracy and fairness problems caused by sparsity.

[0026] Second, as the creative auxiliary evidence of the claims of the present invention, it is also reflected in the following important aspects:

[0027] (1) The expected benefits and commercial value after the transformation of the technical solution of the present invention are as follows:

[0028] After the transformation of the present invention, it can bring greater commercial and economic value and market competitiveness to companies such as retail enterprises, e-commerce platforms, and video platforms.

[0029] First, it can improve the accuracy and pertinence of product recommendations, enhance the user experience, increase the conversion rate and customer stickiness. It is mainly reflected in: ① For users with sparse data, simulate their actual interests and hobbies, discover similar users, overcome the problem of sparse data, and improve the recommendation accuracy rate. ② By optimizing the recommendation model, it can more accurately capture the potential needs and interest preferences of users, improve the purchase intention and consumption amount of users, thereby increasing the revenue of the enterprise. ③ By increasing user satisfaction, that is, increasing user stickiness, it is conducive to the enterprise to maintain and develop a long-term stable customer group and increase market competitiveness.

[0030] Second, by protecting user privacy, enhancing user trust, and further expanding the user group. It is mainly reflected in: ① By using the architecture of federated learning, user data has always been only retained locally and not uploaded to the server, which greatly reduces the concentration and sharing of user data and effectively prevents the risk of privacy leakage. ② Through the pseudo-interaction item sampling technology and local differential privacy technology, the privacy security of users is further protected, and the trust of users in the platform is enhanced. ③ As user trust increases, the reputation and credibility of the platform will also continuously improve, which will attract the attention and participation of more potential users and expand the user group of the platform. A large user base provides a solid foundation for the future development and profitability of the platform.

[0031] Third, it can significantly improve the efficiency of the recommendation strategy and further improve the overall operation efficiency of software development companies. It is mainly reflected in: ① By more accurately capturing user needs, increasing the conversion rate and user repeat purchase rate, and reducing the customer acquisition cost of a single user. ② By better exerting the product value, improving the marketing conversion efficiency, thereby enhancing the overall operation efficiency.

[0032] In short, the present invention can provide more targeted recommendations for users according to user needs, while protecting user privacy and improving the efficiency of recommendations, thereby enabling the company to obtain more and more stable users and enhancing the overall operation efficiency of the company.

[0033] (2) The technical solution of the present invention fills the domestic and foreign technical gaps in the industry:

[0034] The present invention innovates the traditional software development process from three aspects: data optimization, privacy protection, and recommendation performance. It provides recommendation services for users without collecting user data, which is closer to user needs and actual business scenarios, breaks through the traditional centralized construction mode, and effectively improves the company's data processing ability. This solves the limitations of the prior art, fills an important technical gap in the field of federated recommendation, and has strong application value.

[0035] (3) Does the technical solution of the present invention solve the technical problems that people have been eager to solve but have never succeeded in?

[0036] ① In the previous federated recommendation scenario, due to the dispersion of data in different clients, there was a relatively serious data sparsity problem, which affected the recommendation effect. ② Federated learning needs to update the model and exchange data while protecting privacy, which has always been a major technical problem. ③ How to improve the accuracy and fairness of recommendations while protecting privacy has always been a key technical challenge in federated recommendation. The present invention adopts the method of optimal transport and clustering, effectively reducing the training gap between different user groups and solving the technical bottleneck brought by data sparsity. At the same time, a dual privacy protection technology is proposed, including pseudo-interaction item sampling and local differential privacy, effectively improving the security of federated learning and solving the privacy protection problem. In addition, the present invention also takes into account the fairness of recommendations for different users, improving the user experience. Description of the Drawings

[0037] Figure 1 It is a flowchart of the federated recommendation method based on optimal transport between groups and pseudo-sampled difference provided by an embodiment of the present invention.

[0038] Figure 2 It is a block diagram of the structure of the federated recommendation system based on optimal transport between groups and pseudo-sampled difference provided by an embodiment of the present invention.

[0039] Figure 3 It is a cross-device federated learning model diagram provided by an embodiment of the present invention.

[0040] Figure 4 It is a model diagram of data sparsity optimization based on optimal transport between groups provided by an embodiment of the present invention.

[0041] Figure 5 It is a model diagram of federated recommendation under dual privacy protection based on pseudo-sampled difference provided by an embodiment of the present invention;

[0042] Figure 6 It is an effect diagram of federated learning after sparsity optimization using optimal transport provided by an embodiment of the present invention;

[0043] Figure 7It is the effect diagram of privacy protection under different noise intensities and clipping thresholds of differential privacy provided by the embodiments of the present invention. Detailed implementation manners

[0044] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] As Figure 1 shown, a federated recommendation method based on optimal transport between groups and pseudo-sampled differential privacy provided by the embodiments of the present invention includes the following steps:

[0046] S1. Optimize the data sparsity model based on the optimal transport theory, which specifically includes the following content:

[0047] S101. Construct one-hot interaction embedding matrices H D and H A ;

[0048] S102. Solve the Monge-Kantorovich problem with smoothed Sinkhorn divergence;

[0049] S103. Send the embeddings of the dominant users back to the corresponding client of the disadvantaged users, and let the disadvantaged users learn from the dominant users to enhance the training of the disadvantaged users when training the backbone recommendation model;

[0050] S104. Subsequently, combine these two loss functions as the loss function for training the model of the disadvantaged users.

[0051] S2. Dual privacy protection model based on pseudo-sampled differential privacy, which specifically includes the following content:

[0052] S201. Pseudo-interaction item sampling;

[0053] S202. Based on local differential privacy protection technology.

[0054] As Figure 2 shown, a federated recommendation system based on optimal transport between groups and pseudo-sampled differential privacy provided by the embodiments of the present invention includes:

[0055] A construction module, used to construct one-hot interaction embedding matrices of two groups and;

[0056] A solution module, used to solve the Monge-Kantorovich problem with smoothed Sinkhorn divergence;

[0057] A training module, configured to send the embeddings of the advantaged users back to the corresponding disadvantaged user clients, and enable the disadvantaged users to learn from the advantaged users during the training of the backbone recommendation model, so as to enhance the training of the disadvantaged users;

[0058] A combining module, configured to combine the two loss functions as the loss function for training the model of the disadvantaged users subsequently.

[0059] S1, a data sparsity optimization model based on optimal transport theory.

[0060] Further, the specific process of S101 is as follows:

[0061] In an embodiment of the present invention, a one-hot interaction embedding matrix H of two groups is constructed D and H A . First, the user-item interaction records are represented as a binary matrix, where the rows represent users and the columns represent items. If user i has an interaction with item j, the corresponding position in the matrix is 1, otherwise it is 0. Based on this interaction matrix, one-hot interaction embedding matrices are constructed for the advantaged user group A and the disadvantaged user group D respectively.

[0062] Further, the specific process of S102 is as follows:

[0063] The embodiment of the present invention provides a solution to the Monge-Kantorovich problem of smoothing the Sinkhorn divergence: The problem description is given a transportation cost matrix, and the goal is to find a joint probability such that the total transportation cost of transferring from one distribution to another is minimized;

[0064]

[0065] is a cost matrix based on the transportation cost, h D and h A are two embeddings sampled from H D and H A , is a joint probability, X ij represents the possibility of transferring user u in D i to user u in A j , reflecting the similarity between u i and u j , and e is a regularization factor;

[0066] For the goal of clustering each disadvantaged user into a specific advantaged user, the following constraint conditions related to X are designed:

[0067]

[0068] Let S represent the clustering result, Sij Indicates D i Belongs to the cluster centered on A j For the cluster centered on A

[0069] Furthermore, the specific process of S103 is as follows:

[0070] In the embodiment of the present invention, the embedding of the dominant user is sent back to the corresponding client of the disadvantaged user. When training the backbone recommendation model, the disadvantaged user learns from the dominant user to enhance the training of the disadvantaged user; the inter-group loss is as follows:

[0071]

[0072] Where Ti represents the cluster center of Di, and E represents the user embedding; by minimizing L group , the disadvantaged user can learn from similar dominant users, and both will obtain similar distributions, thereby improving the recommendation performance of the disadvantaged user.

[0073] Furthermore, the specific process of S104 is as follows:

[0074] In the embodiment of the present invention, the following combines these two loss functions as the loss function for training the disadvantaged user model:

[0075] L = L utility + L group (4)

[0076] In the above steps, the original server never receives the privacy information of the user, and the third-party server cannot obtain the privacy information of any item because it cannot decrypt the item ID. However, by clustering each disadvantaged user client with its similar dominant user client, local users can allow the disadvantaged user to learn from the dominant user while protecting privacy, thereby effectively alleviating the problem of data sparsity.

[0077] S2, a dual privacy protection model based on pseudo-sampling difference.

[0078] The local subgraph on each user client is constructed based on user and item interaction data. First, an embedding layer is used to obtain the embeddings of the user and the item. Then, a rating predictor module is used to predict the rating given by the user to the item they interact with based on the embeddings of the user and the item. These predicted ratings are compared with the original ratings stored on the user device to calculate the loss function. For user u i , the loss function L i Is calculated as Then use the loss L i To derive the model gradient And the embedding gradient Then these gradients are aggregated on the central server. During the aggregation process, the server coordinates all user devices and calculates the global gradient to update the model and embed parameters in these devices. In each round, the server wakes up a certain number of user clients to calculate gradients locally and send them to the server. After the server receives the gradients of these users, the aggregator in the server aggregates these local gradients into a unified gradient. Then, the server sends the aggregated gradient to each client participating in the training, allowing each client to update its local model parameters with this aggregated gradient. During the model update process, the federated recommendation framework adopts two main privacy protection strategies to protect the privacy information of users.

[0079] Furthermore, the specific process of S201 is as follows:

[0080] Pseudo-interaction item sampling is a privacy protection technology used in the federated recommendation system. Its core idea is to generate some artificially synthesized pseudo-interaction items as negative samples of the training data. In addition, it can also make up for the sparsity of the actual user-item interaction data to a certain extent. Specifically, in the present invention, pseudo-interaction item sampling consists of the following two steps:

[0081] (1) Sample the items that the user has not interacted with but has interacted with before to obtain M pseudo-interaction items, which are assumed that the user has interacted with them before. According to the items that the user has actually interacted with, calculate their gradient statistical features, including the mean and variance. Using the Gaussian distribution, randomly generate the gradients of M pseudo-interaction items according to the mean and variance calculated in the previous step. The gradients of the pseudo-interaction items generated in this way should be similar to the gradient distribution of the real interaction items.

[0082]

[0083] Among them, N represents the Gaussian distribution, μ e is the mean of the real interaction item gradients ∑ e is the covariance of the real interaction item gradients

[0084] (2) Combine the real embedding gradients with the pseudo-item embedding gradients and finally modify the model and embedding gradients on the i-th user as follows:

[0085]

[0086] Furthermore, the specific process of S202 is as follows:

[0087] ​Under the federated recommendation model, each user locally stores their own data and only uploads the encrypted model parameters and gradients to the central server. The central server cannot obtain the original data of the users and can only obtain the parameters and gradients after privacy processing. This can not only effectively protect user privacy but also reduce the computational pressure on the central server. The working principle of local differential privacy is as follows: First, each user locally performs random perturbation processing on the data to be uploaded. This random perturbation process satisfies the definition of differential privacy and can effectively protect user privacy. Then, the user uploads the encrypted data to the server. Finally, the server performs statistical analysis based on this privacy-processed data to obtain some valuable results. This method performs two-step processing on the local gradient on the user client:

[0088] (1) Gradient clipping: The main function of this step is to limit the magnitude of the gradient vector uploaded by each user to the server, thus effectively preventing privacy leakage caused by individual elements being too large. During the training process of a machine learning model, each user client calculates the local gradient g i , which represents the sensitivity of the model parameters to the user's data. Without any processing, these gradient vectors may contain some extremely large elements that may disclose the sensitive information of the users. To solve this problem, the gradient clipping technique limits the L ∞ -norm of the gradient vector. In this process: First, calculate the L i -norm of the gradient vector g ∞ , that is, find the element with the largest absolute value in g i . Then set a predefined threshold δ. If the L i -norm of g ∞ exceeds δ, scale all elements in g i proportionally so that the L ∞ -norm is exactly equal to δ. This process is shown as follows:

[0089]

[0090] Finally, upload the scaled gradient vector to the server for model update. This gradient clipping technique can ensure that the maximum change of each gradient element does not exceed δ, thus effectively preventing privacy leakage caused by elements being too large. At the same time, it also ensures the stability of model training and avoids model oscillation or divergence caused by a few extremely large gradients. It should be noted that the choice of δ needs to balance privacy protection and model performance. A too-small δ value will overly compress the gradient and affect the model convergence speed; while a too-large δ value cannot effectively protect privacy. Therefore, it is usually necessary to reasonably set δ according to the specific application scenario and privacy requirements.

[0091] (2) Noise addition: Its basic principle is to superimpose a certain amount of random noise on the gradient vector after gradient clipping processing, so as to effectively obscure the original gradient information and achieve the purpose of privacy protection. Specifically, a random noise obeying the Laplace distribution is superimposed on each gradient. The mean of the Laplace noise is 0, and the standard deviation is a pre-set privacy budget parameter λ, which is used to control the intensity of the Laplace noise. By adding this zero-mean Laplace noise, the sensitive information in the gradient vector can be effectively obscured. Even if an attacker observes the gradient value uploaded to the server it is very difficult to infer the user's original data. This is because the Laplace noise can fully cover the true value of the original gradient.

[0092]

[0093] The privacy protection ability of the noise addition technique mainly depends on the intensity of the noise, that is, the size of the privacy budget parameter. The smaller the value, the greater the added noise, and the better the privacy protection effect. However, it will also reduce the convergence speed and final performance of model training. Therefore, in the actual use process, a trade-off needs to be made between privacy protection and model performance. After the above processing, by combining the two techniques of gradient clipping and noise addition, a relatively complete local differential privacy protection framework can be formed. The user obtains a local gradient that contains both useful information of the user data and is difficult to be used to infer the original private data.

[0094] This combined use method can give full play to the respective advantages of the two techniques, so as to achieve more comprehensive and reliable privacy protection.

[0095] Finally, this clipped and noise-added gradient value will be uploaded to the server for aggregation to update the model parameters. In this way, while protecting user privacy, the advantages of distributed data can be fully utilized for effective recommendation. The aggregation method is shown as follows:

[0096] where K is the total number of users participating in the training, and G is the finally aggregated gradient. In summary, this method realizes a method for designing and implementing federated recommendation based on optimal transport between groups and pseudo-sampling difference to address the data sparsity problem and protect user privacy data that recommendation systems often face in practical applications, while taking into account multiple goals of recommendation accuracy and fairness.

[0097]

[0098]

[0099] In summary, this method realizes a method for designing and implementing federated recommendation based on optimal transport between groups and pseudo-sampling difference to address the data sparsity problem and protect user privacy data that recommendation systems often face in practical applications, while taking into account multiple goals of recommendation accuracy and fairness. ​

[0100] I. Specific application fields or related products of the present invention

[0101] Based on the experience of past recommendation systems and combined with current user data requirements and company requirements, the present invention is proposed under the premise of meeting privacy protection and fairness.

[0102] The present invention enables users from different regions to understand the system functions and requirements in detail and deeply participate in local model training. The present invention is adopted in an e-commerce shopping platform system, which usually has multiple distributed regional data centers storing purchase behavior data of users in different regions. By using the present invention, the platform can perform federated learning between different data centers, make full use of the purchase preferences of users in each region, and learn a global recommendation model. This can not only improve the recommendation effect, protect user privacy, ensure the user recommendation experience, but also relieve the network transmission pressure between data centers in different regions. In practical applications, this federated recommendation algorithm can help recommend more personalized and accurate products, improving user purchase satisfaction, conversion rate, and purchase experience.

[0103] II. Evidence related to the technical effects obtained in the embodiments of the present invention

[0104] The present invention proposes a federated recommendation method based on optimal transport between groups and pseudo-sampled differential privacy, aiming to solve the challenges of data sparsity and privacy protection in the federated scenario.

[0105] For the data sparsity problem, the present invention uses the optimal transport between groups technology for optimization. By discovering similar dominant users for each disadvantaged user, the disadvantaged users can learn from the corresponding similar dominant users. This method can effectively narrow the training gap between the two user groups, thereby improving the accuracy of the overall recommendation result. At the same time, through the clustering method, the feature vectors of disadvantaged users are aligned with those of dominant users, alleviating the data sparsity problem and improving the generalization ability of the model. Through this method, the optimization of the two metrics of NDCG and HR is as Figure 6 shown. For the user privacy protection problem, the present invention combines pseudo-interaction item sampling and differential privacy technology to ensure the anonymity of user interaction items and protect the gradient information transmitted by users in federated learning, ensuring data and model privacy. The results of differential privacy protection are as Figure 7 shown, where the privacy budget is a metric used to measure privacy loss, describing the ability of the algorithm to protect privacy. The smaller the value, the better the privacy protection. The clipping threshold controls the clipping degree of sensitive data uploaded to the central server. The smaller it is, the better the privacy protection. The noise intensity determines the noise intensity added to the uploaded data. The larger it is, the better the privacy protection.

[0106] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and their modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software, such as firmware.

[0107] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A federated recommendation method based on optimal transmission between groups and pseudo sampling difference, characterized in that: The following steps are involved: S1, data sparsity optimization model based on optimal transmission theory, specifically includes the following contents: S101, construct the one-hot interaction embedding matrix H of the two groups D and H A ; S102, solving the Monge-Kantorovich problem of Sinkhorn divergence smoothing; S103, sending the embedding of the dominant user back to the corresponding disadvantaged user client, and letting the disadvantaged user learn from the dominant user to enhance the training of the disadvantaged user when training the backbone recommendation model; dominant users refer to users who have rich interactions with a large number of items on the platform, while disadvantaged users refer to user groups who have less interactions with items; S104, subsequently combining the two loss functions as the loss function of the disadvantaged user training model; S2, a dual privacy protection model based on pseudo-sampling difference, specifically includes the following contents: S201, pseudo-interaction item sampling; pseudo-interaction item sampling is a technology used to protect privacy in federated recommendation systems, which generates some artificial pseudo-interaction items as negative samples of training data; S202, based on local differential privacy protection technology; in the federated recommendation model, each user saves his or her own data locally, and only encrypts the model parameters and gradients before uploading them to the central server; The Monge-Kantorovich problem for solving Sinkhorn divergence smoothing is described as follows: the problem is described as a given transmission cost matrix, the goal is to find a joint probability that minimizes the total transmission cost of transferring from one distribution to another; is a cost matrix based on transmission cost, h D and h A It is from H D and H A The two embeddings sampled, is a joint probability, X ij Indicates that user u in D i Transmitted to user u in A j The possibility of u i and u j The similarity between them, e is the regularization factor; Cluster each disadvantaged user into a specific advantaged user target and design the following restrictions related to X: The embedding of the dominant user is sent back to the corresponding disadvantaged user client, and when training the backbone recommendation model, the disadvantaged user is allowed to learn from the dominant user to enhance the training of the disadvantaged user; the inter-group loss is as follows: Where T represents the cluster center of Di, E represents the user embedding; by minimizing L group ; The following combines these two loss functions as the loss function of the disadvantaged user training model, where L utility The loss function inherent to the backbone recommendation model and used to optimize the recommendation accuracy: L=L utility +L group (4) 2. The federated recommendation method based on inter-group optimal transmission and pseudo-sampling difference as claimed in claim 1, characterized in that: The one-hot interaction embedding matrix H of the two groups is constructed D and H A : First, the user-item interaction records are represented as a binary matrix, where rows represent users and columns represent items. If user i interacts with item j, the corresponding position in the matrix is ​​1, otherwise it is 0. Based on this interaction matrix, one-hot interaction embedding matrices are constructed for the dominant user group A and the disadvantaged user group D respectively.

3. A federated recommendation system based on inter-group optimal transmission and pseudo-sampling difference implementing the federated recommendation method based on inter-group optimal transmission and pseudo-sampling difference as claimed in any one of claims 1 to 2, characterized in that: The federated recommendation system based on optimal transmission between groups and pseudo sampling difference includes: A building module for constructing the one-hot interaction embedding matrix of two groups; Solving module for solving the Monge-Kantorovich problem with Sinkhorn divergence smoothing; The training module is used to send the embedding of the dominant user back to the corresponding disadvantaged user client. When training the backbone recommendation model, the disadvantaged user can learn from the dominant user to enhance the training of the disadvantaged user. The joint module is used to subsequently combine the two loss functions as the loss function of the disadvantaged user training model.

4. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the federal recommendation method based on optimal transmission between groups and pseudo-sampling difference as described in any one of claims 1-2.

5. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the federal recommendation method based on optimal transmission between groups and pseudo-sampling difference as described in any one of claims 1-2.

6. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the federated recommendation system based on optimal transmission between groups and pseudo sampling difference as described in claim 3.

Images