Recommendation system forgetting method and device based on differential privacy contrast learning

Abstract

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. The method comprises the following steps: according to interaction data between a user and a commodity, obtaining an original interaction graph, a forgotten user initiating a forgetting request and a privacy sensitivity label of the user; performing knowledge migration on the student model by using the teacher model; respectively utilizing the teacher model and the student model to predict a preference result of the user for the commodity, acquiring a loss value based on the preference result, and optimizing the student model based on the loss value; obtaining a disturbance interaction graph according to the privacy sensitivity label and the original interaction graph; and based on the forgotten user and the disturbance interaction diagram, performing fine adjustment on model parameters of the student model to obtain the student model after forgetting optimization, and performing commodity recommendation on the user in a recommendation system. The method is used for solving the problem that an existing recommendation system is difficult to achieve thorough and real forgetting of user data, and the recommendation precision and the real-time response efficiency of the recommendation system are both considered.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically, to a forgetting method and apparatus for recommendation systems based on differential privacy contrastive learning. Background Technology

[0002] With the deep penetration of internet technology and the booming development of the digital economy, personalized recommendation systems have become a core infrastructure for internet services such as e-commerce platforms, social media, and content distribution. By mining massive amounts of historical user behavior data, such as clicks, browsing, purchases, and favorites, recommendation systems accurately construct user interest profiles, thereby achieving a paradigm shift from users finding content to content finding users, significantly improving user experience and business conversion efficiency.

[0003] However, in practical applications of recommender systems, these systems rely on machine learning models to provide personalized recommendations. During training, these models transform users' original behavioral characteristics into complex coupling relationships within a high-dimensional parameter space. Even if the recommender system physically deletes the user's original records at the database level based on user needs, this sensitive information remains latent as shadow data within the machine learning weights. Furthermore, traditional recommender systems commonly employ graph neural networks to mine high-order collaborative features. This allows specific users' privacy information to permeate neighboring nodes and even the global topology through the message passing mechanism of the graph structure, creating a chain-like privacy diffusion effect that is difficult to break. This leads to significant shortcomings in existing forgetting techniques: while global retraining can theoretically achieve complete forgetting, it incurs enormous computational costs and time delays in scenarios with ultra-large-scale recommendation graphs, making it unable to handle frequent and random real-time forgetting requests; other traditional approximate forgetting learning methods are caught in a zero-sum game between privacy isolation and performance preservation, where simple parameter rollback or local fine-tuning often causes drastic fluctuations in the preference representations of non-forgotten users, resulting in a precipitous drop in recommendation accuracy; and while traditional differential privacy-based randomization mechanisms can provide mathematical privacy guarantees, they often overwhelm genuine user interest signals with excessive noise, thereby impairing model interpretability and recommendation satisfaction. Therefore, personalized forgetting methods in recommendation systems remain a key technical bottleneck that urgently needs to be overcome in current personalized recommendation systems. Summary of the Invention

[0004] This invention provides a forgetting method and apparatus for recommendation systems based on differential privacy contrastive learning, which solves the problem that existing recommendation systems are unable to achieve complete and true forgetting of user data, and achieves a balance between recommendation accuracy and real-time response efficiency.

[0005] According to a first aspect of this application, a forgetting method for a recommendation system based on differential privacy contrastive learning is provided, the method comprising: Based on the user-product interaction data, obtain the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the user's privacy sensitivity tag; The student model is used to transfer knowledge from the trained teacher model to obtain a student model that has completed the knowledge transfer. The teacher model and the student model that has completed knowledge transfer are used to predict the user's preference for the product, respectively. The loss value is obtained based on the preference result. The student model that has completed knowledge transfer is optimized based on the loss value to obtain the optimized student model. The original interaction graph is randomly perturbed based on the privacy sensitivity label to obtain a perturbed interaction graph; the model parameters of the optimized student model are fine-tuned based on the forgotten user and the perturbed interaction graph to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

[0006] Understandably, by introducing user privacy-sensitive tags and performing random topological perturbations at the graph structure level, user privacy is protected while avoiding excessive loss of recommendation accuracy caused by traditional one-size-fits-all noise addition. The comparative optimization of teacher and student models, along with a fine-tuning mechanism based on the perturbation interaction graph, not only completely erases the knowledge contributions of forgotten users but also ensures the immediate recovery and stable maintenance of recommendation performance for non-forgotten users. This method effectively resolves the technical contradiction of balancing privacy erasure and performance maintenance, significantly improving the response speed and recommendation accuracy of the recommendation system in complex privacy environments while strictly adhering to the user's right to be forgotten.

[0007] Optionally, the teacher model has several graph convolutional layers; the method further includes: The original interaction graph between the user and the product is input into the teacher model, and the teacher model is trained to obtain a trained teacher model; wherein, during the training process, the user embedding vector and product embedding vector of each graph convolutional layer are obtained.

[0008] Understandably, by constructing a teacher model with a multi-layered graph convolutional structure, user embedding vectors and item embedding vectors are acquired layer by layer during training. Each graph convolutional layer effectively integrates the feature information of neighboring nodes of the user and item embedding vectors through a weighted aggregation mechanism. This allows the teacher model to start from local interactions and gradually capture global collaborative filtering signals, significantly improving the semantic richness and discriminative power of the embedding representations. Simultaneously, by acquiring the intermediate embeddings from each graph convolutional layer, multi-layered and fine-grained feature inputs are provided for subsequent knowledge transfer, thereby enhancing the student model's learning effect on user preference patterns and ultimately improving the prediction accuracy and generalization performance of the recommendation system.

[0009] Optionally, the user embedding vector of the current graph convolutional layer in the teacher model is obtained as follows: The user corresponding to the user embedding vector of the current graph convolutional layer is taken as the first user; the product that has interacted with the first user is taken as the first product; the product embedding vector of the previous graph convolutional layer corresponding to the first product is obtained as the first product embedding vector; the first product embedding vector is obtained by weighted summation of all the first product embedding vectors; wherein the weight of the first product embedding vector is calculated based on the number of the first products and the number of users who have interacted with the first products.

[0010] Understandably, this invention, through the design of a hierarchical graph convolution aggregation mechanism, not only aggregates the feature information of the user's historical interaction with the product when generating the user embedding vector of the current graph convolution layer, but also introduces a dynamic weight calculation method, comprehensively considering the number of the first product and the number of users who have interacted with the first product. This avoids the excessive dominance of the high-order nodes on the embedding representation, enabling the user embedding vector to more accurately characterize the user's personalized preferences. This provides higher-quality feature input for subsequent privacy protection and knowledge transfer tasks, ultimately improving the overall performance and stability of the recommendation system.

[0011] Optionally, the product embedding vector of the current graph convolutional layer in the teacher model is obtained as follows: The product corresponding to the product embedding vector of the current graph convolutional layer is taken as the second product; the user who has interacted with the second product is taken as the second user; the user embedding vector of the previous graph convolutional layer corresponding to the second user is obtained as the second user embedding vector; the weighted sum of all the second user embedding vectors is used to obtain the product embedding vector of the second product in the current graph convolutional layer; wherein the weight of the second user embedding vector is calculated based on the number of second users and the number of products that have interacted with the second user.

[0012] Understandably, by designing a product-oriented inverse graph convolution aggregation mechanism, the features of all users who have interacted with the product are aggregated when generating the product embedding vector of the current layer. A dynamic weight calculation method is introduced, where the weights are jointly determined by the number of second users and the number of products that user has interacted with. This effectively balances the contribution of different users to the product representation and avoids the excessive dominance of a few users in the product embedding. This not only enables the product embedding vector to more comprehensively reflect the diverse preference characteristics of its audience, but also improves the robustness and discriminativeness of the product representation, laying a solid foundation for the recommendation system to achieve accurate matching.

[0013] Optionally, before the step of transferring knowledge from the trained teacher model to the student model, the method further includes: Based on the trained teacher model, predict the user's first preference rating for products that have been interacted with, and predict the user's second preference rating for products that have not been interacted with. The first preference score, the second preference score, and the model parameters in the trained teacher model are input into a preset first loss function to optimize the trained teacher model and obtain an optimized teacher model. The process of transferring knowledge from the trained teacher model to the student model specifically involves: The optimized teacher model is used to transfer knowledge from the student model.

[0014] Understandably, introducing an optimization step for the teacher model before knowledge transfer—by constructing a first loss function that incorporates different sample preference scores and model parameters—optimizes the trained teacher model. This not only strengthens the teacher model's ability to distinguish between explicit and implicit user preferences but also effectively suppresses overfitting by constraining the model parameters, making the teacher model's knowledge representation more concise and generalizable. The optimized teacher model, as the source of knowledge transfer, provides clearer and more stable supervision signals to the student model, significantly improving the efficiency of knowledge distillation and the learning effect of the student model. This lays a higher-quality model foundation for subsequent privacy protection and forgetting operations.

[0015] Optionally, the step of using a trained teacher model to transfer knowledge to a student model to obtain a student model that has completed the knowledge transfer includes: Based on the trained teacher model, the user corresponding to the current user embedding vector to be transferred is determined as the third user, the product corresponding to the current product embedding vector to be transferred is determined as the fourth product, and other users who have interacted with the fourth product are obtained as associated users. Based on the privacy sensitivity tags of the third user and the associated user, the differential privacy budget parameters corresponding to the third user and the associated user are determined respectively; Based on the differential privacy budget parameters corresponding to the third user, perturbation noise is injected into the user embedding vector of the third user to obtain a user fuzzy embedding vector; and based on the differential privacy budget parameters corresponding to the associated user, perturbation noise is injected into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector. The user fuzzy embedding vector and the product fuzzy embedding vector are used as supervision signals to guide the training of the student model, resulting in a student model that has completed knowledge transfer.

[0016] Understandably, differential privacy protection mechanisms are deeply integrated with personalized privacy needs during the knowledge transfer phase. By dynamically obtaining differentiated differential privacy budget parameters based on the privacy sensitivity tags of users and the users associated with their interactive products, and by injecting targeted noise into the embedding vector of the teacher model, precise adaptation of privacy protection strength is achieved. This not only ensures that the privacy information of highly sensitive users is strictly obfuscated during knowledge transfer, but also avoids excessive loss of effective information due to excessive noise for low-sensitivity users.

[0017] Optionally, the step of injecting perturbation noise into the user embedding vector of the third user based on the differential privacy budget parameter corresponding to the third user to obtain a user fuzzy embedding vector includes: The first perturbation noise is generated based on the preset hyperparameters of the Laplace distribution, the random variables of the Laplace distribution, and the differential privacy budget parameters of the third user. The first perturbation noise is then injected into the user embedding vector of the third user to obtain the user fuzzy embedding vector. The step of injecting perturbation noise into the product embedding vector of the fourth product based on the differential privacy budget parameter corresponding to the associated user to obtain the product fuzzy embedding vector includes: A second perturbation noise is generated based on the hyperparameters, the random variables, and the differential privacy budget parameters of the associated users, and the second perturbation noise is injected into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector.

[0018] Understandably, the Laplace perturbation mechanism based on personalized privacy budget parameters is adopted in the knowledge transfer stage. By configuring differentiated differential privacy budget parameters for third-party users and associated users respectively, targeted noise injection is achieved into the user and item embedding vectors. This ensures that the perturbation strength of the item embedding is determined by the privacy needs of its associated users, thereby blocking the path of individual privacy leakage through embedding representation from the underlying feature level. This maximizes the preservation of the preference signals of low-sensitivity users, so that the generated student model has both inherent privacy resistance and can maintain high recommendation accuracy, providing an ideal safe manifold initial state for subsequent forgetting operations.

[0019] Optionally, the perturbation interaction graph specifically refers to: Based on the user's privacy sensitivity tags, obtain the differential privacy budget parameters corresponding to the user; The corresponding bit-flip probability is obtained based on the user's differential privacy budget parameters; Obtain the total number of products that the user has interacted with in the original interaction graph; The product of the bit-flipping probability and the total number of products is the first value, and the number of products that the user did not interact with in the original interaction graph is the second value. If the user and the product have interacted in the original interaction graph, then in the perturbation interaction graph, the complementary probabilities of the bit flip probabilities corresponding to the user follow a Bernoulli distribution. If the user and the product have not interacted in the original interaction graph, then the ratio of the first value to the second value follows a Bernoulli distribution.

[0020] Understandably, by randomly removing real interaction edges in the perturbation interaction graph with complementary probabilities and introducing false interactions into non-interaction edges with sampling probabilities based on user activity, the statistical characteristics of user distribution are maintained macroscopically. This blocks the link between privacy information at the underlying physical structure level. It not only provides each user with a level of protection that matches their privacy needs, but also provides a natural strong negative sample for the subsequent comparison and fine-tuning stage, effectively improving the security and efficiency of the forgetting operation.

[0021] Optionally, the step of predicting user preferences for goods using the teacher model and the student model that has completed knowledge transfer, obtaining a loss value based on the preference results, and optimizing the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model includes: Based on the teacher model, predict the user's third preference rating for products that have been interacted with, and predict the user's fourth preference rating for products that have not been interacted with. Based on the student model that has completed knowledge transfer, predict the user's fifth preference rating for products that have been interacted with, and predict the user's sixth preference rating for products that have not been interacted with. The third preference score and the fifth preference score are input into a preset mean squared error function to obtain a first loss value; the fourth preference score and the sixth preference score are input into the mean squared error function to obtain a second loss value; The student model that has completed knowledge transfer is optimized based on the first loss value and the second loss value to obtain the optimized student model.

[0022] Understandably, a comparative optimization stage was introduced after knowledge transfer. By calculating the mean squared error of preference scores for teacher and student models that had and had not interacted, a two-way distillation loss mechanism was constructed. This mechanism not only constrains the student model to accurately fit the teacher's judgment of the user's explicit preferences, but also forces it to align with the teacher model's perception of the ranking of items that the user did not interact with. This effectively avoids the bias problem in the recommendation list caused by ignoring the distribution of items that did not interact in traditional distillation. While inheriting the knowledge of the teacher model, the student model enhances its ability to distinguish the boundaries of user preferences, significantly improving the accuracy and personalization of the recommendation results. This lays a robust model foundation for subsequent privacy forgetting operations.

[0023] Optionally, the fine-tuning of the model parameters of the optimized student model based on the forgotten user and the perturbation interaction graph includes: Based on the forgotten users and the perturbation interaction graph, obtain the interaction data between the non-forgotten users and the products; The interaction data of the non-forgotten users and the model parameters in the optimized student model are input into a preset second loss function to calculate the first intermediate value; Based on the optimized student model, the seventh preference rating of non-forgotten users for products they have interacted with is predicted, and the eighth preference rating of non-forgotten users for products they have not interacted with is predicted. The second intermediate value is obtained based on the seventh preference score, the eighth preference score, and the preset comparison fine-tuning loss function; The model parameters of the optimized student model are fine-tuned based on the first intermediate value and the second intermediate value to obtain a student model that has completed forgetting optimization.

[0024] Understandably, by constraining the ranking performance of the student model on the non-forgotten user group through the second loss function, and introducing a contrastive fine-tuning loss function, the virtual interactions generated during the topology perturbation stage are used as strong negative samples for contrastive learning. This allows for the precise location and removal of knowledge traces related to forgotten users in the parameter space. This not only achieves the complete erasure of specific user information, but also enhances the ability to identify the true preferences of the remaining users through contrastive learning. This effectively avoids recommendation performance fluctuations caused by forgetting operations and ensures that the student model maintains high-precision online service capabilities while exercising its right to be forgotten.

[0025] According to a second aspect of this application, a forgetting device for a recommendation system based on differential privacy contrastive learning is provided, the device comprising: The acquisition module is used to acquire the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the privacy sensitivity tag of the user based on the interaction data between the user and the product. The knowledge transfer module is used to transfer knowledge from the trained teacher model to the student model, resulting in a student model that has completed the knowledge transfer. The student model optimization module is used to predict user preferences for products using the teacher model and the student model that has completed knowledge transfer, respectively, obtain a loss value based on the preference results, and optimize the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model. The forget-recommendation module is used to perform random topological perturbation on the original interaction graph based on the privacy-sensitive label to obtain a perturbed interaction graph; and to fine-tune the model parameters of the optimized student model based on the forgotten user and the perturbed interaction graph to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

[0026] Based on any of the above aspects, the forgetting method and apparatus for a recommendation system based on differential privacy contrastive learning provided in this application can achieve the following benefits: This approach achieves synergistic optimization of personalized privacy protection and recommendation accuracy, avoiding the one-size-fits-all accuracy loss inherent in traditional differential privacy: By introducing user privacy sensitivity tags, differentiated processing is implemented during the knowledge transfer stage. When transferring knowledge from the teacher model to the student model, perturbation noise is injected into the embedding vector based on the user's sensitivity tags, ensuring strict protection of the privacy of users initiating forget requests. When constructing the perturbation interaction graph, the graph topology is randomized based on bit flipping probabilities. Users with higher sensitivity have a greater probability of their interaction edges being flipped, achieving a precise match between the strength of privacy protection and the user's actual needs, thus maximizing the preservation of the recommendation system's ability to characterize user preferences.

[0027] By employing knowledge transfer and forgetting fine-tuning, a secure manifold space with inherent privacy resistance is constructed: through personalized differential privacy knowledge transfer, the knowledge of the teacher model is transferred to the student model in a fuzzy form. The student model does not learn the original interaction pattern, but rather the knowledge after topological perturbation and embedding noise. This allows the student model to naturally form a desensitized manifold for sensitive user information in the parameter space. Even if an attacker attempts to reverse-engineer the student model, what is recovered is a fuzzy feature obfuscated by noise, thus blocking the privacy leakage link from the underlying architecture.

[0028] A contrastive fine-tuning mechanism based on perturbation topology is introduced to solve the challenge of balancing privacy erasure and performance maintenance: upon receiving a forgetting request, fine-tuning is performed based on the perturbation interaction graph and the interaction data between non-forgotten users and products, and optimization is carried out using the real interactions of non-forgotten users, which accelerates the student model's targeted forgetting of the forgotten user's associated information; at the same time, contrastive learning is used to enhance the student model's ability to identify the real interests of the remaining users, ensuring the immediate recovery and stable maintenance of recommendation performance during the fine-tuning process. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating a forgetting method for a recommendation system based on differential privacy contrastive learning, provided in this embodiment.

[0031] Figure 2 The flowchart for knowledge transfer provided in this embodiment.

[0032] Figure 3 The flowchart for the optimized student model provided in this embodiment.

[0033] Figure 4 The flowchart for fine-tuning the student model provided in this embodiment.

[0034] Figure 5 This is a schematic diagram of the functional modules of a forgetting device for a recommendation system based on differential privacy contrastive learning, provided in this embodiment. Detailed Implementation

[0035] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate the following embodiments, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0036] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0037] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0038] In existing systems, recommender systems face a contradiction between privacy preservation and performance loss when processing user data corresponding to users who initiate forgetting requests: user information is implicitly embedded in model parameters and spreads through graph structures, resulting in excessively high global retraining costs for traditional forgetting methods, and approximate forgetting causes a decline in recommender system performance, making it difficult to achieve a balance between thorough privacy protection and high-performance recommendation at low cost.

[0039] This embodiment provides a technical solution that can solve the above problems. The specific implementation of this application will be described in detail below with reference to the accompanying drawings.

[0040] like Figure 1 As shown, this embodiment provides a forgetting method for a recommendation system based on differential privacy contrastive learning. The method can be further divided into the following steps: S100. Based on the interaction data between the user and the product, obtain the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the privacy sensitivity tag of the user; Understandably, the user-product interaction data refers to the interaction data between users and products in the recommendation system, such as purchases and clicks, which can generally be obtained through relevant logs and / or databases of the recommendation system. In this embodiment, based on the user-product interaction data, the original interaction graph between users and products is obtained, transforming the original interaction data from a discrete, unstructured record format into graph structure data with topological relationships and high-order semantics. This provides a standardized computational foundation for subsequent mining of implicit paths between users and products and capturing collaborative filtering signals.

[0041] Specifically, the total user set in the recommendation system can be obtained. Commodity Collection User-product interaction dataset The total user set Includes several users The product set Includes several products Based on the total user set Commodity Collection Dataset of user-product interactions Construct the original interaction graph ;wherein, in the original interaction graph In this system, each user and each product corresponds to a node. The interaction state between a user and a product is set as follows: When the user interacts with the product, then ,otherwise, Then the dataset of interactions between each user and the product. It can be represented as: Then, in the original interaction diagram In this diagram, each user and each product they have interacted with forms an edge. All users and all products can form an interaction graph, i.e.: Understandably, users in a recommendation system have the right to be forgotten. This right is exercised by data subjects who actively request its removal from the recommendation system. Specifically, users who wish to exercise this right can initiate a forgetting request. The recommendation system can then use the user's interaction data with products to identify these requests and thus determine which user initiated the forgetting request. Preferably, in this embodiment, the forgetting users who initiated forgetting requests are aggregated to form a forgetting user set. .

[0042] Understandably, each user has a corresponding privacy sensitivity tag to reflect their privacy preferences regarding product interaction information. If a user initiates a forget request to the recommendation system, the user's privacy sensitivity tag has a higher sensitivity; if the user does not initiate a forget request to the recommendation system, the user's privacy sensitivity tag has a lower sensitivity. Specifically, the privacy sensitivity tag for each user can be defined based on forget requests in the user's interaction data with products.

[0043] S200. Use the trained teacher model to transfer knowledge to the student model to obtain a student model that has completed the knowledge transfer. In this embodiment, the teacher model can adopt a Graph Convolutional Network (GCN) architecture, wherein the teacher model may include an input layer, several graph convolutional layers, a fully connected layer, and an output layer connected in sequence. It is necessary to initialize the user embedding vectors and product embedding vectors of the several graph convolutional layers in the teacher model, and then base them on the original interaction graph. The teacher model is trained to obtain a well-trained teacher model.

[0044] Specifically, the teacher model has several graph convolutional layers; the method further includes: The original interaction graph between the user and the product is input into the teacher model, and the teacher model is trained to obtain a trained teacher model; wherein, during the training process, the user embedding vector and product embedding vector of each graph convolutional layer are obtained.

[0045] Understandably, the teacher model includes several graph convolutional layers, each capable of processing based on the original interaction graph. The corresponding user embedding vectors and product embedding vectors are obtained. These vectors guide the teacher model's high-quality global recommendation work, enabling it to learn knowledge such as the ability to accurately recommend products to users. Furthermore, they serve as a transfer vector for subsequent knowledge transfer in the student model. The user embedding vectors and product embedding vectors from each graph convolutional layer are saved and passed to the student model's differential privacy knowledge transfer process for learning. This allows the student model to construct a privacy-resistant secure manifold in the parameter space, inheriting knowledge from the teacher model while eliminating the risk of highly sensitive data leakage from the underlying architecture.

[0046] Preferably, when obtaining the user embedding vector and product embedding vector for each graph convolutional layer, a matrix processing method is used to obtain them accurately. Specifically, based on the original interaction graph... Obtain the adjacency matrix between users and products. Meanwhile, in order to eliminate the influence of graph size on feature distribution in the original interaction graph, the adjacency matrix is ​​modified. After performing symmetric normalization, the normalized Laplacian matrix is ​​obtained. : in, The degree matrix is ​​a preset degree matrix. elements in The connection variables that represent the nodes corresponding to users and / or products in the original interaction graph; Based on the matrix processing method described above, the user embedding vector and product embedding vector of each graph convolutional layer are obtained.

[0047] Specifically, the method for obtaining the user embedding vector of the current graph convolutional layer in the teacher model is as follows: The user corresponding to the user embedding vector of the current graph convolutional layer is taken as the first user; the product that has interacted with the first user is taken as the first product; the product embedding vector of the previous graph convolutional layer corresponding to the first product is obtained as the first product embedding vector; the first product embedding vector is obtained by weighted summation of all the first product embedding vectors; wherein the weight of the first product embedding vector is calculated based on the number of the first products and the number of users who have interacted with the first products.

[0048] For example, in this embodiment, the teacher model has Layer graph convolutional layer, the first user in the layer ... Layer User Embedding Vector for: in, , indicating the number of layers in the graph convolutional layer; Represents the first set of goods. Indicates the quantity of the first item; This represents the set of users who have interacted with the first product. This indicates the number of users who have interacted with the first product. This represents the embedding vector of the first product.

[0049] Specifically, the method for obtaining the product embedding vector of the current graph convolutional layer in the teacher model is as follows: The product corresponding to the product embedding vector of the current graph convolutional layer is taken as the second product; the user who has interacted with the second product is taken as the second user; the user embedding vector of the previous graph convolutional layer corresponding to the second user is obtained as the second user embedding vector; the weighted sum of all the second user embedding vectors is used to obtain the product embedding vector of the second product in the current graph convolutional layer; wherein the weight of the second user embedding vector is calculated based on the number of second users and the number of products that have interacted with the second user.

[0050] For example, the second product in the first Product embedding vectors of layers for: in, Describes the second user set. This indicates the number of the second user; This represents the set of products that have been interacted with by the second user. This indicates the number of items that have been interacted with by the second user; This represents the second user embedding vector.

[0051] Preferably, after obtaining the teacher model After generating user and item embedding vectors for each graph convolutional layer, a weighted average is applied to both. This captures both local interaction features between users and items, as well as knowledge from global collaborative filtering, generating final user and item representation vectors for comprehensive knowledge. This provides global features for subsequent teacher model optimization and knowledge transfer. The user representation vectors... and the commodity representation vector Represented as: in, This represents the inference contribution of each graph convolutional layer.

[0052] Specifically, before the step of transferring knowledge from the trained teacher model to the student model, the method further includes: Based on the trained teacher model, predict the user's first preference rating for products that have been interacted with, and predict the user's second preference rating for products that have not been interacted with. The first preference score, the second preference score, and the model parameters in the trained teacher model are input into a preset first loss function to optimize the trained teacher model and obtain an optimized teacher model. Step S200, which involves using the trained teacher model to transfer knowledge to the student model, specifically includes: The optimized teacher model is used to transfer knowledge from the student model.

[0053] In this embodiment, by using the optimized teacher model to transfer knowledge to the student model, the recommendation ability of the teacher model can be optimized, making the knowledge learned by the teacher model more accurate. This results in the student model performing better in recommendation capabilities after learning the recommendation knowledge from the teacher model. Preferably, the first loss function can be the Bayesian Personalized Ranking (BPR) loss function. Specifically, the first loss function is: Among them, commodities In order to connect with users Products that have had historical interactions can be considered positive samples of observed interactions; products In order to connect with users Products that have not had any historical interaction can be called negative samples of observed interaction; Indicates the user in the recommendation system and goods and goods The set of triples; It is the sigmoid activation function; This indicates that the rating is based on the first preference. Indicates the second preference rating; The preset regularization coefficients are used. The model parameters of the teacher model are represented.

[0054] Specifically, such as Figure 2 As shown, the step of using a trained teacher model to transfer knowledge from a student model to obtain a student model that has completed the knowledge transfer includes: S210. Based on the trained teacher model, the user corresponding to the current user embedding vector to be transferred is determined as the third user, the product corresponding to the current product embedding vector to be transferred is determined as the fourth product, and other users who have interacted with the fourth product are obtained as associated users. S220. Based on the privacy sensitivity tags of the third user and the associated user, determine the differential privacy budget parameters corresponding to the third user and the associated user respectively; In this embodiment, a specific differential privacy budget parameter can be set for each user's privacy sensitivity tag. The higher the sensitivity of the user's privacy sensitivity tag, the smaller the value of the differential privacy budget parameter, which represents the greater the intensity of the injected noise.

[0055] S230. Based on the differential privacy budget parameter corresponding to the third user, inject perturbation noise into the user embedding vector of the third user to obtain a user fuzzy embedding vector; and based on the differential privacy budget parameter corresponding to the associated user, inject perturbation noise into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector. In this embodiment, by introducing a perturbation noise injection mechanism based on user differential privacy budget parameters, the privacy protection strength is precisely matched with individual needs, thereby blocking the path of individual privacy leakage through product nodes at the source. While ensuring strict anonymization of sensitive information, the preference signals of low-sensitivity users are preserved to the maximum extent, laying a solid foundation for the subsequent construction of a secure manifold space for student models that combines privacy resistance and recommendation accuracy.

[0056] Specifically, the step of injecting perturbation noise into the user embedding vector of the third user based on the differential privacy budget parameter corresponding to the third user to obtain the user fuzzy embedding vector includes: The first perturbation noise is generated based on the preset hyperparameters of the Laplace distribution, the random variables of the Laplace distribution, and the differential privacy budget parameters of the third user. The first perturbation noise is then injected into the user embedding vector of the third user to obtain the user fuzzy embedding vector. The step of injecting perturbation noise into the product embedding vector of the fourth product based on the differential privacy budget parameter corresponding to the associated user to obtain the product fuzzy embedding vector includes: A second perturbation noise is generated based on the hyperparameters, the random variables, and the differential privacy budget parameters of the associated users, and the second perturbation noise is injected into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector.

[0057] For example, the user fuzzy representation vector and the fuzzy representation vector of the commodity ,include: in, Let be a random variable with a Laplace distribution. The preset hyperparameters of the Laplace distribution, This represents the differential privacy budget parameter corresponding to the third user; This represents the set of users associated with each other.

[0058] S240. Using the user fuzzy embedding vector and the product fuzzy embedding vector as supervision signals, the student model is trained to obtain a student model that has completed knowledge transfer.

[0059] Understandably, in this embodiment, the architecture of the student model can be similar to that of the teacher model, or it can adopt a more lightweight network architecture. The user fuzzy embedding vectors and product fuzzy embedding vectors obtained through knowledge transfer from the user embedding vectors and product embedding vectors of the teacher model can serve as supervisory signals to guide the training of the student model. This enables the student model to make accurate product recommendations to users in the future, while simultaneously constructing a secure manifold space with inherent privacy resistance, while ensuring the accuracy of the student model's recommendations.

[0060] S300. Predict user preferences for products using the teacher model and the student model that has completed knowledge transfer, respectively. Obtain a loss value based on the preference results. Optimize the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model. In this embodiment, a comparative optimization step is introduced after knowledge transfer in the student model to construct a dual-path distillation loss mechanism. This not only forces the student model to accurately fit the teacher's judgment of the user's explicit preferences, but also aligns it with its cognitive ranking of the sample space. This effectively prevents the recommendation list from being biased, significantly improves the ranking consistency and generalization ability of the student model, and lays a robust model foundation for subsequent privacy forgetting operations.

[0061] Specifically, such as Figure 3 As shown, the process of predicting user preferences for products using the teacher model and the student model that has completed knowledge transfer, obtaining a loss value based on the preference results, and optimizing the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model may include the following steps: S310. Based on the teacher model, predict the user's third preference rating for products that have been interacted with, and predict the user's fourth preference rating for products that have not been interacted with. In this embodiment, the third preference score and the fourth preference score can be predicted and obtained based on the trained teacher model or the optimized teacher model. Preferably, in this embodiment, the third preference score and the fourth preference score can be predicted and obtained based on the optimized teacher model.

[0062] S320. Based on the student model that has completed knowledge transfer, predict the user's fifth preference rating for products that have been interacted with, and predict the user's sixth preference rating for products that have not been interacted with. S330. Input the third preference score and the fifth preference score into a preset mean squared error function to obtain a first loss value; input the fourth preference score and the sixth preference score into the mean squared error function to obtain a second loss value; S340. Optimize the student model that has completed knowledge transfer based on the first loss value and the second loss value to obtain the optimized student model.

[0063] In this embodiment, knowledge transfer optimization is achieved by minimizing the difference in preference scores between the teacher model and the student model on sample triples. For example, the total divergence loss function can be used to optimize the student model that has completed knowledge transfer, wherein the total divergence loss function... for: in This indicates the fifth preference rating. This indicates the third preference rating. This indicates the sixth preference rating. This indicates the fourth preference rating; Let be the mean square error function.

[0064] S400. Based on the privacy sensitivity label, the original interaction graph is randomly topologically perturbed to obtain a perturbed interaction graph; based on the forgotten user and the perturbed interaction graph, the model parameters of the optimized student model are fine-tuned to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

[0065] Understandably, the perturbation interaction graph is based on privacy-sensitive labels to randomly perturb the original interaction graph, which can block privacy leakage at the graph structure level. By randomly perturbing real edges and adding fake edges, topology-level privacy obfuscation is achieved, effectively defending against member reasoning attacks based on graph structure.

[0066] Specifically, the perturbation interaction graph is as follows: Based on the user's privacy sensitivity tags, obtain the differential privacy budget parameters corresponding to the user; The corresponding bit-flip probability is obtained based on the user's differential privacy budget parameters; Obtain the total number of products that the user has interacted with in the original interaction graph; The first value is the product of the bit-flipping probability and the total number of products. The second value is the number of products that the user did not interact with in the original interaction graph. If the user and the product have interacted in the original interaction graph, then in the perturbation interaction graph, the complementary probabilities of the bit flip probabilities corresponding to the user follow a Bernoulli distribution. If the user and the product have not interacted in the original interaction graph, then the ratio of the first value to the second value follows a Bernoulli distribution.

[0067] For example, the bit flip probability It can be: in, This is a truncation function. To define the lower bound of the truncation function, preferably, It can be 0; To define the lower bound of the truncation function, preferably, You can take 0.5.

[0068] For example, the perturbation interaction graph Chinese users and goods Interaction status between for: in, and Both represent Bernoulli distributions; Indicates the user In the original interaction diagram The collection of products that have been interacted with in the process. Indicates the user In the original interaction diagram The total number of products that have been interacted with in the process; This represents the product of the bit-flipping probability and the total number of items; This represents the set of all items in the recommendation system. Indicates the user In the original interaction diagram The number of products that have not had any interaction.

[0069] Understandably, this invention fine-tunes the model parameters of the optimized student model based on forgotten users and perturbation interaction graphs, enabling precise location and removal of knowledge traces related to forgotten users in the model parameter space of the student model, while effectively maintaining the recommendation performance of non-forgotten users, avoiding accuracy oscillations caused by forgetting operations, and ensuring the stability of the student model's online service under the premise of strictly adhering to the right to be forgotten.

[0070] Specifically, such as Figure 4 The step of fine-tuning the model parameters of the optimized student model based on the forgotten user and the perturbation interaction graph may include the following steps: S410. Obtain interaction data between non-forgotten users and products based on the forgotten users and the perturbation interaction graph; In this embodiment, the forgotten user identifier set can be used as a reference. Total user set Obtain the non-forgotten user set ; and based on the perturbation interaction graph Obtain a non-forgotten user-product interaction dataset ,commodity In order to connect with users Products that have had historical interactions In order to connect with users Products for which there has been no prior interaction.

[0071] S420. Input the interaction data of the non-forgotten user and the model parameters in the optimized student model into a preset second loss function to calculate the first intermediate value; S430. Based on the optimized student model, predict the seventh preference rating of the non-forgotten user for the products with which they have interacted, and predict the eighth preference rating of the non-forgotten user for the products with which they have not interacted. S440. Obtain a second intermediate value based on the seventh preference score, the eighth preference score, and a preset comparison fine-tuning loss function; S450. Based on the first intermediate value and the second intermediate value, the model parameters of the optimized student model are fine-tuned to obtain a student model that has completed forgetting optimization.

[0072] For example, the second loss function can be the Bayesian personalized ranking loss function; the model parameters of the fine-tuned student model for: in, This represents a function that takes the minimum value of the independent variable. This represents the Bayesian personalized ranking loss function, which considers several triples. The first intermediate value is obtained by summing the corresponding Bayesian personalized ranking loss values. This represents the set of spurious interactions corresponding to the addition of spurious interaction edges to the perturbation interaction graph. This represents the set of real interactions between users and products that are not forgotten. For non-forgotten users, products This refers to products that have been genuinely interacted with by users who have not been forgotten. This indicates that the product represents a fake interaction by a user who is not actually forgotten. This represents the model parameters of the student model before optimization. This refers to the contrastive fine-tuning loss function, which allows us to obtain the corresponding second intermediate value. It is the sigmoid activation function. This represents the seventh preference rating. This represents the eighth preference score.

[0073] Understandably, in practical applications, using a student model that has completed forgetting optimization to recommend products to users in a recommendation system can resolve the contradiction between privacy compliance and the performance of online recommendation services. It can accurately isolate user privacy in complex graph structures and quickly perform forgetting operations on users who initiate forgetting requests through lightweight fine-tuning, ensuring that the recommendation system maintains high generalization accuracy in recommendation tasks during dynamic updates.

[0074] like Figure 5 As shown in the illustration, this application also provides a forgetting device for a recommendation system based on differential privacy contrastive learning. Optionally, the device includes: The modules are: Acquisition Module 511, Knowledge Transfer Module 512, Student Model Optimization Module 513, and Forgetting-Recommendation Module 514, among which: The acquisition module 511 is used to acquire the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the privacy sensitivity tag of the user based on the interaction data between the user and the product. In this embodiment, the acquisition module 511 can be used to perform... Figure 1 For a detailed description of the acquisition module 511, please refer to the description of step S100 shown.

[0075] The knowledge transfer module 512 is used to transfer knowledge from the trained teacher model to the student model, resulting in a student model that has completed the knowledge transfer. In this embodiment, the knowledge transfer module 512 can be used to execute Figure 1For a detailed description of the knowledge transfer module 512, please refer to the description of step S200 shown.

[0076] Specifically, the knowledge transfer module 512 is further configured to, based on the trained teacher model, identify the user corresponding to the current user embedding vector to be transferred as the third user, identify the product corresponding to the current product embedding vector to be transferred as the fourth product, and obtain other users who have interacted with the fourth product as associated users; determine the differential privacy budget parameters corresponding to the third user and the associated users based on the privacy sensitivity tags of the third user and the associated users respectively; inject perturbation noise into the user embedding vector of the third user according to the differential privacy budget parameters corresponding to the third user to obtain a user fuzzy embedding vector; and inject perturbation noise into the product embedding vector of the fourth product according to the differential privacy budget parameters corresponding to the associated users to obtain a product fuzzy embedding vector; and use the user fuzzy embedding vector and the product fuzzy embedding vector as supervision signals to guide the student model to train, thereby obtaining a student model that has completed knowledge transfer.

[0077] In this embodiment, the knowledge transfer module 512 can also be used to execute Figure 2 For a detailed description of the knowledge transfer module 512, please refer to the description of steps S210-S240 shown.

[0078] The student model optimization module 513 is used to predict the user's preference for the product using the teacher model and the student model that has completed knowledge transfer, respectively, obtain a loss value based on the preference result, and optimize the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model. In this embodiment, the student model optimization module 513 can be used to perform... Figure 1 For a detailed description of the student model optimization module 513, see step S300 shown below. For a detailed description of step S300, please refer to the description of step S300.

[0079] Specifically, the student model optimization module 513 is further configured to, based on the teacher model, predict the user's third preference rating for products with which they have interacted, and predict the user's fourth preference rating for products without which they have not interacted; based on the student model that has completed knowledge transfer, predict the user's fifth preference rating for products with which they have interacted, and predict the user's sixth preference rating for products without which they have not interacted; input the third preference rating and the fifth preference rating into a preset mean squared error function to obtain a first loss value; input the fourth preference rating and the sixth preference rating into the mean squared error function to obtain a second loss value; and optimize the student model that has completed knowledge transfer based on the first loss value and the second loss value to obtain an optimized student model.

[0080] In this embodiment, the student model optimization module 513 can also be used to perform... Figure 3 For a detailed description of the student model optimization module 513, see steps S310-S340 shown below. For further details on steps S310-S340, please refer to the description of steps S310-S340.

[0081] The forgetting-recommendation module 514 is used to perform random topological perturbation on the original interaction graph based on the privacy sensitivity label to obtain a perturbed interaction graph; and to fine-tune the model parameters of the optimized student model based on the forgotten user and the perturbed interaction graph to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

[0082] In this embodiment, the forget-recommendation module 514 can be used to perform... Figure 1 For a detailed description of the forget-recommendation module 514, please refer to the description of step S400 shown.

[0083] Specifically, the forgetting-recommendation module 514 is further configured to obtain interaction data between non-forgotten users and products based on the forgotten users and the perturbation interaction graph; input the interaction data of the non-forgotten users and the model parameters in the optimized student model into a preset second loss function to calculate a first intermediate value; based on the optimized student model, predict the seventh preference rating of non-forgotten users for products that have been interacted with, and predict the eighth preference rating of non-forgotten users for products that have not been interacted with; obtain a second intermediate value based on the seventh preference rating, the eighth preference rating, and a preset comparison fine-tuning loss function; and fine-tune the model parameters of the optimized student model based on the first intermediate value and the second intermediate value to obtain a student model that has completed forgetting optimization.

[0084] In this embodiment, the forget-recommendation module 514 may also be used to perform Figure 4For a detailed description of the forget-recommendation module 514, see steps S410-S450 shown. Further details regarding steps S410-S450 can be found in the description of steps S410-S450.

[0085] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A forgetting method for recommendation systems based on differential privacy contrastive learning, characterized in that, The method includes: Based on the user-product interaction data, obtain the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the user's privacy sensitivity tag; The student model is used to transfer knowledge from the trained teacher model to obtain a student model that has completed the knowledge transfer. The teacher model and the student model that has completed knowledge transfer are used to predict the user's preference for the product, respectively. The loss value is obtained based on the preference result. The student model that has completed knowledge transfer is optimized based on the loss value to obtain the optimized student model. The original interaction graph is randomly perturbed based on the privacy sensitivity label to obtain a perturbed interaction graph; the model parameters of the optimized student model are fine-tuned based on the forgotten user and the perturbed interaction graph to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

2. The method according to claim 1, characterized in that, The teacher model has several graph convolutional layers; the method further includes: The original interaction graph between the user and the product is input into the teacher model, and the teacher model is trained to obtain a trained teacher model; wherein, during the training process, the user embedding vector and product embedding vector of each graph convolutional layer are obtained.

3. The method according to claim 2, characterized in that, The method for obtaining the user embedding vector of the current graph convolutional layer in the teacher model is as follows: The user corresponding to the user embedding vector of the current graph convolutional layer is taken as the first user; the product that has interacted with the first user is taken as the first product; the product embedding vector of the previous graph convolutional layer corresponding to the first product is taken as the first product embedding vector; and all the first product embedding vectors are weighted and summed to obtain the user embedding vector of the first user in the current graph convolutional layer. The weight of the first product embedding vector is calculated based on the number of the first product and the number of users who have interacted with the first product.

4. The method according to claim 2, characterized in that, The method for obtaining the product embedding vector of the current graph convolutional layer in the teacher model is as follows: The product corresponding to the product embedding vector of the current graph convolutional layer is taken as the second product; the user who has interacted with the second product is taken as the second user; the user embedding vector of the previous graph convolutional layer corresponding to the second user is obtained as the second user embedding vector; the weighted sum of all the second user embedding vectors is used to obtain the product embedding vector of the second product in the current graph convolutional layer; The weights of the second user embedding vector are calculated based on the number of the second user and the number of products that the second user has interacted with.

5. The method according to claim 1, characterized in that, Before the step of transferring knowledge from the trained teacher model to the student model, the method further includes: Based on the trained teacher model, predict the user's first preference rating for products that have been interacted with, and predict the user's second preference rating for products that have not been interacted with. The first preference score, the second preference score, and the model parameters in the trained teacher model are input into a preset first loss function to optimize the trained teacher model and obtain an optimized teacher model. The process of transferring knowledge from the trained teacher model to the student model specifically involves: The optimized teacher model is used to transfer knowledge from the student model.

6. The method according to claim 2, characterized in that, The process of transferring knowledge from a trained teacher model to a student model to obtain a student model that has completed the knowledge transfer includes: Based on the trained teacher model, the user corresponding to the current user embedding vector to be transferred is determined as the third user, the product corresponding to the current product embedding vector to be transferred is determined as the fourth product, and other users who have interacted with the fourth product are obtained as associated users. Based on the privacy sensitivity tags of the third user and the associated user, the differential privacy budget parameters corresponding to the third user and the associated user are determined respectively; Based on the differential privacy budget parameters corresponding to the third user, perturbation noise is injected into the user embedding vector of the third user to obtain a user fuzzy embedding vector; and based on the differential privacy budget parameters corresponding to the associated user, perturbation noise is injected into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector. The user fuzzy embedding vector and the product fuzzy embedding vector are used as supervision signals to guide the training of the student model, resulting in a student model that has completed knowledge transfer.

7. The method according to claim 6, characterized in that, The step of injecting perturbation noise into the user embedding vector of the third user based on the differential privacy budget parameter corresponding to the third user to obtain the user fuzzy embedding vector includes: The first perturbation noise is generated based on the preset hyperparameters of the Laplace distribution, the random variables of the Laplace distribution, and the differential privacy budget parameters of the third user. The first perturbation noise is then injected into the user embedding vector of the third user to obtain the user fuzzy embedding vector. The step of injecting perturbation noise into the product embedding vector of the fourth product based on the differential privacy budget parameter corresponding to the associated user to obtain the product fuzzy embedding vector includes: A second perturbation noise is generated based on the hyperparameters, the random variables, and the differential privacy budget parameters of the associated users, and the second perturbation noise is injected into the product embedding vector of the fourth product to obtain a product fuzzy embedding vector.

8. The method according to claim 1, characterized in that, The perturbation interaction graph is specifically as follows: Based on the user's privacy sensitivity tags, obtain the differential privacy budget parameters corresponding to the user; The corresponding bit-flip probability is obtained based on the user's differential privacy budget parameters; Obtain the total number of products that the user has interacted with in the original interaction graph; The product of the bit-flipping probability and the total number of products is the first value, and the number of products that the user did not interact with in the original interaction graph is the second value. If the user and the product have interacted in the original interaction graph, then in the perturbation interaction graph, the complementary probabilities of the bit flip probabilities corresponding to the user follow a Bernoulli distribution. If the user and the product have not interacted in the original interaction graph, then the ratio of the first value to the second value follows a Bernoulli distribution.

9. The method according to claim 1, characterized in that, The process involves predicting user preferences for products using both the teacher model and the student model that has completed knowledge transfer, obtaining a loss value based on the preference results, and optimizing the student model based on the loss value to obtain an optimized student model. This includes: Based on the teacher model, predict the user's third preference rating for products that have been interacted with, and predict the user's fourth preference rating for products that have not been interacted with. Based on the student model that has completed knowledge transfer, predict the user's fifth preference rating for products that have been interacted with, and predict the user's sixth preference rating for products that have not been interacted with. The third preference score and the fifth preference score are input into a preset mean squared error function to obtain a first loss value; the fourth preference score and the sixth preference score are input into the mean squared error function to obtain a second loss value; The student model that has completed knowledge transfer is optimized based on the first loss value and the second loss value to obtain the optimized student model.

10. The method according to claim 1, characterized in that, The fine-tuning of the model parameters of the optimized student model based on the forgotten user and the perturbation interaction graph includes: Based on the forgotten users and the perturbation interaction graph, obtain the interaction data between the non-forgotten users and the products; The interaction data of the non-forgotten users and the model parameters in the optimized student model are input into a preset second loss function to calculate the first intermediate value; Based on the optimized student model, the seventh preference rating of non-forgotten users for products they have interacted with is predicted, and the eighth preference rating of non-forgotten users for products they have not interacted with is predicted. The second intermediate value is obtained based on the seventh preference score, the eighth preference score, and the preset comparison fine-tuning loss function; The model parameters of the optimized student model are fine-tuned based on the first intermediate value and the second intermediate value to obtain a student model that has completed forgetting optimization.

11. A forgetting device for a recommendation system based on differential privacy contrastive learning, characterized in that, The device includes: The acquisition module is used to acquire the original interaction graph between the user and the product, the forgotten user who initiated the forgetting request, and the privacy sensitivity tag of the user based on the interaction data between the user and the product. The knowledge transfer module is used to transfer knowledge from the trained teacher model to the student model, resulting in a student model that has completed the knowledge transfer. The student model optimization module is used to predict user preferences for products using the teacher model and the student model that has completed knowledge transfer, respectively, obtain a loss value based on the preference results, and optimize the student model that has completed knowledge transfer based on the loss value to obtain an optimized student model. The forget-recommendation module is used to perform random topological perturbation on the original interaction graph based on the privacy-sensitive label to obtain a perturbed interaction graph; and to fine-tune the model parameters of the optimized student model based on the forgotten user and the perturbed interaction graph to obtain a student model that has completed forgetting optimization, which is used to recommend products to users in the recommendation system.

Images