An intent-driven federated graph enhanced cross-domain recommendation method under privacy protection

Abstract

The application provides an intention-driven federated graph enhanced cross-domain recommendation method under privacy protection, comprising the following steps: S1, extracting a set of local intention prototypes representing the behavior patterns of a local user group; S2, each domain client performs disturbance processing on the local intention prototypes based on an adaptive differential privacy mechanism of intention popularity awareness, uploads the intention prototypes after adding noise to a central server, and generates global intention prototypes; S3, each domain client receives the global intention prototypes issued by the central server, performs cross-domain potential edge inference based on the global awareness representation to generate an enhanced graph structure, and optimizes the local recommendation model through double-view contrast learning, and finally outputs the recommendation result for the target user by the optimized local recommendation model. The application effectively overcomes the client drift problem caused by Non-IID data, so that the model has strong global generalization ability and accurately meets the personalized needs of local users.

Description

Technical Field

[0001] This invention relates to the field of cross-domain recommendation technology, and more particularly to a privacy-preserving intent-driven federated graph-enhanced cross-domain recommendation method. Background Technology

[0002] With the explosive growth of internet information, recommender systems have become a core technology for alleviating information overload. However, in practical applications, single-domain recommender systems generally face severe data sparsity and cold-start problems, making it difficult for models to accurately capture users' true preferences. Cross-domain recommender technology has emerged to address this issue, aiming to introduce rich knowledge from the source domain to assist in target domain recommendations and alleviate the data sparsity problem. Federated learning, as a privacy-preserving distributed machine learning paradigm, allows clients to collaboratively train models while retaining the original data locally. In recent years, the combination of federated learning and cross-domain recommender has provided a new technical path for breaking down data silos and achieving knowledge transfer under privacy compliance.

[0003] Existing federated cross-domain recommendation methods typically employ the following technical solutions: each client trains a recommendation model using local data, uploads the model parameters or gradients to a central server for aggregation, and the server distributes the global model to the clients for iterative updates. For example, some solutions achieve knowledge transfer by sharing cross-domain alignment modules for user embeddings or item embeddings, while others apply global differential privacy noise to the client to defend against reverse engineering attacks targeting the original data. These methods achieve cross-domain collaboration with privacy protection to a certain extent; however, due to the stringent data isolation environment in federated scenarios, significant technical bottlenecks still exist in actual deployment.

[0004] The shortcomings of existing technologies are mainly reflected in two aspects. First, the contradiction between the non-independent and identically distributed problem and the local sparse graph representation capability is prominent. Due to significant differences in user group behavior patterns across different clients, the local data distribution often exhibits huge deviations. The global model struggles to achieve an effective balance between local personalization and global generalization, leading to severe client drift. Second, a sharp conflict exists between existing privacy protection mechanisms and the preservation of intent semantics. To defend against feature reverse inference attacks, traditional methods typically apply indiscriminate numerical perturbations to the data uploaded by each client. This practice severely disrupts the original distribution structure of user intent in the high-level semantic space, causing significant semantic deviations when the central server performs cross-domain aggregation. Cross-domain intent alignment thus fails, resulting in a technical dilemma where protecting privacy inevitably sacrifices recommendation accuracy. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a privacy-preserving intention-driven federated graph-enhanced cross-domain recommendation method to address the prominent contradiction between the non-independent and identically distributed problem and the local sparse graph representation capability in existing technologies.

[0006] The technical means employed in this invention are as follows:

[0007] A privacy-preserving, intent-driven federated graph-enhanced cross-domain recommendation method includes the following steps: S1. Each domain client encodes the local user-item interaction graph using a graph neural network and extracts a set of local intent prototypes representing the behavior patterns of local user groups through a soft clustering mechanism. S2. Each domain client perturbs the local intent prototype based on an adaptive differential privacy mechanism that is aware of intent popularity, and uploads the noisy intent prototype to the central server. The central server aggregates the noisy intent prototypes uploaded by all clients into a global prototype pool, and performs a clustering algorithm on the global prototype pool to generate a global intent prototype. S3. Each domain client receives the global intent prototype issued by the central server, reconstructs the global perception representation of users and items using the global intent prototype, performs cross-domain potential edge inference based on the global perception representation to generate an enhanced graph structure, optimizes the local recommendation model through dual-view comparison learning, and finally outputs the recommendation result for the target user by the optimized local recommendation model.

[0008] Furthermore, S1 specifically includes the following steps: S11. On the m-th local client, initialize the user node embedding representation. and item node embedding representation and define The local intention prototype matrix is ​​composed of learnable intention prototype vectors. ; S12. In the l-th layer of this map neural network, the current representation of the user node is calculated through message passing. and the current representation of the item node The calculation formula is:

[0009]

[0010] in, Represents the i-th user The neighborhood group, Represents the j-th item The neighborhood group, They are respectively the ( Layer users With items Embedded representation; S13. Calculate the current representation of the user node. With each local intent prototype The correlation scores between them yield the soft clustering probability distribution. The calculation formula is:

[0011] S14. Based on the soft clustering probability distribution pair The intent prototypes are weighted and summed to generate an aggregated representation containing higher-order intent information. The calculation formula is:

[0012] S15. Extracting local graph topological features from graph neural networks. With the aforementioned polymerization characterization Add them together to obtain the final user representation of the l-th layer. The formula is as follows:

[0013] S16. The local graph recommendation model is trained end-to-end by minimizing the Bayesian personalized ranking loss, and the local intent prototype matrix is ​​optimized during backpropagation. .

[0014] Furthermore, the intention popularity-aware adaptive differential privacy mechanism in S2 includes the following steps: S21. Calculate the k-th intent prototype using the soft clustering probability distribution of user nodes on the intent prototype. Local intent popularity weight and to The relative popularity coefficient is obtained by normalization. The formula is as follows:

[0015]

[0016] S22. Based on the relative popularity coefficient, design an adaptive noise variance scheduling function and calculate the specific Gaussian noise standard deviation of the k-th intention prototype. The formula is:

[0017] in, Based on privacy budget parameters, Sensitivity hyperparameters for controlling noise attenuation rate; S23. Apply an adaptive threshold to perform L2 norm pruning on the intent prototype to obtain the pruned intent prototype. The formula is:

[0018] S24. Inject adaptive Gaussian noise into the trimmed intent prototype to generate a noisy intent prototype with privacy protection. The formula is as follows:

[0019] in, This represents the noise vector sampled from a multidimensional Gaussian distribution. The adaptive noise standard deviation calculated in S22; S25. Each client will submit the noisy intent prototype set. Uploaded to the central server.

[0020] Furthermore, the specific steps of the clustering algorithm for generating global intent prototypes in S2 are as follows: The central server receives all Uploaded by each client These noisy prototypes constitute the global prototype pool. ; in the global prototype pool Clustering is performed, and the objective function that minimizes the clustering error is solved to obtain K global intent prototypes. ;

[0021] The clustering error is minimized as follows: .

[0022] Furthermore, the S3 method for reconstructing the global perception representation of users and items using the global intent prototype specifically includes the following steps: S31. Obtain the final embedded representation of the local user node after passing through all graph neural network layers, and convert the final layer representation... abbreviated as And the final embedded representation of the local item node, abbreviated as ; S32, Compute user nodes In global intent prototype Attention probability distribution and item nodes In global intent prototype Attention probability distribution The calculation formula is:

[0023]

[0024] in, Temperature hyperparameters for controlling the smoothness of attention distribution; S33. Reconstruct the globally perceived user representation based on the attention probability distribution. and global awareness item representation The formula is as follows: .

[0025] Furthermore, the cross-domain potential edge inference and enhanced graph structure generation in S3 includes the following steps: S34. Calculate the cross-domain semantic affinity score between any user-item pair. The formula is as follows:

[0026] in, , The projection matrix is ​​learnable. For bias terms, This is the mapping weight vector; S35. Add independent and identically distributed Gumbel noise to the affinity score. The Sigmoid function is used for continuous relaxation to calculate the continuous probability of candidate edges being activated. And it is binarized into discrete pseudo-edges using a direct-pass estimator. The formula is:

[0027]

[0028] S36. A residual topology fusion mechanism is used to combine the generated pseudo-edge matrix with the locally observed real discrete adjacency matrix. The images are then merged to generate the final enhanced graph view. The formula is as follows: .

[0029] Furthermore, the dual-view comparison learning in S3 includes the following steps: S37. The original local discrete graph... With the generated enhanced graph The original node representations are obtained by inputting them into a shared local graph neural network encoder. ) and enhanced node representation ( ); S38. Calculate the node-level structural contrast loss, using the representation generated from the original graph as the anchor point, the representation generated from the augmented graph as the positive sample, and other nodes of the same type within the same batch as negative samples. The user contrast loss is calculated using this method. Compared with the loss of the items The formula is:

[0030]

[0031] The total structural contrast loss is: ; in, For the current training batch, Represents the cosine similarity function. To control the intensity of punishment for difficult negative samples, a contrastive learning temperature parameter was used. S39. Using Bayesian personalized ranking loss The formula for optimizing the local recommendation main task is:

[0032] S310. Construct the end-to-end joint optimization objective function as follows:

[0033] in, To compare the learning of weight hyperparameters, The L2 regularization coefficient is... This represents all learnable parameters in the model.

[0034] The present invention also provides a storage medium comprising a stored program, wherein, when the program is executed, it performs any of the above-described privacy-preserving intent-driven federated graph enhanced cross-domain recommendation methods.

[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes any of the above-described privacy-preserving intent-driven federated graph enhanced cross-domain recommendation methods through the computer program.

[0036] Compared with the prior art, the present invention has the following advantages: This invention innovatively employs intent prototypes as a medium for cross-domain interaction. Through the synergy of global intent aggregation and local contrastive learning, it allows models in each domain to retain their local personalized network weights to cope with differences in data distribution while aligning towards global consensus in the macro-semantic space. This mechanism effectively overcomes the client drift problem caused by Non-IID data, enabling the model to possess both strong global generalization capabilities and accurately meet the personalized needs of local users.

[0037] This invention designs a semantically preserved privacy protection mechanism for intent prototype interactions. While effectively obfuscating features and successfully defending against reverse inference attacks on the original data, it maintains the relative semantic structure of the intent vector in the representation space to the greatest extent. This mechanism completely avoids the cross-domain aggregation bias caused by traditional noise addition, breaks down the conflict between privacy protection and recommendation accuracy, and achieves a win-win situation for security and accuracy in cross-domain federated recommendation. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a diagram of the method architecture of the present invention. Detailed Implementation

[0040] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0041] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 the invention 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 a 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.

[0042] like Figure 1 As shown, this invention provides a privacy-preserving intent-driven federated graph-enhanced cross-domain recommendation method, comprising the following steps: S1. Each domain client locally encodes the local user-item interaction graph using a graph neural network to obtain node features. It then extracts a set of local intent prototypes representing the behavioral patterns of local user groups through a soft clustering mechanism. S11. On the m-th local client, initialize the user node embedding representation. and item node embedding representation and define The local intention prototype matrix is ​​composed of learnable intention prototype vectors. This vector set serves as a globally shared parameter for the local model, aiming to characterize cluster centers of potential local user behaviors in a high-dimensional latent space. These latent behaviors include purchase motivation and aesthetic preferences.

[0043] S12. In the l-th layer of this map neural network, the current representation of the user node is calculated through message passing. and the current representation of the item node The calculation formula is:

[0044]

[0045] in, Represents the i-th user The neighborhood group, Represents the j-th item The neighborhood group, They are respectively the ( Layer users With items Embedded representation; S13. Calculate the current representation of the user node. With each local intent prototype The correlation score between them indicates whether the user node belongs to the first... Soft clustering probability distribution of each intention The calculation is performed using the inner product combined with the Softmax function:

[0046] S14. Based on the soft clustering probability distribution pair The intent prototypes are weighted and summed to generate an aggregated representation containing higher-order intent information. This enables knowledge mapping from individual behavior to group patterns. Aggregated representation. The calculation formula is:

[0047] S15. Extracting local graph topological features from graph neural networks. With the aforementioned polymerization characterization Add them together to obtain the final user representation of the l-th layer. The formula is as follows:

[0048] S16. The local graph recommendation model is trained end-to-end by minimizing the Bayesian personalized ranking loss. During backpropagation, the gradient updates not only the node embeddings. It will also send back and adaptively optimize the intent prototype matrix. This causes it to continuously approach the true distribution center.

[0049] S2. On the client side, noise is dynamically added according to the popularity of intents. The popularity of each intent is evaluated. More noise is added to niche / long-tail intents, while more semantics are preserved for popular intents, thus achieving adaptive local differential privacy. After the server receives the noisy prototype, it re-divides the multi-domain prototype pool through a clustering algorithm to accurately align the semantic misalignments between different clients and generate a global intent prototype.

[0050] The intent popularity-aware adaptive differential privacy mechanism includes the following steps: S21. Upload the local intent prototype to the federated server from each domain client. ( Before identifying the client, to prevent malicious attackers from inferring local privacy data through intent distribution, we perform local differential privacy processing on the client side. We calculate the k-th intent prototype using the soft clustering probability distribution of user nodes on the intent prototype. Local intent popularity weight and to The relative popularity coefficient is obtained by normalization. The formula is as follows:

[0051]

[0052] This coefficient reflects the breadth of the intended audience. The smaller the value, the longer-tailed and more personalized the intention, and the higher the risk of privacy breaches.

[0053] S22. Based on the relative popularity coefficient, design an adaptive noise variance scheduling function and calculate the specific Gaussian noise standard deviation of the k-th intention prototype. The formula is:

[0054] in, Based on privacy budget parameters, This is a sensitivity hyperparameter for controlling the noise attenuation rate; through this function, the system can apply a stronger noise mask to high-risk niche intentions, while retaining more original semantic information for mass intentions.

[0055] S23. Before injecting noise, to limit local sensitivity, an adaptive threshold is used to perform L2 norm pruning on the intention prototype, resulting in the pruned intention prototype. The formula is:

[0056] in, This is the preset cropping threshold.

[0057] S24. Inject adaptive Gaussian noise into the trimmed intent prototype to generate a noisy intent prototype with privacy protection. The formula is as follows:

[0058] in, This represents the noise vector sampled from a multidimensional Gaussian distribution. The adaptive noise standard deviation calculated in S22; S25. Each client will submit the noisy intent prototype set. Uploaded to the central server.

[0059] The specific steps of the clustering algorithm for generating global intent prototypes are as follows: The central server receives all Uploaded by each client A noisy intent prototype is generated and then globally aggregated. Because different clients independently extract intents locally, there is often a semantic misalignment issue in the indexing order of the intent prototypes (i.e., the order of intents extracted by different clients). This could represent a very different preference space.

[0060] To solve this problem, the server does not use a simple bit-coordinate averaging, but instead uses all... Uploaded by each client Individual intent prototypes are aggregated into a global prototype pool. ; in the global prototype pool Perform clustering to extract K global semantic centers are obtained by solving the objective function that minimizes the clustering error. ;

[0061] The clustering error is minimized as follows: .

[0062] Through clustering and aggregation, the server can adaptively group local intents from different domains that are semantically similar into the same cluster, thereby calculating the center of the aligned global intent cluster and sending it out as a global intent prototype.

[0063] S3. Each domain client receives the global intent prototype issued by the central server, reconstructs the global perception representation of users and items using the global intent prototype, performs cross-domain potential edge inference based on the global perception representation to generate an enhanced graph structure, optimizes the local recommendation model through dual-view comparison learning, and finally outputs the recommendation result for the target user by the optimized local recommendation model.

[0064] Reconstructing the global perception representation of users and items using the global intent prototype specifically includes the following steps: S31. Obtain the final embedding representation of the local user node after passing through all graph neural network layers. For simplicity, the final layer representation will be... abbreviated as And the final embedded representation of the local item node, abbreviated as ; S32. Using the Softmax function and inner product operation, calculate the user node... In global intent prototype Attention probability distribution and item nodes In global intent prototype Attention probability distribution The calculation formula is:

[0065]

[0066] in, Temperature hyperparameters for controlling the smoothness of attention distribution; S33. Based on the attention probability distribution, local users and items are jointly mapped to a unified global intent latent space, reconstructing the globally perceived user representation. and global awareness item representation The formula is as follows: .

[0067] Cross-domain potential edge inference and enhanced graph structure generation include the following steps: S34. Based on the reconstructed global perception representation, a feature fusion network is introduced to evaluate the potential connection probability between non-interacting nodes. The cross-domain semantic affinity score between any user-item pair is calculated. The formula is as follows:

[0068] in, , The projection matrix is ​​learnable. For bias terms, This is the mapped weight vector; the score This constitutes a dense global semantic scoring matrix. .

[0069] S35. To introduce randomness for contrastive learning during graph generation and to ensure end-to-end differentiability of the discrete sampling process, we directly... As a distribution parameter, it is superimposed with independent and identically distributed Gumbel noise. The Sigmoid function is used for continuous relaxation to calculate the continuous probability of candidate edges being activated. And it is binarized into discrete pseudo-edges using a direct-pass estimator. The formula is:

[0070]

[0071] S36. To avoid the generated pseudo-edges from corrupting the actual interaction data, a residual topology fusion mechanism is used to combine the generated pseudo-edge matrix with the locally observed real discrete adjacency matrix. The images are then merged to generate the final enhanced graph view. The formula is as follows: .

[0072] Dual-view comparison learning includes the following steps: S37. To ensure the model can effectively absorb the collaborative signals brought by the global intent, while avoiding overfitting to noise in the augmented graph, a dual-view graph contrastive learning paradigm is introduced. The local original discrete graph... With the generated enhanced graph The original node representations are obtained by inputting them into a shared local graph neural network encoder and passing them through a message passing mechanism. ) and enhanced node representation ( ); S38. Calculate the node-level structural contrast loss, using the representation generated from the original graph as the anchor point, the representation generated from the augmented graph as the positive sample, and other nodes of the same type within the same batch as negative samples. Based on the InfoNCE framework, calculate the user contrast loss. Compared with the loss of the items The formula is:

[0073]

[0074] The total structural contrast loss is: ; in, For the current training batch, Represents the cosine similarity function. To control the intensity of punishment for difficult negative samples, a contrastive learning temperature parameter was used. S39. In the local recommendation main task, Bayesian personalized ranking loss is adopted. The formula for optimizing the local recommendation main task is:

[0075] in, This represents the observed interaction data. For users Positive sample items that have been interacted with For negative sample items that have not been interacted with, This refers to the Sigmoid function.

[0076] S310. Combining the recommendation task loss with the contrastive learning loss, an end-to-end joint optimization objective function is constructed as follows:

[0077] in, To compare the learning of weight hyperparameters, The L2 regularization coefficient is... This represents all learnable parameters in the model, including the parameters of the recommendation model and the projection moments of the feature fusion network. By minimizing This drives the adaptive enhancement of local graph structures and the synergistic evolution of recommended representations.

[0078] Example This invention uses the classic recommendation system Amazon dataset, selecting different business domains: Movies, Videos, and Games. During the joint training process between local clients and the central server, the system employs the PyTorch framework for model construction and the Adam optimizer for end-to-end parameter updates. To ensure the alignment of multi-domain features and model convergence, the system adopts the following preferred hyperparameter configuration: the initial embedding dimension for user and item nodes is uniformly set to 64, and the local training batch size is set to 1024. Learning Rate and Regularization: The initial learning rate is optimized using a grid search within the range of 1e-4 to 5e-3; the number of global intent prototypes is selected from the set {10, 20, 50, 100}; the temperature parameters controlling the smoothness of the contrastive loss between probabilistic sampling and InfoNCE are selected from {0.1, 0.2, 0.5, 1.0}; the contrastive learning loss weight ratio is dynamically balanced within {1e-3, 1e-2, 0.1, 0.5, 1.0}. Federated Communication Parameters: The number of local training epochs (LocalEpochs) executed before each communication by the local client is set to 5, and the maximum number of total federated global communication epochs is set to 100. The test module calculates and outputs the following three ranking evaluation metrics: MRR, HR@10, and NDCG@10, to verify the model's performance gains. Through the above experiments, the present invention can effectively extract global intent in this multi-domain business scenario, and significantly improve the MRR, HR@10 and NDCG@10 indicators while strictly protecting the data privacy of each domain, thus proving the industrial applicability and technological advancement of the present invention.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A privacy-preserving intent-driven federated graph-enhanced cross-domain recommendation method, characterized in that, Includes the following steps: S1. Each domain client encodes the local user-item interaction graph using a graph neural network and extracts a set of local intent prototypes representing the behavior patterns of local user groups through a soft clustering mechanism. S2. Each domain client perturbs the local intent prototype based on an adaptive differential privacy mechanism that is aware of intent popularity, and uploads the noisy intent prototype to the central server. The central server aggregates the noisy intent prototypes uploaded by all clients into a global prototype pool, and performs a clustering algorithm on the global prototype pool to generate a global intent prototype. S3. Each domain client receives the global intent prototype issued by the central server, reconstructs the global perception representation of users and items using the global intent prototype, performs cross-domain potential edge inference based on the global perception representation to generate an enhanced graph structure, optimizes the local recommendation model through dual-view comparison learning, and finally outputs the recommendation result for the target user by the optimized local recommendation model.

2. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 1, characterized in that, S1 specifically includes the following steps: S11. On the m-th local client, initialize the user node embedding representation. and item node embedding representation and define The local intention prototype matrix is ​​composed of learnable intention prototype vectors. ; S12. In the l-th layer of this map neural network, the current representation of the user node is calculated through message passing. and the current representation of the item node The calculation formula is: in, Represents the i-th user The neighborhood group, Represents the j-th item The neighborhood group, They are respectively the ( Layer users With items Embedded representation; S13. Calculate the current representation of the user node. With each local intent prototype The correlation scores between them yield the soft clustering probability distribution. The calculation formula is: S14. Based on the soft clustering probability distribution pair The intent prototypes are weighted and summed to generate an aggregated representation containing higher-order intent information. The calculation formula is: S15. Extracting local graph topological features from graph neural networks. With the aforementioned polymerization characterization Add them together to obtain the final user representation of the l-th layer. The formula is as follows: S16. The local graph recommendation model is trained end-to-end by minimizing the Bayesian personalized ranking loss, and the local intent prototype matrix is ​​optimized during backpropagation. .

3. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 2, characterized in that, The intention popularity-aware adaptive differential privacy mechanism in S2 includes the following steps: S21. Calculate the k-th intent prototype using the soft clustering probability distribution of user nodes on the intent prototype. In local user sets Intent popularity weight and to The relative popularity coefficient is obtained by normalization. The formula is as follows: S22. Based on the relative popularity coefficient, design an adaptive noise variance scheduling function and calculate the specific Gaussian noise standard deviation of the k-th intention prototype. The formula is: in, Based on privacy budget parameters, The sensitivity hyperparameter for controlling noise attenuation rate; S23. Apply an adaptive threshold to perform L2 norm pruning on the intent prototype to obtain the k-th pruned intent prototype. The formula is: in, This is the preset cropping threshold; S24. Inject adaptive Gaussian noise into the trimmed intention prototype to generate the k-th noisy intention prototype under privacy protection. The formula is as follows: in, This represents the noise vector sampled from a multidimensional Gaussian distribution. The adaptive noise standard deviation calculated in S22; S25, Each client sets all K noisy intent prototypes. Uploaded to the central server.

4. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 1, characterized in that, The specific steps of the clustering algorithm for generating global intent prototypes in S2 are as follows: The central server receives all Uploaded by each client These noisy prototypes constitute the global prototype pool. ; In the global prototype pool Clustering is performed, and the objective function that minimizes the clustering error is solved to obtain K global intent prototypes. ; The clustering error is minimized as follows: 。 5. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 1, characterized in that, The S3 method for reconstructing the global perception representation of users and items using the global intent prototype includes the following steps: S31. Obtain the final embedded representation of the local user node after passing through all graph neural network layers, and convert the final layer representation... abbreviated as And the final embedded representation of the local item node, abbreviated as ; S32, Compute user nodes In global intent prototype Attention probability distribution and item nodes In global intent prototype Attention probability distribution The calculation formula is: in, Temperature hyperparameters for controlling the smoothness of attention distribution; S33. Reconstruct the globally perceived user representation based on the attention probability distribution. and global awareness item representation The formula is as follows: 。 6. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 5, characterized in that, Cross-domain potential edge inference and enhanced graph structure generation in S3 includes the following steps: S34. Calculate the cross-domain semantic affinity score between any user-item pair. The formula is as follows: in, , The projection matrix is ​​learnable. For bias terms, This is the mapping weight vector; S35. Add independent and identically distributed Gumbel noise to the affinity score. The Sigmoid function is used for continuous relaxation to calculate the continuous probability of candidate edges being activated. And it is binarized into discrete pseudo-edges using a direct-pass estimator. The formula is: S36. Employ a residual topology fusion mechanism to combine the generated pseudo-edge matrix with the locally observed real discrete adjacency matrix. The images are then merged to generate the final enhanced graph view. The formula is as follows: 。 7. The privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to claim 6, characterized in that, The dual-view comparison learning in S3 includes the following steps: S37. The original local discrete graph... With the generated enhanced graph The original node representations are obtained by inputting them into a shared local graph neural network encoder. ) and enhanced node representation ( ); S38. Calculate the node-level structural contrast loss, using the representation generated from the original graph as the anchor point, the representation generated from the augmented graph as the positive sample, and other nodes of the same type within the same batch as negative samples. The user contrast loss is calculated using this method. Compared with the loss of the items The formula is: The total structural contrast loss is: ; in, For the current training batch, Represents the cosine similarity function. To control the intensity of punishment for difficult negative samples, a contrastive learning temperature parameter was used. S39. Using Bayesian personalized ranking loss The formula for optimizing the local recommendation main task is: S310. Construct the end-to-end joint optimization objective function as follows: in, To compare the learning of weight hyperparameters, The L2 regularization coefficient is... This represents all learnable parameters in the model.

8. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it performs the privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the privacy-preserving intent-driven federated graph enhanced cross-domain recommendation method according to any one of claims 1 to 7 through the computer program.

Images