CN-121981807-A - Recommendation system forgetting method and device based on differential privacy contrast learning

CN121981807ACN 121981807 ACN121981807 ACN 121981807ACN-121981807-A

Abstract

The application relates to the field of machine learning, in particular to a recommendation system forgetting method and device based on differential privacy contrast learning. The method comprises the steps of obtaining an original interaction diagram, a forgetting user initiating a forgetting request and privacy sensitivity labels of the user according to interaction data between the user and commodities, conducting knowledge migration on a student model through a teacher model, predicting preference results of the user on the commodities through the teacher model and the student model respectively, obtaining loss values based on the preference results, optimizing the student model based on the loss values, obtaining a disturbance interaction diagram according to the privacy sensitivity labels and the original interaction diagram, conducting fine adjustment on model parameters of the student model based on the forgetting user and the disturbance interaction diagram, and obtaining the student model with the forgetting optimized function, wherein the student model is used for recommending commodities to the user in a recommendation system. The method is used for solving the problem that the traditional recommendation system is difficult to realize thorough and real forgetting of user data, and realizing both recommendation precision and real-time response efficiency of the recommendation system.

Inventors

ZHAO JIE
XIE ZHUOLIN
ZHANG CUIHONG
WU XIUZHU

Assignees

广东工业大学

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (11)

1. A recommendation system forgetting method based on differential privacy contrast learning is characterized by comprising the following steps: acquiring an original interaction diagram between a user and a commodity, a forgetting user initiating a forgetting request and a privacy sensitivity label of the user according to interaction data between the user and the commodity; performing knowledge migration on the student model by using the trained teacher model to obtain a student model with knowledge migration completed; Predicting the preference result of the user for the commodity by using the teacher model and the student model completing the knowledge migration respectively, acquiring a loss value based on the preference result, and optimizing the student model completing the knowledge migration based on the loss value to obtain an optimized student model; And fine tuning model parameters of the optimized student model based on the forgotten user and the disturbance interaction graph to obtain a student model with forgotten optimization, wherein the student model is used for recommending commodities to the user in a recommendation system.
2. The method of claim 1, wherein the teacher model is provided with a plurality of graph convolution layers, the method further comprising: And inputting the original interaction diagram between the user and the commodity into the teacher model, and training the teacher model to obtain a trained teacher model, wherein the user embedded vector and the commodity embedded vector of each diagram roll layer are obtained in the training process.
3. The method of claim 2, wherein the obtaining a user embedded vector of a current graph convolution layer in the teacher model is: The method comprises the steps of taking a user corresponding to a user embedded vector of a current graph convolution layer as a first user, taking commodities interacted with the first user as first commodities, obtaining commodity embedded vectors of a previous graph convolution layer corresponding to the first commodities as first commodity embedded vectors, carrying out weighted summation on all the first commodity embedded vectors to obtain the user embedded vectors of the first user on the current graph convolution layer, wherein the weight of the first commodity embedded vectors is calculated based on the number of the first commodities and the number of users interacted with the first commodities.
4. The method of claim 2, wherein the obtaining method of the commodity embedding vector of the current graph convolution layer in the teacher model is as follows: The commodity embedding method comprises the steps of taking commodities corresponding to commodity embedding vectors of a current graph convolution layer as second commodities, taking users interacted with the second commodities as second users, obtaining user embedding vectors of a previous graph convolution layer corresponding to the second users as second user embedding vectors, carrying out weighted summation on all the second user embedding vectors to obtain commodity embedding vectors of the second commodities on the current graph convolution layer, wherein the weight of the second user embedding vectors is calculated based on the number of the second users and the number of the commodities interacted with the second users.
5. The method of claim 1, wherein prior to the step of knowledge migration to the student model using the trained teacher model, the method further comprises: predicting a first preference score of the user for the goods with the interaction and a second preference score of the user for the goods without the interaction based on the trained teacher model; Inputting the first preference scores, the second preference scores and model parameters in the trained teacher model into a preset first loss function, and optimizing the trained teacher model to obtain an optimized teacher model; The training teacher model is utilized to carry out knowledge migration on the student model, and the training teacher model specifically comprises the following steps: and carrying out knowledge migration on the student model by using the optimized teacher model.
6. The method according to claim 2, wherein the performing knowledge migration on the student model by using the trained teacher model to obtain a student model with completed knowledge migration comprises: based on a trained teacher model, determining a user corresponding to a user embedded vector to be migrated as a third user, determining a commodity corresponding to a commodity embedded vector to be migrated as a fourth commodity, and acquiring other users interacted with the fourth commodity as associated users; Based on privacy sensitivity labels of the third user and the associated user, respectively determining differential privacy budget parameters corresponding to the third user and the associated user; Injecting disturbance noise into the user embedded vector of the third user according to the differential privacy budget parameters corresponding to the third user to obtain a user fuzzy embedded vector; injecting disturbance noise into the commodity embedding vector of the fourth commodity according to the differential privacy budget parameters corresponding to the associated user to obtain a commodity fuzzy embedding vector; and guiding the student model to train by using the user fuzzy embedded vector and the commodity fuzzy embedded vector as supervision signals to obtain the student model for completing knowledge migration.
7. The method of claim 6, wherein the injecting disturbance noise into the user-embedded vector of the third user according to the differential privacy budget parameter corresponding to the third user to obtain the user-blurred embedded vector comprises: Generating first disturbance noise based on a preset super parameter of Laplace distribution, a random variable of Laplace distribution and a differential privacy budget parameter of the third user, and injecting the first disturbance noise into a user embedded vector of the third user to obtain a user fuzzy embedded vector; Injecting disturbance noise into the commodity embedding vector of the fourth commodity according to the differential privacy budget parameters corresponding to the associated user to obtain a commodity fuzzy embedding vector, comprising: Generating second disturbance noise based on the super parameter, the random variable and the differential privacy budget parameter of the associated user, and injecting the second disturbance noise into the commodity embedded vector of the fourth commodity to obtain a commodity fuzzy embedded vector.
8. The method according to claim 1, wherein the disturbance interaction map is specifically: acquiring differential privacy budget parameters corresponding to the user according to the privacy sensitivity label of the user; acquiring corresponding bit flip probability according to the differential privacy budget parameters of the user; acquiring the total number of commodities interacted by the user in an original interaction diagram; taking the product of the bit flip probability and the total number of commodities as a first numerical value, and taking the number of commodities which are not interacted by the user in the original interaction diagram as a second numerical value; if the user interacts with the commodity in the original interaction diagram, the complementary probability of the bit flip probability corresponding to the user accords with Bernoulli distribution in the disturbance interaction diagram; And if the user and the commodity are not interacted in the original interaction diagram, the ratio of the first numerical value to the second numerical value accords with Bernoulli distribution.
9. The method of claim 1, wherein predicting the preference result of the user for the commodity by using the teacher model and the student model completing the knowledge migration, respectively, obtaining a loss value based on the preference result, optimizing the student model completing the knowledge migration based on the loss value, and obtaining an optimized student model, comprises: predicting a third preference score of the user for the goods which have interacted based on the teacher model, and predicting a fourth preference score of the user for the goods which have not interacted; Predicting a fifth preference score of the user for the goods with the interaction based on the student model with the knowledge migration completed, and predicting a sixth preference score of the user for the goods without the interaction; Inputting the third preference score and the fifth preference score into a preset mean square error function to obtain a first loss value; inputting the fourth preference score and the sixth preference score into the mean square error function to obtain a second loss value; and optimizing the student model for completing knowledge migration based on the first loss value and the second loss value to obtain an optimized student model.
10. The method of claim 1, wherein the fine tuning of model parameters of the optimized student model based on the forgotten user, the disturbance interaction map comprises: acquiring interaction data of the non-forgetting user and the commodity according to the forgetting user and the disturbance interaction graph; Inputting the interaction data of the non-forgetting user and the model parameters of the optimized student model into a preset second loss function to calculate to obtain a first intermediate value; Predicting a seventh preference score of the non-forgetting user on the commodity with the interaction based on the optimized student model, and predicting an eighth preference score of the non-forgetting user on the commodity with the interaction; acquiring a second intermediate value based on the seventh preference score, the eighth preference score and a preset contrast fine tuning loss function; And fine tuning the model parameters of the optimized student model based on the first intermediate value and the second intermediate value to obtain the student model with forgetting optimization.
11. A recommendation system forgetting device based on differential privacy contrast learning, the device comprising: the acquisition module is used for acquiring an original interaction diagram between a user and a commodity, a forgetting user initiating a forgetting request and a privacy sensitivity label of the user according to interaction data between the user and the commodity; The knowledge migration module is used for carrying out knowledge migration on the student model by utilizing the trained teacher model to obtain a student model with knowledge migration completed; The student model optimization module is used for predicting preference results of the user on commodities by using the teacher model and the student models completing knowledge migration respectively, acquiring loss values based on the preference results, and optimizing the student models completing knowledge migration based on the loss values to obtain optimized student models; the forgetting-recommending module is used for carrying out random topological disturbance on the original interaction diagram based on the privacy sensitivity label to obtain a disturbance interaction diagram, carrying out fine adjustment on model parameters of the optimized student model based on the forgetting user and the disturbance interaction diagram to obtain a student model with the forgetting optimization, and carrying out commodity recommendation on the user in a recommending system.

Description

Recommendation system forgetting method and device based on differential privacy contrast learning Technical Field The invention relates to the field of machine learning, in particular to a recommendation system forgetting method and device based on differential privacy contrast learning. Background With the rapid development of deep penetration and digital economy of internet technology, personalized recommendation systems have become core infrastructures of internet services such as e-commerce platforms, social media and content distribution. The recommendation system accurately constructs user interest portraits by mining massive user historical behavior data, such as interactive records of clicking, browsing, purchasing, collecting and the like, so that the paradigm transition from the user content searching to the content user searching is realized, and the user experience and business conversion efficiency are remarkably improved. However, in practical use of the recommendation system, the recommendation system performs personalized recommendation on the user based on a machine learning model, the machine learning model converts original behavior characteristics of the user into complex coupling relations in a high-dimensional parameter space in a training process, and even if the recommendation system physically deletes original records of the user at a database level according to user requirements, the sensitive information still remains in shadow data form in weight of machine learning. In addition, the traditional recommendation system generally adopts a graph neural network to mine high-order cooperative characteristics, and privacy information of a specific user can permeate to neighbor nodes and even global topology through a message transmission mechanism of a graph structure, so that a chain type privacy diffusion effect which is difficult to cut off is formed. The method has the advantages that although the global retraining mode in the prior art can realize theoretical complete forgetting, extremely large calculation cost and time delay are needed to bear in a very large-scale recommendation graph scene, frequent and random real-time forgetting requests cannot be met, other traditional approximate forgetting learning methods are deep in privacy isolation and performance maintenance zero and game, simple parameter rollback or local fine tuning often causes severe concussion of preference characterization of non-forgetting user groups, recommendation precision cliff breaking drop is caused, and a randomization mechanism based on traditional differential privacy can provide privacy guarantee in a mathematical level, but often submerges real user interest signals due to excessive noise, so that the interpretability and recommendation satisfaction of the model are damaged. Therefore, the personalized forgetting method in the recommendation system still becomes a key technical bottleneck to be broken through in the current personalized recommendation. Disclosure of Invention The invention provides a recommendation system forgetting method and device based on differential privacy contrast learning, which are used for solving the problem that the traditional recommendation system is difficult to realize thorough and real forgetting of user data and realizing the consideration of the recommendation precision and the real-time response efficiency of the recommendation system. According to a first aspect of the present application, there is provided a recommendation system forgetting method based on differential privacy contrast learning, the method comprising: acquiring an original interaction diagram between a user and a commodity, a forgetting user initiating a forgetting request and a privacy sensitivity label of the user according to interaction data between the user and the commodity; performing knowledge migration on the student model by using the trained teacher model to obtain a student model with knowledge migration completed; Predicting the preference result of the user for the commodity by using the teacher model and the student model completing the knowledge migration respectively, acquiring a loss value based on the preference result, and optimizing the student model completing the knowledge migration based on the loss value to obtain an optimized student model; And fine tuning model parameters of the optimized student model based on the forgotten user and the disturbance interaction graph to obtain a student model with forgotten optimization, wherein the student model is used for recommending commodities to the user in a recommendation system. It can be understood that by introducing the user privacy sensitivity label, random topology disturbance is performed on the image structure level, so that the user privacy is protected, and the excessive loss of recommendation precision caused by traditional one-cut noise is avoided. The method effectively solves the technical contradiction that