Information recommendation method and device based on privacy protection

Abstract

The present invention provides an information recommendation method and device based on privacy protection. The method determines a baseline recommendation model and a privacy protection user obtained by training based on sample user data of all sample users. After deleting the sample user data of the privacy protection user, noise user data corresponding to the privacy protection user is generated based on the baseline recommendation model. The baseline recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to construct an information recommendation model. The knowledge learned by the baseline recommendation model based on the sample user data of the privacy protection user in its previous training process can be obfuscated, while maintaining the information recommendation capability for users other than the privacy protection user. Then, based on the information recommendation model, product information recommendation is performed for the user data of each non-privacy protection user and the product information of the product to be recommended, thereby realizing accurate information recommendation in the privacy protection mode.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic information processing, and in particular to an information recommendation method and device based on privacy protection. Background Art

[0002] In the field of e-commerce, recommendation systems play a vital role. Through intelligent algorithms and technologies, they provide personalized and accurate product and service recommendations, greatly improving the user's shopping experience. Especially in today's Internet world where everyone can be a dealer and there is a wide variety of product information, these systems analyze users' key behavioral data, including browsing history, purchase records, search keywords, product usage records, etc., and deeply explore users' preferences and interests, so that they can tailor recommendation content for each user and increase user stickiness.

[0003] Current recommendation systems use methods including collaborative filtering and content recommendation, as well as deep learning methods that can automatically learn key information in user data to adapt to more complex recommendation scenarios. Most of these methods require the collection and analysis of user personal information and behavior data, including sensitive information such as user personal interests, purchase history, and click behavior. Among them, for deep learning methods, although analyzing user data can improve the recommendation accuracy of the recommendation system, it also brings potential privacy risks. In particular, deep learning models need to learn a large amount of user privacy data and internalize the learning content into model parameters. For users who are concerned about privacy, there is a sense of insecurity about privacy leakage. Therefore, it is necessary to provide an information recommendation method based on privacy protection, which can enable the information recommendation model to quickly and specifically completely delete the user sensitive information that has been learned without affecting the overall model performance for users with privacy protection needs, thereby protecting the user's privacy while taking into account the performance of the recommendation system. Summary of the invention

[0004] The present invention provides an information recommendation method and device based on privacy protection, which are used to solve the defects of the privacy leakage problem in the prior art.

[0005] The present invention provides an information recommendation method based on privacy protection, comprising:

[0006] Get user data of each user in the system;

[0007] Inputting the user data of each user and the product information of the product to be recommended into the information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and performing product information recommendation based on the recommendation probability of the product to be recommended for each user;

[0008] The information recommendation model is trained based on the following steps:

[0009] Determine a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users;

[0010] After deleting the sample user data of the privacy-preserving user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy-preserving user based on the benchmark recommendation model;

[0011] The benchmark recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0012] According to a privacy protection-based information recommendation method provided by the present invention, generating the noise user data corresponding to the privacy protection user based on the benchmark recommendation model specifically includes:

[0013] Randomly add customized interest tags to each product; wherein, if the customized interest tag of any product is 1, it means that the privacy protection user is interested in any product, and if the customized interest tag of any product is 0, it means that the privacy protection user is not interested in any product;

[0014] Input the current noise data corresponding to the privacy protection user and the product information of each product into the benchmark recommendation model, obtain the recommendation probability of each product for the privacy protection user output by the benchmark recommendation model, determine the model loss based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product, and adjust the current noise data corresponding to the privacy protection user based on the model loss until the model loss is less than a preset value;

[0015] The current noise data corresponding to the privacy protection user is determined as the noise user data corresponding to the privacy protection user.

[0016] According to a privacy protection-based information recommendation method provided by the present invention, the method of determining the model loss based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product specifically includes:

[0017] The model loss is calculated using a cross entropy loss function based on the recommendation probability of each product for the privacy-preserving user and the customized interest tags of each product.

[0018] According to a privacy protection-based information recommendation method provided by the present invention, the benchmark recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model, specifically comprising:

[0019] Select some sample users from the current sample users as secondary training users;

[0020] The baseline recommendation model is trained based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain an initial confusion model;

[0021] The initial confusion model is trained based on the sample user data of the secondary training users to obtain the information recommendation model.

[0022] According to a privacy protection-based information recommendation method provided by the present invention, the baseline recommendation model is trained based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain an initial confusion model, specifically comprising:

[0023] For any product, the sample user data of the secondary training user or the noise user data corresponding to the privacy protection user and the product information of any product are input into the benchmark recommendation model to obtain the recommendation probability of any product for the secondary training user or the privacy protection user output by the benchmark recommendation model;

[0024] Calculate the model loss of the baseline recommendation model based on the recommendation probability of each product for the secondary training user or the privacy protection user, and the real interest tag of each product for the secondary training user or the customized interest tag of each product for the privacy protection user;

[0025] The model parameters of the baseline recommendation model are adjusted based on the model loss of the baseline recommendation model to obtain the initial confusion model.

[0026] According to a privacy protection-based information recommendation method provided by the present invention, the initial confusion model is trained based on the sample user data of the secondary training user to obtain the information recommendation model, specifically comprising:

[0027] For any product, the sample user data of the secondary training user and the product information of any product are input into the initial confusion model to obtain the recommendation probability of any product for the secondary training user output by the initial confusion model;

[0028] Calculating the model loss of the initial confusion model based on the recommendation probability of each product for the secondary training user and the real interest tag of each product for the secondary training user;

[0029] The model parameters of the initial confusion model are adjusted based on the model loss of the initial confusion model to obtain the information recommendation model.

[0030] According to an information recommendation method based on privacy protection provided by the present invention, the user data of any user includes the user attribute data of the any user and the historical behavior data within a preset time period.

[0031] The present invention also provides an information recommendation device based on privacy protection, comprising:

[0032] A user data acquisition unit, used to acquire user data of each user in the system;

[0033] An information recommendation unit, configured to input the user data of each user and the product information of the product to be recommended into an information recommendation model, obtain the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommend product information based on the recommendation probability of the product to be recommended for each user;

[0034] The information recommendation model is trained based on the following steps:

[0035] Determine a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users;

[0036] After deleting the sample user data of the privacy-preserving user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy-preserving user based on the benchmark recommendation model;

[0037] The benchmark recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0038] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the information recommendation method based on privacy protection as described above is implemented.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the information recommendation method based on privacy protection as described in any one of the above is implemented.

[0040] The present invention also provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, the information recommendation method based on privacy protection as described above is implemented.

[0041] The privacy protection-based information recommendation method and device provided by the present invention determine a baseline recommendation model obtained by training based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users. After deleting the sample user data of the privacy protection user, the model parameters of the baseline recommendation model are fixed, and noise user data corresponding to the privacy protection user is generated based on the baseline recommendation model, so that the baseline recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to construct an information recommendation model. The knowledge learned by the baseline recommendation model based on the sample user data of the privacy protection user in its previous training process can be obfuscated, and the influence of the sample user data of the privacy protection user on the baseline recommendation model can be quickly and targetedly erased. At the same time, the information recommendation capability of the baseline recommendation model for other users other than the privacy protection user is maintained, and then product information recommendation is performed on the user data of each non-privacy protection user and the product information of the product to be recommended based on the information recommendation model, thereby realizing accurate information recommendation in the privacy protection mode. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 It is a flowchart of the information recommendation method based on privacy protection provided by the present invention;

[0044] Figure 2 It is a flowchart of the information recommendation model training method provided by the present invention;

[0045] Figure 3 It is a flow chart of the noise user data generating method provided by the present invention;

[0046] Figure 4 is a structural schematic diagram of an information recommendation device based on privacy protection provided by the present invention;

[0047] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Figure 1 is a flow chart of the information recommendation method based on privacy protection provided by the present invention, such as Figure 1 As shown, the method includes:

[0050] Step 110, obtaining user data of each user in the system;

[0051] Step 120, inputting the user data of each user and the product information of the product to be recommended into the information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and performing product information recommendation based on the recommendation probability of the product to be recommended for each user;

[0052] Among them, Figure 2 As shown, the information recommendation model is trained based on the following steps:

[0053] Step 210, determining a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who has made a privacy protection request among the sample users;

[0054] Step 220, after deleting the sample user data of the privacy protection user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy protection user based on the benchmark recommendation model;

[0055] Step 230: training the benchmark recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0056] Specifically, before officially using the information recommendation model to recommend products to each system registered user, the information recommendation model can be trained in advance to improve the information recommendation model's ability to understand user information and product information, so as to recommend products that users are more interested in. Among them, a lot of user privacy information is needed in the training process of the information recommendation model, so that the learned knowledge about user interests is internalized into the model parameters, so there is an insecurity of privacy leakage for users who pay attention to privacy. In order to protect the privacy of users while taking into account the performance of the recommendation system, a user data "forget" function can be provided for each user, so that the user can delete the knowledge about his user data in the information recommendation model at the same time after deleting his user data, so as to fully protect his privacy. Among them, in the initial stage, the user data of all system registered users can be saved in the server, and a benchmark recommendation model can be trained on the server side using the sample user data of all sample users. In some embodiments, the user data / sample user data of any user / sample user includes the user attribute data of the corresponding user and the historical behavior data within a preset time period, wherein the historical behavior data includes the preset number of product information visited (e.g., purchased, collected or browsed) within the preset time period. In other embodiments, the baseline recommendation model may include an embedding layer, a feature extraction layer, and a classification layer. The embedding layer is used to extract word embedding vectors from the input sample user's user data and the product information of any product, which can be constructed based on the word2vec model; the feature extraction layer is used to perform semantic extraction on the word embedding vectors of the sample user's user data and the word embedding vectors of the product information of the product, respectively, to obtain the user feature vector of the sample user and the product feature vector of the product, which can be constructed based on the Transformer model; the classification layer can classify based on the user feature vector of the sample user and the product feature vector of the product to determine the recommendation probability of the product for the sample user, which can be constructed based on a multi-layer perceptron.

[0057] When some sample users make a privacy protection request, they can be marked as privacy protection users, and the sample user data of the privacy protection users can be immediately deleted from the server. Subsequently, it is necessary to erase the influence of the sample user data of the privacy protection user on the baseline recommendation model, so that even if the baseline recommendation model encounters the sample user data of the privacy protection user again, it is difficult to make accurate information recommendations, thereby ensuring that the baseline recommendation model completely "forgets" the sample user data of the privacy protection user. It should be noted that, considering that the privacy protection user deletes its sample user data immediately after making a privacy protection request, it is necessary to erase the influence of the sample user data of the privacy protection user on the baseline recommendation model without knowing the sample user data of the privacy protection user. At the same time, considering the application scenario requirements of the information recommendation model, it is necessary to improve the efficiency of the above process, so as to efficiently obtain the information recommendation model that erases the influence of the sample user data of the privacy protection user.

[0058] Specifically, after deleting the sample user data of the privacy-preserving user, the model parameters of the baseline recommendation model can be fixed, that is, the baseline recommendation model is kept unchanged, and the noise user data corresponding to the privacy-preserving user is generated based on the baseline recommendation model. Subsequently, the baseline recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy-preserving user to obtain the information recommendation model. Here, the sample user data of some sample users can be used in combination with the noise user data corresponding to the privacy-preserving user to train the baseline recommendation model, thereby improving the efficiency of this process. It should be noted that if there are multiple privacy-preserving users, the noise user data corresponding to each privacy-preserving user is generated respectively. Among them, the noise user data corresponding to the privacy-preserving user can interfere with the information recommendation accuracy of the baseline recommendation model for the privacy-preserving user in the above training process, and confuse the knowledge learned by the baseline recommendation model based on the sample user data of the privacy-preserving user in its previous training process, while the sample user data of the sample user is used to maintain the information recommendation capability of the baseline recommendation model for other users other than the privacy-preserving user, so as to take into account the information recommendation accuracy of the trained information recommendation model.

[0059] In some embodiments, Figure 3 As shown in the figure, the baseline recommendation model can be used to generate noisy user data corresponding to privacy-preserving users based on the following steps:

[0060] Step 310, randomly adding customized interest tags to each product; wherein, if the customized interest tag of any product is 1, it means that the privacy protection user is interested in any product, and if the customized interest tag of any product is 0, it means that the privacy protection user is not interested in any product;

[0061] Step 320: input the current noise data corresponding to the privacy protection user and the product information of each product into the benchmark recommendation model, obtain the recommendation probability of each product output by the benchmark recommendation model for the privacy protection user, determine the model loss based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product, and adjust the current noise data corresponding to the privacy protection user based on the model loss until the model loss is less than a preset value;

[0062] Step 330: determine that the current noise data corresponding to the privacy protection user is the noise user data corresponding to the privacy protection user.

[0063] Here, for any privacy protection user, a customized interest tag corresponding to the privacy protection user can be randomly added to each product, that is, when there are multiple privacy protection users, the customized interest tags of the same product for different privacy protection users can be different. Among them, if the customized interest tag of any product for the privacy protection user is 1, it means that the privacy protection user is interested in the product, and if the customized interest tag of any product for the privacy protection user is 0, it means that the privacy protection user is not interested in the product.

[0064] For any privacy-preserving user, the noise user data of the privacy-preserving user can be generated in a manner similar to model training, which can be referred to as noise data training later. Specifically, in the initial stage, the initial noise data corresponding to the privacy-preserving user can be generated in a randomized manner as the current noise data in the first round of noise data training, wherein the dimension of the initial noise data is the same as the dimension of the sample user data of the sample user. In each round of noise data training, the current noise data corresponding to the privacy-preserving user and the product information of any product are input into the benchmark recommendation model to obtain the recommendation probability of the product for the privacy-preserving user output by the benchmark recommendation model. After obtaining the recommendation probability of each product for the privacy-preserving user output by the benchmark recommendation model, the model loss of the current round of noise data training process can be determined based on the recommendation probability of each product for the privacy-preserving user and the customized interest tags of each product for the privacy-preserving user, thereby adjusting the current noise data corresponding to the privacy-preserving user based on the model loss. Repeat the above training process until the model loss of a certain round of noise data training process is less than a preset value, and the current noise data corresponding to the privacy-preserving user in this round of noise data training process is used as the noise user data corresponding to the privacy-preserving user.

[0065] In some embodiments, the model loss of the current round of noise data training process can be calculated using a cross entropy loss function based on the recommendation probability of each product for the privacy-preserving user and the customized interest tags of each product for the privacy-preserving user.

[0066] In other embodiments, if the model loss of the current round of noise data training process is greater than a preset value, the current noise data corresponding to the privacy protection user can be randomly adjusted. For example, part of the current noise data corresponding to the privacy protection user can be randomly selected for randomization processing, wherein the number of data to be randomly processed in the current noise data can be determined according to the set learning rate, and the higher the learning rate, the more data to be randomly processed.

[0067] After obtaining the noise user data corresponding to each privacy-preserving user, the baseline recommendation model can be trained based on the sample user data of each current sample user and the noise user data corresponding to each privacy-preserving user, thereby obtaining an information recommendation model. In some embodiments, in order to quickly and specifically completely delete the sample user data of the privacy-preserving user that has been learned without affecting the overall model performance, thereby protecting the privacy of the user while taking into account the performance of the recommendation system, some sample users can be first selected from the current sample users as secondary training users, and then the baseline recommendation model can be trained based on the sample user data of the secondary training users and the noise user data corresponding to each privacy-preserving user, confusing the knowledge learned by the baseline recommendation model based on the sample user data of the privacy-preserving user during its previous training process, while maintaining the information recommendation capability of the baseline recommendation model for users other than the privacy-preserving user to a certain extent, thereby obtaining an initial confusion model. In order to further improve the information recommendation capability of the initial confusion model for users other than the privacy-preserving user, the initial confusion model can be trained based on the sample user data of the secondary training users alone to obtain an information recommendation model.

[0068] In some embodiments, in the process of training the benchmark recommendation model based on the sample user data of the secondary training users and the noise user data corresponding to each privacy protection user, for each round of training, for any product, the sample user data of the secondary training users or the noise user data corresponding to the privacy protection user and the product information of the product can be input into the benchmark recommendation model together, and the recommendation probability of the product for the corresponding secondary training user or the corresponding privacy protection user output by the benchmark recommendation model is obtained. Based on the recommendation probability of each product for each secondary training user or each privacy protection user, and the real interest tags of each product for the corresponding secondary training user or the customized interest tags of each product for the corresponding privacy protection user, the model loss of the benchmark recommendation model in the current round of training is calculated, so as to adjust the model parameters of the benchmark recommendation model based on the model loss of the benchmark recommendation model in the current round of training. Among them, the model loss can be calculated based on the cross entropy loss function. Repeat the above training process until the model loss of the benchmark recommendation model in a certain round of training is less than the first preset loss value, and then the current benchmark recommendation model can be used as the initial confusion model.

[0069] In other embodiments, in the process of training the initial confusion model based on the sample user data of the above-mentioned secondary training users, for each round of training, for any product, the sample user data of the secondary training users and the product information of the product can be input into the initial confusion model together to obtain the recommendation probability of the product for the corresponding secondary training user output by the initial confusion model. Based on the recommendation probability of each product for each secondary training user, and the real interest tags of each product for the corresponding secondary training user, the model loss of the initial confusion model in the current round of training is calculated, and the model parameters of the initial confusion model are adjusted based on the model loss of the initial confusion model in the current round of training. Among them, the model loss can also be calculated based on the cross entropy loss function. Repeat the above training process until the model loss of the initial hybrid model in a certain round of training is less than the second preset loss value, and then the current initial confusion model can be used as an information recommendation model.

[0070] After obtaining the information recommendation model, the user data of each user in the system (excluding privacy protection users) can be obtained, and the user data of each user and the product information of the product to be recommended can be input into the information recommendation model to obtain the recommendation probability of the product to be recommended for each user output by the information recommendation model. Based on the recommendation probability of the product to be recommended for each user, product information recommendation can be performed to achieve accurate information recommendation under the privacy protection mode.

[0071] In summary, the method provided by the embodiment of the present invention determines a baseline recommendation model obtained by training based on the sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users. After deleting the sample user data of the privacy protection user, the model parameters of the baseline recommendation model are fixed, and noise user data corresponding to the privacy protection user is generated based on the baseline recommendation model, so that the baseline recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to construct an information recommendation model. The knowledge learned by the baseline recommendation model based on the sample user data of the privacy protection user in its previous training process can be confused, and the influence of the sample user data of the privacy protection user on the baseline recommendation model can be quickly and targetedly erased. At the same time, the information recommendation capability of the baseline recommendation model for other users other than the privacy protection user is maintained, and then product information recommendation is performed on the user data of each non-privacy protection user and the product information of the product to be recommended based on the information recommendation model, thereby realizing accurate information recommendation in the privacy protection mode.

[0072] The following is a description of an information recommendation device based on privacy protection provided by the present invention. The information recommendation device based on privacy protection described below and the information recommendation method based on privacy protection described above can be referenced to each other.

[0073] Based on any of the above embodiments, Figure 4 is a schematic diagram of the structure of the information recommendation device based on privacy protection provided by the present invention. Figure 4 As shown, the device comprises:

[0074] A user data acquisition unit 410 is used to acquire user data of each user in the system;

[0075] The information recommendation unit 420 is used to input the user data of each user and the product information of the product to be recommended into the information recommendation model, obtain the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommend product information based on the recommendation probability of the product to be recommended for each user;

[0076] The information recommendation model is trained based on the following steps:

[0077] Determine a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users;

[0078] After deleting the sample user data of the privacy-preserving user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy-preserving user based on the benchmark recommendation model;

[0079] The benchmark recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0080] The device provided by the embodiment of the present invention determines a baseline recommendation model obtained by training based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users. After deleting the sample user data of the privacy protection user, the model parameters of the baseline recommendation model are fixed, and noise user data corresponding to the privacy protection user is generated based on the baseline recommendation model, so that the baseline recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to construct an information recommendation model. The knowledge learned by the baseline recommendation model based on the sample user data of the privacy protection user in its previous training process can be obfuscated, and the influence of the sample user data of the privacy protection user on the baseline recommendation model can be quickly and targetedly erased. At the same time, the information recommendation capability of the baseline recommendation model for other users other than the privacy protection user is maintained, and then product information recommendation is performed on the user data of each non-privacy protection user and the product information of the product to be recommended based on the information recommendation model, thereby realizing accurate information recommendation in the privacy protection mode.

[0081] Based on any of the foregoing embodiments, generating the noisy user data corresponding to the privacy-preserving user based on the benchmark recommendation model specifically includes:

[0082] Randomly add customized interest tags to each product; wherein, if the customized interest tag of any product is 1, it means that the privacy protection user is interested in any product, and if the customized interest tag of any product is 0, it means that the privacy protection user is not interested in any product;

[0083] Input the current noise data corresponding to the privacy protection user and the product information of each product into the benchmark recommendation model, obtain the recommendation probability of each product for the privacy protection user output by the benchmark recommendation model, determine the model loss based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product, and adjust the current noise data corresponding to the privacy protection user based on the model loss until the model loss is less than a preset value;

[0084] The current noise data corresponding to the privacy protection user is determined as the noise user data corresponding to the privacy protection user.

[0085] Based on any of the above embodiments, determining the model loss based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product specifically includes:

[0086] The model loss is calculated using a cross entropy loss function based on the recommendation probability of each product for the privacy-preserving user and the customized interest tags of each product.

[0087] Based on any of the above embodiments, the training of the benchmark recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model specifically includes:

[0088] Select some sample users from the current sample users as secondary training users;

[0089] The baseline recommendation model is trained based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain an initial confusion model;

[0090] The initial confusion model is trained based on the sample user data of the secondary training users to obtain the information recommendation model.

[0091] Based on any of the above embodiments, the training of the benchmark recommendation model based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain the initial confusion model specifically includes:

[0092] For any product, the sample user data of the secondary training user or the noise user data corresponding to the privacy protection user and the product information of any product are input into the benchmark recommendation model to obtain the recommendation probability of any product for the secondary training user or the privacy protection user output by the benchmark recommendation model;

[0093] Calculate the model loss of the baseline recommendation model based on the recommendation probability of each product for the secondary training user or the privacy protection user, and the real interest tag of each product for the secondary training user or the customized interest tag of each product for the privacy protection user;

[0094] The model parameters of the baseline recommendation model are adjusted based on the model loss of the baseline recommendation model to obtain the initial confusion model.

[0095] Based on any of the above embodiments, the training of the initial confusion model based on the sample user data of the secondary training user to obtain the information recommendation model specifically includes:

[0096] For any product, the sample user data of the secondary training user and the product information of any product are input into the initial confusion model to obtain the recommendation probability of any product for the secondary training user output by the initial confusion model;

[0097] Calculating the model loss of the initial confusion model based on the recommendation probability of each product for the secondary training user and the real interest tag of each product for the secondary training user;

[0098] The model parameters of the initial confusion model are adjusted based on the model loss of the initial confusion model to obtain the information recommendation model.

[0099] Based on any of the above embodiments, the user data of any user includes user attribute data of the any user and historical behavior data within a preset time period.

[0100] Figure 5 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor (processor) 510, a memory (memory) 520, a communication interface (Communications Interface) 530 and a communication bus 540, wherein the processor 510, the memory 520, and the communication interface 530 communicate with each other through the communication bus 540. The processor 510 can call the logic instructions in the memory 520 to execute the information recommendation method based on privacy protection, which includes: obtaining user data of each user in the system; inputting the user data of each user and the product information of the product to be recommended into the information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommending product information based on the recommendation probability of the product to be recommended for each user; wherein the information recommendation model is trained based on the following steps: determining a baseline recommendation model trained based on the sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users; after deleting the sample user data of the privacy protection user, fixing the model parameters of the baseline recommendation model, and generating noise user data corresponding to the privacy protection user based on the baseline recommendation model; training the baseline recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0101] In addition, the logic instructions in the above-mentioned memory 520 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0102] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the information recommendation method based on privacy protection provided by the above methods, the method comprising: obtaining user data of each user in the system; inputting the user data of each user and the product information of the product to be recommended into an information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommending product information based on the recommendation probability of the product to be recommended for each user; wherein the information recommendation model is trained based on the following steps: determining a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users; after deleting the sample user data of the privacy protection user, fixing the model parameters of the baseline recommendation model, and generating noise user data corresponding to the privacy protection user based on the baseline recommendation model; training the baseline recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model.

[0103] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the above-mentioned privacy-protection-based information recommendation method, the method comprising: obtaining user data of each user in the system; inputting the user data of each user and the product information of the product to be recommended into an information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommending product information based on the recommendation probability of the product to be recommended for each user; wherein the information recommendation model is trained based on the following steps: determining a baseline recommendation model trained based on sample user data of all sample users and a privacy-protection user who makes a privacy protection request among the sample users; after deleting the sample user data of the privacy-protection user, fixing the model parameters of the baseline recommendation model, and generating noise user data corresponding to the privacy-protection user based on the baseline recommendation model; training the baseline recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy-protection user to obtain the information recommendation model.

[0104] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0105] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A privacy protection-based information recommendation method, characterized in that: include: Get user data of each user in the system; Inputting the user data of each user and the product information of the product to be recommended into the information recommendation model, obtaining the recommendation probability of the product to be recommended for each user output by the information recommendation model, and performing product information recommendation based on the recommendation probability of the product to be recommended for each user; The information recommendation model is trained based on the following steps: Determine a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users; After deleting the sample user data of the privacy-preserving user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy-preserving user based on the benchmark recommendation model; The reference recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model; The generating the noisy user data corresponding to the privacy-preserving user based on the benchmark recommendation model specifically includes: Randomly add customized interest tags to each product; wherein, if the customized interest tag of any product is 1, it means that the privacy protection user is interested in any product, and if the customized interest tag of any product is 0, it means that the privacy protection user is not interested in any product; The current noise data corresponding to the privacy protection user and the product information of each product are input into the benchmark recommendation model, and after obtaining the recommendation probability of each product output by the benchmark recommendation model for the privacy protection user, the model loss is determined based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product, and the current noise data corresponding to the privacy protection user is adjusted based on the model loss until the model loss is less than a preset value; wherein, the initial noise data corresponding to the privacy protection user is generated in a randomized manner; The current noise data corresponding to the privacy protection user is determined as the noise user data corresponding to the privacy protection user.

2. The information recommendation method based on privacy protection according to claim 1, characterized in that: The determining of the model loss based on the recommendation probability of each product for the privacy-preserving user and the customized interest tag of each product specifically includes: The model loss is calculated using a cross entropy loss function based on the recommendation probability of each product for the privacy-preserving user and the customized interest tags of each product.

3. The information recommendation method based on privacy protection according to claim 1, characterized in that: The step of training the benchmark recommendation model based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model specifically includes: Select some sample users from the current sample users as secondary training users; The baseline recommendation model is trained based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain an initial confusion model; The initial confusion model is trained based on the sample user data of the secondary training users to obtain the information recommendation model.

4. The information recommendation method based on privacy protection according to claim 3 is characterized in that: The training of the benchmark recommendation model based on the sample user data of the secondary training user and the noise user data corresponding to the privacy protection user to obtain an initial confusion model specifically includes: For any product, the sample user data of the secondary training user or the noise user data corresponding to the privacy protection user and the product information of the any product are input into the benchmark recommendation model to obtain the recommendation probability of the any product for the secondary training user or the privacy protection user output by the benchmark recommendation model; Calculate the model loss of the baseline recommendation model based on the recommendation probability of each product for the secondary training user or the privacy protection user, and the real interest tag of each product for the secondary training user or the customized interest tag of each product for the privacy protection user; The model parameters of the baseline recommendation model are adjusted based on the model loss of the baseline recommendation model to obtain the initial confusion model.

5. The information recommendation method based on privacy protection according to claim 3 is characterized in that: The training of the initial confusion model based on the sample user data of the secondary training user to obtain the information recommendation model specifically includes: For any product, the sample user data of the secondary training user and the product information of any product are input into the initial confusion model to obtain the recommendation probability of any product for the secondary training user output by the initial confusion model; Calculating the model loss of the initial confusion model based on the recommendation probability of each product for the secondary training user and the real interest tag of each product for the secondary training user; The model parameters of the initial confusion model are adjusted based on the model loss of the initial confusion model to obtain the information recommendation model.

6. The information recommendation method based on privacy protection according to claim 1, characterized in that: The user data of any user includes the user attribute data of the any user and the historical behavior data within a preset time period.

7. A privacy protection-based information recommendation method, characterized in that: include: A user data acquisition unit, used to acquire user data of each user in the system; An information recommendation unit, configured to input the user data of each user and the product information of the product to be recommended into an information recommendation model, obtain the recommendation probability of the product to be recommended for each user output by the information recommendation model, and recommend product information based on the recommendation probability of the product to be recommended for each user; The information recommendation model is trained based on the following steps: Determine a baseline recommendation model trained based on sample user data of all sample users and a privacy protection user who makes a privacy protection request among the sample users; After deleting the sample user data of the privacy-preserving user, fixing the model parameters of the benchmark recommendation model, and generating noise user data corresponding to the privacy-preserving user based on the benchmark recommendation model; The reference recommendation model is trained based on the sample user data of each current sample user and the noise user data corresponding to the privacy protection user to obtain the information recommendation model; The generating the noisy user data corresponding to the privacy-preserving user based on the benchmark recommendation model specifically includes: Randomly add customized interest tags to each product; wherein, if the customized interest tag of any product is 1, it means that the privacy protection user is interested in any product, and if the customized interest tag of any product is 0, it means that the privacy protection user is not interested in any product; The current noise data corresponding to the privacy protection user and the product information of each product are input into the benchmark recommendation model, and after obtaining the recommendation probability of each product output by the benchmark recommendation model for the privacy protection user, the model loss is determined based on the recommendation probability of each product for the privacy protection user and the customized interest tag of each product, and the current noise data corresponding to the privacy protection user is adjusted based on the model loss until the model loss is less than a preset value; wherein, the initial noise data corresponding to the privacy protection user is generated in a randomized manner; The current noise data corresponding to the privacy protection user is determined as the noise user data corresponding to the privacy protection user.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the information recommendation method based on privacy protection as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the information recommendation method based on privacy protection as claimed in any one of claims 1 to 6 is implemented.

Images