Federal map neural network-based intention perception fair recommendation method

Abstract

The invention discloses an intention perception fair recommendation method based on a federated map neural network, and relates to the technical field of recommendation systems and federated map neural networks, in particular to the intention perception fair recommendation method based on the federated map neural network. According to the technology, sequence modeling is performed on a user behavior sequence, a global intention center is generated in combination with K-means clustering, and cross-client semantic similarity driven user embedding extension is realized, so that neighbor discovery of potential intention perception is completed, and the problems of cold start and data sparsity are effectively relieved. Under a federated learning framework, a joint optimization mechanism fusing double-layer comparative learning, sorting optimization and fairness constraint is constructed, and differential privacy noise is introduced to enhance privacy protection capability. The method supports a cooperative working mode of client local training and server security aggregation, and cooperative improvement of recommendation accuracy and group fairness is realized while data privacy and system expandability are guaranteed. Experimental results show that compared with an existing fair recommendation model, the method has better fairness and recommendation performance.

Description

Technical Field

[0001] This invention relates to the fields of recommender systems and privacy protection technology, specifically an intent-aware fair recommendation method based on federated graph neural networks. Background Technology

[0002] In an era of information overload, recommender systems, by analyzing user behavior and preferences to provide personalized product and service recommendations, have become a crucial bridge connecting users and merchants. However, traditional recommender systems rely on centralized data storage and processing models, which, in the context of surging data volumes, not only face the dilemma of data silos but also pose serious privacy risks. Furthermore, these methods often neglect to delve into the deeper intentions of users, making it difficult to accurately match users' true needs, and they suffer from significant deficiencies in fairness, ultimately severely impacting user experience and platform stickiness. To alleviate these problems, many researchers have introduced federated learning into social recommendation tasks, proposing a federated social recommendation framework combining Graph Neural Networks (GNNs). Through distributed model training that does not share raw data, privacy risks are effectively reduced. However, existing recommender systems still face multiple challenges, including security, fairness, and cold start. Among these, fairness and cold start are particularly prominent. On the one hand, differences in user behavior and biased data distribution lead to systemic unfairness between different groups, such as active and inactive users, resulting in significant differences in recommendation accuracy and uneven exposure opportunities, which seriously undermines the fairness of the system. On the other hand, the data silo effect in a federated environment further exacerbates the cold start problem. New users and long-tail users who lack historical behavioral data find it difficult to obtain high-quality personalized recommendations. Traditional collaborative filtering methods have limited performance in data-sparse scenarios. Although some studies have attempted to improve the situation by introducing fairness constraints or utilizing auxiliary information, they have still failed to fully explore the core factor of users' potential intentions—characterizing interests and preferences and their evolution—to achieve accurate recommendations.

[0003] With the explosive growth of internet information, recommender systems have become an important technical means to alleviate information overload and improve user experience, and are widely used in e-commerce, content recommendation, and online services. While existing recommender methods based on federated graph neural networks have made some progress in privacy protection and cross-platform collaboration, they still have significant shortcomings: First, these methods often rely on statistical co-occurrence relationships of user-item interactions for neighbor expansion, lacking effective mining of users' deep semantic intent and interest evolution characteristics. This leads to persistent cold-start problems for new users or long-tail items in data-sparse scenarios, resulting in a significant decline in recommendation performance. Second, existing methods generally lack systematic modeling of user group differences and sensitive attributes, and lack explicit fairness constraint mechanisms. During federated training, this can easily amplify the difference in recommendation quality between active and inactive users, causing cross-group fairness imbalances, which in turn affect the overall fairness and long-term sustainable development of the recommender system. Therefore, there is an urgent need for an innovative recommender method within the federated learning framework that can simultaneously mine users' latent intent, alleviate the cold-start problem, and balance recommendation accuracy and group fairness. Summary of the Invention

[0004] The purpose of this invention is to provide an intent-aware fair recommendation method based on federated graph neural networks. This method aims to address the problems in existing federated recommendation systems, such as over-reliance on statistical co-occurrence relationships, difficulty in characterizing deep semantic intent of users leading to significant performance degradation in data-sparse and cold-start scenarios, and the lack of group fairness constraints resulting in imbalanced recommendation quality among different user groups. This invention integrates sequence intent modeling, private graph expansion, and fairness optimization mechanisms within a federated learning framework, achieving a synergistic improvement in recommendation accuracy, cold-start mitigation capabilities, and group fairness without disclosing original user data.

[0005] This invention is achieved through the following technical solution:

[0006] An intent-aware fair recommendation method based on federated graph neural networks, the technique comprising:

[0007] Step 1: Preprocess and construct user profiles from raw user behavior data. Convert raw rating data and interaction records into a standardized data format suitable for federated learning, and clean outlier and missing data. Construct user profiles based on users' historical interaction frequency and activity characteristics, group users and segment clients, and generate user interaction sequences to provide basic data support for subsequent model training and fairness evaluation.

[0008] Step 2: Construct a federated graph neural network recommendation model for distributed privacy-constrained scenarios. Through hierarchical embedding representation and message passing mechanisms, user nodes and task nodes (or item nodes) are jointly modeled. This achieves model parameter synchronization and unified management within the federated environment without sharing original user data, providing a model foundation for subsequent intent modeling and graph structure expansion.

[0009] Step 3: Introduce the Intent-aware Neighbor EmbeddingExtension (INEE) module to perform intent modeling and cluster analysis on user behavior sequences, extract the potential semantic intent of users in the process of selecting crowdsourcing tasks; construct semantic-level neighbor relationships between different users based on intent similarity, realize cross-client neighbor expansion, thereby enhancing the user-task interaction graph structure and alleviating the problems of data sparsity and cold start.

[0010] Step 4: Introduce multiple privacy protection and secure communication mechanisms during model training and information interaction. RSA digital signatures and encryption technology are used to securely protect the data identifiers transmitted across clients. A secure aggregation protocol is used to reliably summarize model parameters. Differential privacy noise is introduced during the local model update phase on the client side to prevent the inference of sensitive user information from model parameters, thus ensuring the security and privacy of the federated training process.

[0011] Step 5: Construct a multi-objective joint optimization training mechanism. The task matching prediction loss, ranking optimization loss, instance-level and intent-level two-layer contrastive learning loss, and group fairness constraints are modeled in a unified manner, and the model is trained through a joint optimization strategy. This improves the accuracy, robustness, and fairness among different user groups while protecting user privacy.

[0012] The beneficial effects of this invention are: it achieves efficient personalized recommendations while protecting user data privacy, and simultaneously promotes in-depth mining of intent perception, technical protection of privacy and security, and system optimization of recommendation fairness. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the intent-aware fair recommendation method based on federated graph neural networks provided in an embodiment of the present invention.

[0015] Figure 2 This is a model training framework diagram of the intent-aware fair recommendation method based on federated graph neural networks provided in this embodiment of the invention;

[0016] Figure 3 This is a performance comparison analysis of the embodiments of the present invention and existing designs on the ML-100K dataset as training cycles;

[0017] Figure 4 This is a performance comparison analysis of the embodiments of the present invention and existing designs on the Amazon Movies dataset as training cycles;

[0018] Figure 5 This is a performance comparison analysis of the embodiments of the present invention and existing designs on the ML-1M dataset as training cycles change;

[0019] Figure 6 This is an ablation experiment provided by an embodiment of the present invention: to study the impact of different components on recommendation performance and fairness on an Amazon dataset. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0021] like Figure 1 , Figure 2 As shown, the intent-aware fair recommendation method based on federated graph neural networks provided in this example includes the following steps:

[0022] Step 1: To construct a dataset with fairness awareness and reduce noise interference, this invention first preprocesses the raw data and constructs user profiles. Specifically, the rating data is converted into a standardized format suitable for federated learning, and user profiles are generated based on user activity to support subsequent fairness assessment and optimization. This module covers steps such as data cleaning, user grouping, client sampling, and interaction sequence generation.

[0023] In this step, the raw TSV logs are first cleaned and standardized, converted into a four-tuple format (UserID, ItemID, Rating, Timestamp), and only the Top-K popular items are retained to alleviate data sparsity. Then, based on the average number of user interactions... Users are divided into active and inactive user groups, as defined below:

[0024]

[0025]

[0026] in Indicates user The historical interaction set is used as a sensitive attribute in the fairness analysis. To simulate the partial participation characteristics in federated learning, users are randomly shuffled and sampled under a fixed random seed to construct training clients, while independent test sets are constructed for both user groups. Finally, each user's interaction records are sorted chronologically and truncated to the 20 most recent items to generate a compact behavioral sequence that reflects cold-start behavior while preserving temporal dynamics.

[0027] Step 2: This module aims to build an efficient graph neural network architecture for distributed environments. It adopts a hierarchical embedding learning strategy to complete the representation learning of users and items while strictly protecting user privacy. At the same time, it supports unified model parameter management in a federated environment, laying a solid foundation for the implementation of subsequent intent modeling and privacy protection mechanisms.

[0028] In this step, our solution uses a Graph Attention Network (GAT) to achieve weighted aggregation of neighbor information. Specifically, it utilizes user historical interaction records and neighbor node information, first concatenating historical interaction item embeddings with neighbor node embeddings, then calculating attention scores using the LeakyReLU activation function and a trainable weight matrix. After removing invalid neighbors using a masking mechanism, attention weights are obtained through softmax normalization, ultimately achieving dynamic weighted aggregation of neighbor information, effectively improving the discriminativeness and generalization ability of the representation. In each round of federated training, the client first encrypts the item IDs of the user's historical interactions using RSA signature and uploads them to the server. After receiving the signature, the server constructs an item-user index table based on this information, retrieves and returns user-similar embedding representations including neighbor embeddings and intent-based embeddings. The client then trains the local GAT model based on this representation, uploads the encrypted gradients to the server for secure aggregation and parameter updates, and finally, the server distributes the aggregated and optimized global model parameters to each client, completing a single round of federated training.

[0029] Step 3: The Intent-Based Neighbor Embedding Extension Module (INEE) is introduced to deeply fuse explicit interaction signals with implicit semantic signals, thereby significantly alleviating the data sparsity problem in the cold start scenario and improving the accuracy and fairness of recommendations.

[0030] In this step, the INEE module achieves the transition from explicit interaction to implicit intent through three steps: sequence intent encoding, dynamic intent clustering, and intent-driven neighbor enhancement.

[0031] Sequence intent encoding: On the client side, a self-attention-based SASRec model is first used to encode the user's local interaction sequences. Encode the data to obtain a time-series-based vector representation E, and then add the corresponding positional encoding P to form the initial input matrix:

[0032]

[0033] Sequence through Layered Transformer encoder to capture global dependencies:

[0034]

[0035] The multi-head self-attention mechanism is as follows:

[0036]

[0037]

[0038]

[0039] No. The hidden state vector at each position is denoted as . , The time step of the sequence. Indicates user The sequence length, ultimately, after After the layer encoder, the sequence output is pooled to obtain the user's intent representation vector. :

[0040]

[0041] .

[0042] Dynamic intent clustering: Intent vectors generated locally by each client. After being uploaded to the server via a privacy protection mechanism, the server uses K-means to adaptively cluster cross-client intent vectors to mine the global latent intent space.

[0043]

[0044] in For the first Each intent center is obtained through iterative updates. To reduce communication and computational overhead, the server updates the cluster centers every τ rounds, ensuring that the intent space dynamically adapts to changes in data distribution.

[0045] Intent-driven neighbor enhancement: On the client side, dynamic discovery and embedding fusion of soft neighbors across clients are achieved through intent similarity metrics, thereby overlaying a semantic-level implicit neighbor structure on the explicit graph to compensate for the insufficient expressive power of the user-interaction graph when data is sparse. Firstly, for users... Calculate its intended representation With global intent center Cosine similarity:

[0046]

[0047] Take the first one with the highest similarity Intent clusters: ,in This indicates sorting by similarity score in descending order. For users The selected set of intent clusters, whose cluster labels are denoted as For each cluster Sampling at most within this cluster Each user is a semantic neighbor:

[0048]

[0049] While preserving the privacy protection features of private graph extensions, this module introduces globally transferable semantic signals, deepening the neighbor extension from the semantic level of "interaction subject" relationships to the semantic level of "interaction motivation". This significantly improves the overall performance of federated recommendation systems in terms of fairness, robustness and generalization ability, and provides a richer representational foundation for subsequent graph neural network reasoning.

[0050] Step 4: In the privacy protection and fairness assurance stage, this invention is implemented from three aspects: secure communication, privacy protection during model training, and fairness constraints.

[0051] In this step, during the data identification and transmission phase, the client first generates a public-private key pair using the RSA asymmetric encryption algorithm. It then uses a unified public key to digitally sign and encrypt the user's historical interaction item IDs, constructing an encrypted mapping dictionary to achieve a two-way mapping between the encrypted signature and the original item ID. The encrypted signature is uploaded to the server as a unique identifier. The server uses this signature to construct a global cipher2userid inverted index to obtain the user's neighbor information and securely distributes it to the client in dictionary form. Throughout this process, the user's original interaction data remains encrypted, ensuring that the server cannot infer the true item information.

[0052] Secondly, during the model training phase, the system introduces differential privacy technology to perturb the user gradient, preventing the server from deducing user features from the aggregated gradient. The client computes the gradient locally. Then, proceed first. Paradigm clipping limits the gradient to a threshold. Then, Laplace noise is injected into the cropping result: .in, For differential privacy budgeting, the smaller the value, the greater the noise and the stronger the privacy protection.

[0053] To maintain fairness in recommendations while protecting privacy, the system further considers group differences based on user-sensitive attributes (such as activity level). Let the user set be divided into active user sets based on activity level. With inactive user group , Let be the root mean square error of user u, where a lower value indicates a more accurate prediction. Then, the prediction accuracy for a single user u in the current batch is denoted as: Let the group average performance be defined as... The measure of fairness difference between the two groups is:

[0054]

[0055]

[0056] To prevent the disclosure of individual user information in fairness statistics, the system calculates... and Gaussian noise is introduced at this time:

[0057]

[0058] in For each user's loss in the current batch, Given the standard deviation of the Gaussian noise (used to control noise level), the fairness regularization term is finally defined as: .in Control the weight of fairness constraints in the total loss. Adjust the penalty strength for group performance differences (set to 1 here). By... By jointly optimizing the main task loss, the system can significantly improve the performance balance among groups while improving recommendation accuracy.

[0059] Step 5: Regarding ranking optimization, to improve the quality of Top-K recommendations, the Bayesian Personalized Ranking (BPR) framework is introduced. For any user... Sample positive samples from its set of interactive items. and negative samples Construct triples , For the preference scores predicted by the model, the BPR loss is defined as: .

[0060] In the representation learning part, while a single user view can enhance personalized representations, it is insufficient to suppress group differences. Therefore, this study introduces multi-view contrastive learning, combining instance-level and intent-level losses to jointly optimize the discriminativeness and robustness of representations at both the individual and semantic levels. First, instance-level contrastive learning employs InfoNCE loss, generating two embedded views of user behavior through random perturbations and view transformations. and ( This brings the representational distance between different views for the same user closer together.

[0061]

[0062] in Here, sim(⋅,⋅) represents the cosine similarity, where is the temperature parameter. Secondly, intent-level contrastive learning utilizes intent labels obtained through K-means clustering. Optimize representation consistency for users within the same cluster:

[0063]

[0064] Instance-level contrastive loss Loss compared to intention level The weighted summation yields the final comparative learning objective:

[0065]

[0066] in The relative weights of the two contrastive losses are controlled. To ensure training stability, the contrastive learning module is not involved in the model warm-up phase (approximately the first 5 rounds) and is only introduced after convergence.

[0067] Regarding fairness constraints, as can be seen from the privacy protection and secure communication module, the definition of fairness loss is... for:

[0068]

[0069] Meanwhile, an adaptive adjustment coefficient is introduced under the differential privacy framework. and Dynamically scale the gradient for different users to reduce the gap in recommendation quality between groups:

[0070]

[0071] Finally, the jointly optimized quaternary loss function is:

[0072]

[0073] in, Based on the score, predict loss. and These are the weight coefficients for BPR and contrastive learning loss, respectively. During optimization, the SGD optimizer is employed, combined with gradient pruning and differential privacy noise injection mechanisms to achieve a synergistic improvement in accuracy and fairness, significantly enhancing the model's generalization and robustness.

[0074] Depend on Figure 3 As can be seen, on the ML-100K dataset, the loss curve of our proposed model decreases more rapidly during both the training and validation phases, and the final convergence value is significantly lower than that of the existing F²PGNN design: the training loss decreases from 0.052 to 0.050, and the validation loss decreases from 0.0575 to 0.0550; in the testing phase, the RMSE decreases from 1.24 to 1.19, with an error reduction of approximately 4.0% and smaller fluctuations, indicating that the model can achieve efficient fitting and stable prediction in small-scale data scenarios. Figure 4 As shown, on the Amazon Movies dataset, our proposed model also exhibits faster convergence speed and lower loss value, with the RMSE decreasing from 1.20 to 1.16 compared to the existing design F²PGNN, representing an improvement of approximately 3.3%. Especially in this scenario with high data sparsity, the loss curve of our method continues to decline steadily, fully demonstrating its robustness and generalization ability in sparse data environments. Figure 5 It can be inferred that on the ML-1M dataset, our proposed model maintains a stable advantage in large-scale data scenarios: both training and validation losses are reduced compared to the baseline, and the final test RMSE decreases from 1.22 to 1.18, an improvement of approximately 3.2%. More importantly, during the training process from the 10th to the 50th epoch, the curve remains smooth and oscillating, avoiding the fairness constraints inherent in F²PGNN. Overfitting phenomena observed in changing or complex scenarios validate the robustness and cross-domain generalization ability of the proposed architecture. In summary, the proposed method significantly outperforms existing state-of-the-art baselines in key metrics such as convergence speed, final loss, prediction accuracy, and robustness, providing a new paradigm that is accurate, robust, and easy to tune for practical deployment of recommendation systems.

[0075] like Figure 6As shown, due to computational resource constraints, we conducted ablation studies on a medium-sized Amazon dataset. Removing the contrastive learning module (Test 1), the intent clustering component (Test 2), differential privacy noise (Test 3), and the fairness regularizer (Test 4) all resulted in a consistent decrease in model performance. Among these, the lack of intent clustering caused the most significant drop in accuracy, followed by the removal of contrastive learning. The RMSE gap between active and inactive users also widened significantly, indicating that intent semantics and contrastive alignment are crucial for building accurate and sparsity-resistant representations. In contrast, removing only the fairness regularizer or differential privacy noise caused only a slight loss in accuracy, suggesting that the injected noise was well calibrated and that the fairness constraint maintained model utility with minimal overhead. From a fairness perspective, the complete model achieved the lowest validation set dissimilarity, while removing the fairness regularizer significantly increased inter-group dissimilarity. Notably, discarding intent clustering or contrastive learning also increased dissimilarity, revealing their implicit role in stabilizing predictions across different user groups. Overall, intent clustering delivers the greatest accuracy gain; contrastive learning enhances representation robustness while supporting fairness; differential privacy preserves model utility under noise injection; and the fairness regularizer directly suppresses group-level bias. These components work together to achieve a good balance between accuracy, fairness, and privacy protection.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for intent-aware fair recommendation based on federated graph neural networks, characterized in that, Comprise the following steps: Step 1, preprocessing of original user behavior data and user portrait construction; the original score data and interaction record are converted into a standardized data format suitable for federated learning, and the abnormal data and missing data are cleaned up; based on the user historical interaction frequency and activity characteristics, the user portrait is constructed, the user is grouped and the client is divided, and the user interaction sequence is generated, which provides basic data support for subsequent model training and fairness evaluation; Step 2, building a federated graph neural network recommendation model for distributed privacy restricted scenarios, through hierarchical embedding representation and message passing mechanism, jointly modeling user nodes and task nodes (or item nodes), realizing model parameter synchronization and unified management in federated environment without sharing user original data, providing model basis for subsequent intent modeling and graph structure expansion; Step 3, introducing intent-aware neighbor embedding extension module (INEE), modeling and clustering analysis of user behavior sequence, extracting the potential semantic intent of users in the process of selecting crowdsourcing tasks; Based on the similarity of intent, the semantic level neighbor relationship between different users is constructed, realizing neighbor expansion across clients, so as to enhance the user-task interaction graph structure and alleviate the problem of data sparsity and cold start; Step 4, introducing multiple privacy protection and secure communication mechanisms in the process of model training and information interaction; through RSA digital signature and encryption technology, the data identification transmitted across clients is safely protected, the model parameters are reliably aggregated by combining secure aggregation protocol, and differential privacy noise is introduced in the local model updating stage of client, preventing the user sensitive information from being inferred through model parameters, and ensuring the security and privacy of federated training process; Step 5, building a multi-objective joint optimization training mechanism; task matching prediction loss, ranking optimization loss, instance level and intent level double-layer contrastive learning loss and group fairness constraint are modeled, and the model training is completed through joint optimization strategy, which improves the accuracy, robustness and fairness of different user groups in the premise of protecting user privacy.

2. The federated graph neural network based intent-aware fair recommendation method according to claim 1, characterized in that, The data preprocessing in step 1 and the user portrait construction include: cleaning and standardizing the original score data and the interaction record, and processing the abnormal values and the missing values; generating a user portrait based on the user historical interaction frequency and the activity degree characteristics, and dividing the users into an active user group and an inactive user group to form a user behavior sequence arranged in time sequence, thereby providing basic data for subsequent model training and fairness evaluation; the user grouping is to divide the user set U into an active user group and an inactive user group according to the historical interaction times of the users , and the definition is as follows: where, denotes the set of historical interactions of user u, and the partitioning results will be used as sensitive attributes for fairness evaluation.

3. The federated graph neural network based intent-aware fair recommendation method according to claim 1, characterized in that, The construction of federated graph neural network model in step 2 includes: jointly modeling user nodes and item nodes through hierarchical embedding representation and message passing mechanism; in each round of federated training, the client encrypts the item ID of user historical interaction by RSA signature, and uploads the encrypted information to the server; the server receives and constructs the item-user index table, retrieves and returns the embedding representation similar to the user, including neighbor embedding and intent-based embedding, so as to realize model parameter synchronization and unified management in federated environment under the premise of protecting user privacy, and provide basis for subsequent intent modeling and neighbor expansion.

4. The federated graph neural network based intent-aware fair recommendation method according to claim 1, characterized in that, The IEEE module introduced in step 3 is a three-step realization of the transition from explicit interaction to implicit intent through sequence intent encoding, dynamic intent clustering and intent-driven neighbor enhancement; in this step, sequence intent encoding uses the SASRec model based on the self-attention mechanism to encode the user's local interaction sequence into an intent vector, and adds position encoding to retain the timing information; dynamic intent clustering uses the K-means algorithm on the server side to generate global intent centers , where is the th intent center, and is updated once every τ rounds of iteration to adapt to changes in data distribution; intent-driven neighbor expansion selects the top most similar clusters by calculating the cosine similarity between the user intent vector and the global intent center , and samples users from each cluster as semantic neighbors to form implicit neighbor embeddings.

5. The federated graph neural network based intent-aware fair recommendation method according to claim 1, characterized in that, The multiple privacy protection and secure communication mechanism described in step 4 includes: using RSA digital signature and encryption technology to protect the data identification transmitted across the client to prevent data tampering and illegal access; combining a secure aggregation protocol to encrypt and aggregate the model parameters uploaded by each client to ensure that the server cannot obtain the original parameter information of a single client; and introducing gradient clipping and differential privacy noise in the local model updating process of the client to prevent attackers from inferring sensitive data of the user through model parameters or gradient information; specifically, after the client calculates the gradient g locally, it first clips the gradient within a threshold norm, limits the gradient to a threshold , then injects Laplace noise into the clipping result to obtain the perturbed gradient: wherein To prevent the leakage of individual user's performance information in the fairness evaluation process, the system introduces Gaussian noise when calculating the average prediction performance of different user groups in the form of: wherein represents the loss value of the user in the current batch, is the standard deviation of the Gaussian noise, used to control the noise intensity.

6. The federated graph neural network based intent-aware fair recommendation method according to claim 1, characterized in that, In step 5, the recommendation fairness and multi-objective joint optimization training including double-layer contrastive learning view are introduced; The fairness of a recommendation system aims to ensure that different user groups enjoy the same quality of recommendation service without systematic bias caused by sensitive attributes such as gender, age, and user activity. Our invention focuses on the most popular research paradigm of group fairness, with the core goal of minimizing the performance difference between different groups. Specifically, assuming that a user set is divided into mutually disjoint groups based on a sensitive attribute, the average recommendation performance of the th group is defined as: wherein, represents a recommendation performance indicator of the user ; then, the group fairness at the system level can be quantified by the maximum group difference: Obviously, The smaller the difference, the closer the recommendation quality of each group, and the higher the group fairness of the system. To explicitly control this difference in the model training phase, we introduce a fairness regularization term: wherein the hyperparameters controlling the weight of the fairness constraint in the overall loss, adjusting the punishment intensity of the difference value; combining the regular term with the main task loss for optimization, so as to effectively suppress the performance deviation between groups while maintaining the overall recommendation performance, and realize the win-win of "performance-fairness"; in the fairness multi-task optimization module of the application, a four-element loss joint optimization framework is innovatively constructed, score prediction, ranking optimization, double-layer contrast learning and fairness constraint are fused in the federated learning paradigm, and the performance and fairness are synergistically enhanced. In terms of ranking optimization, to improve the quality of Top-K recommendations, the Bayesian Personalized Ranking (BPR) framework is introduced; for any user Sample positive samples from its set of interactive items. and negative samples Construct triples BPR loss is defined as: wherein, is the predicted preference score for the model, is a sigmoid function; in the representation learning part, although the single user view can enhance the personalized representation, it is not enough to suppress the group difference; therefore, this study introduces multi-view contrastive learning, combining instance-level and intent-level losses to jointly optimize the discriminability and robustness of the representation at the individual and semantic levels; first, the instance-level contrastive learning optimizes the consistency of the same user's different view representations using the InfoNCE loss, and the calculation formula is as follows: The intent level contrastive learning optimizes the consistency of the same cluster user representation by K-means clustering label: and the two are weighted and summed to form the final contrastive learning objective ; group fairness regularizer defined as where is a weight, is a penalty exponent; further obtaining a jointly optimized total loss function, defined as: wherein , , BPR loss, contrastive learning loss and weight coefficient of fairness constraint, respectively; in order to ensure the stability of training, the contrastive learning module does not participate in the model warm-up stage (about the first 5 rounds), and is introduced only after convergence.

Images