A social recommendation method and device based on large model enhancement and a graph neural network

Abstract

The application discloses a social recommendation method and device based on a large model enhancement and a graph neural network, comprising: collecting unstructured social corpus of a user, inputting the de-identified unstructured social corpus into a large language model to generate a user comprehensive description; constructing a heterogeneous social graph, encoding the user comprehensive description into a semantic vector as an initial feature of a node of a graph neural network, propagating and iterating on the graph network to output a high-order topological vector; obtaining a recommendation request and a current social context of a target user, recalling a candidate user set meeting a hard constraint condition from all users based on the high-order topological vector; inputting the comprehensive description of the target user and the candidate user and the current context into the large language model to obtain a semantic matching score; deriving a shadow intelligent agent based on the user comprehensive description to obtain an interaction simulation score through virtual dialogue deduction, and calculating a structural similarity score based on the high-order topological vector; and fusing the semantic matching score, the interaction simulation score and the structural similarity score to obtain a final matching score, and outputting a recommendation list after sorting.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a social recommendation method and apparatus based on large model enhancement and graph neural networks. Background Technology

[0002] With the widespread adoption of mobile internet and the continuous diversification of social scenarios, social applications have extensively penetrated into comprehensive platforms encompassing diverse scenarios such as stranger networking, interest groups, local events, academic collaboration, and professional networking. Against this backdrop, intelligent social recommendation systems have become a core functional module for enhancing user activity, promoting high-quality relationship connections, and strengthening platform retention. The core task of social recommendation is to accurately filter potential partners from a massive user base who match the target user's current social intentions—whether it's teammates for a game, conversation partners for a book club, running buddies for the weekend, or companions for academic collaboration or industry exchange—the quality of recommendations directly impacts user experience and the vitality of the platform's social network.

[0003] Most current mainstream recommendation technologies follow collaborative filtering or matrix factorization paradigms, but when faced with complex social semantic understanding and dynamic graph structure modeling, the limitations of their underlying technical architecture are gradually exposed. On the one hand, traditional methods are inadequate in handling large-scale unstructured text generated by users in group chats, dynamics, and private chats, often relying on shallow statistical features or explicit labels, resulting in inaccurate capture of users' deep interests, personality traits, and immediate social intentions. On the other hand, the essence of social relationships is a complex graph structure, and traditional models struggle to effectively capture higher-order neighborhood information and community topological features beyond the second order, limiting the semantic relevance and structural rationality of recommendations. Furthermore, existing systems lack adaptability when facing users' ever-changing social contexts, with recommendation results often being static and undifferentiated black-box outputs, and it is difficult to utilize users' implicit feedback for model iteration in real time. Although large models have shown great potential in natural language understanding and information extraction, and graph neural networks have been fully validated in graph structure representation learning, existing social recommendation systems still face fragmented underlying architectures. Summary of the Invention

[0004] The purpose of this invention is to provide a social recommendation method and apparatus based on large model enhancement and graph neural networks, aiming to solve the technical problems existing in the current social recommendation scheme, such as coarse semantic characterization of user profiles, difficulty in capturing high-order social structures, and lack of context awareness and interpretability.

[0005] To achieve the above objectives, this invention provides a social recommendation method based on large model augmentation and graph neural networks, the method comprising: Unstructured social corpora generated by users on social platforms are collected and de-identified to obtain semantic feature representations. These semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user. A heterogeneous social graph is constructed with users as nodes, friends and interactive behaviors between users as edges. The comprehensive description of users is encoded into semantic vectors as the initial features of nodes in the graph neural network. Message passing and iterative updates are performed on the heterogeneous social graph through the graph neural network, and the high-order topology vectors of each user are output. Receive recommendation requests initiated by target users and obtain the target user's current social context. Based on the target user's high-order topological vector, recall a set of candidate users that meet the preset hard constraints from all users. Input the comprehensive user description of the target user, the current social context, and the comprehensive user description of the candidate users in the candidate user set into the large language model to obtain the semantic matching score; The virtual dialogue simulation is performed by deriving a shadow agent based on the comprehensive user description of the target user and the candidate user, and the interaction simulation score is obtained. The structural similarity score is calculated based on the high-order topological vectors of the target user and the candidate user. The semantic matching score, interaction simulation score, and structural similarity score are combined to calculate the final matching score for the corresponding candidate user. The final matching scores of all candidate users are then sorted to generate a recommendation list, which is then output to the target user.

[0006] To achieve the above objectives, the present invention also provides a social recommendation device based on large model augmentation and graph neural networks, the device comprising: The comprehensive description generation unit is used to collect unstructured social corpora generated by users on social platforms and perform de-identification processing to obtain semantic feature representations. The semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user. The topology vector generation unit is used to construct a heterogeneous social graph with users as nodes, friends and interactive behaviors between users as edges. The comprehensive description of the users is encoded into semantic vectors as the initial features of the nodes of the graph neural network. The graph neural network is used to perform message passing and iterative updates on the heterogeneous social graph, and outputs the high-order topology vectors of each user. The vector recall unit is used to receive recommendation requests initiated by target users and obtain the current social context of target users. Based on the high-order topological vector of target users, it recalls a set of candidate users that meet the preset hard constraints from the full user base. The semantic matching unit is used to input the comprehensive user description of the target user, the current social context, and the comprehensive user descriptions of the candidate users in the candidate user set into the large language model to obtain the semantic matching score. The computing unit is used to derive a shadow agent based on the comprehensive user description of the target user and the candidate user to perform virtual dialogue inference, obtain the interaction simulation score, and calculate the structural similarity score based on the high-order topological vector of the target user and the candidate user. The fusion and sorting unit is used to fuse the semantic matching score, interaction simulation score, and structural similarity score to calculate the final matching score of the corresponding candidate user. The final matching scores of all candidate users are sorted to generate a recommendation list and output it to the target user.

[0007] To achieve the above objectives, the present invention also proposes a social recommendation device based on large model augmentation and graph neural networks, including a processor, a memory, and a computer program stored in the memory. The computer program is executed by the processor to implement the steps of a social recommendation method based on large model augmentation and graph neural networks as described in the above embodiments.

[0008] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of a social recommendation method based on large model enhancement and graph neural networks as described in the above embodiments.

[0009] Beneficial effects: The above solution collects unstructured social corpora from users and generates comprehensive user descriptions using a large language model. It then uses a graph neural network to perform message passing on a heterogeneous social graph to output high-order topological vectors. Finally, through a complete chain of vector recall, large-scale model semantic matching, shadow agent simulation, and structural similarity fusion and ranking, it achieves the collaborative optimization of multi-dimensional semantic profile construction, high-order social structure modeling, and contextualized fine ranking, thereby improving the accuracy, interpretability, and contextual adaptability of recommendation results. Compared to traditional solutions, this approach leverages the semantic understanding capabilities of a large-scale model to complete user profiles from unstructured corpora, expanding the profile dimensions from simple explicit labels to encompass rich semantic information such as interests, preferences, social needs, and communication styles, thus enhancing the completeness of user profiles. Furthermore, by using graph neural networks to propagate and aggregate within the social graph structure, it overcomes the limitation of traditional collaborative filtering, which can only model second-order co-occurrence relationships, thereby enhancing the relevance of recommendations. Simultaneously, the large-scale model performs semantic matching based on the current social context and outputs recommendation reasons, giving the recommendation results context-awareness and interpretability, improving the credibility and conversion rate of recommendations. In addition, by integrating semantic matching scores, interaction simulation scores, and structural similarity scores for comprehensive ranking, it takes into account multiple dimensions such as semantic matching, dynamic interpersonal attraction, and social structural similarity, further improving the accuracy of recommendations.

[0010] By generating a high-dimensional semantic concept graph through concept-level parsing of unstructured social corpora, and applying Laplace noise perturbations to entity slots, location slots, and event slots according to a local differential privacy budget, effective desensitization of sensitive entities such as phone numbers, ID numbers, and addresses is achieved while preserving core social features such as interests, preferences, social needs, and communication styles. This ensures that user corpora undergo privacy protection processing before entering the large model service, achieving explicit decoupling between unstructured text de-identification and high-dimensional semantic extraction. Therefore, it is possible to retain the semantic features required for recommendation while ensuring user privacy and security, achieving an effective balance between privacy protection and recommendation quality, reducing the risk of direct exposure of raw social corpora, and meeting increasingly stringent privacy compliance requirements.

[0011] By constructing a heterogeneous social graph based on friend relationships, group co-occurrence relationships, and diverse interactive behaviors, different initial weights are assigned to different types of social relationships. A comprehensive edge weight update formula integrating interaction frequency, time decay, and interaction quality is established, enabling the graph neural network to perform message passing on a graph structure that accurately reflects the strength and dynamic evolution of real-world social relationships. On the one hand, the differentiated weights of different types of edges (private chat and live chat have higher weights than likes) allow the model to distinguish between strong and weak relationships, improving the accuracy of social structure learning. On the other hand, the dynamic update mechanism of edge weights, which increases with interaction frequency, decays with time intervals, and is adjusted by interaction quality, allows the graph structure to respond to changes in social relationships in real time, enhancing the model's ability to track the dynamic evolution of users' social circles. This provides a more accurate graph structure input for subsequent high-order topological vector learning, thereby improving the accuracy of recall and the relevance of recommendation results.

[0012] By introducing semantic-topological bidirectional spatial alignment constraints during the training phase, the semantic vectors and higher-order topological vectors of the same user are forced to align in the low-dimensional space. This contrastive learning mechanism eliminates the representational bias between the semantic space and the structural space, enabling vector-based similarity retrieval to consider not only the relevance of social structures but also the fit of deep semantics. The bidirectional alignment constraints ensure that the two types of representations maintain consistency in the shared space. On the one hand, this allows graph neural networks to alleviate the structural sparsity problem of cold-start users by leveraging the semantic priors of large models. On the other hand, it allows the semantic representations of large models to be supplemented and enhanced by social graph structural information. The user representations obtained after the two are optimized together have stronger stability and matching robustness in cold-start, weak connection, and cross-circle recommendation scenarios.

[0013] By employing a near-nearest neighbor index for vector retrieval, combined with hard-constraint filtering and a maximum marginal relevance diversity sampling strategy based on community tags for candidate set reorganization, a three-layer candidate set generation mechanism is constructed: vector recall → hard-constraint filtering → diversity sampling. The near-nearest neighbor index ensures retrieval efficiency at the scale of hundreds of millions of users, enabling real-time response in the recall process. Multi-dimensional hard constraints, such as region, age, gender, and blacklists, guarantee the compliance and security of the recall results. Maximum marginal relevance diversity sampling based on community tags improves the community diversity of the recall results while maintaining relevance, preventing the recommendation list from being overly concentrated on a single user type and enriching the coverage of the candidate set. The synergistic effect of these three mechanisms achieves a good balance between efficiency, security, and diversity in the recall stage, providing high-quality candidate input for the subsequent fine-tuning stage, thereby improving the overall recommendation performance.

[0014] Based on a comprehensive description of the target user and candidate users, two temporary role-playing instances are derived to conduct multi-round virtual dialogues. The evaluation model uses a three-dimensional quantitative scoring system, evaluating dialogue resonance, communication fluency, and icebreaker probability, combined with the cosine similarity of higher-order topological vectors as a structural similarity score. This approach simulates the real interaction process between two parties in a target social context through the role-playing capabilities of a large model, enabling the recommendation system to assess dynamic interpersonal attraction that static text matching cannot reflect: topic resonance measures the degree of alignment of interests between the two parties, communication fluency assesses the natural coherence of the dialogue, and icebreaker probability predicts the likelihood of the first contact turning into effective interaction. This approach mines dynamic features at the interaction level through virtual social simulation, making the recommendation results closer to the interactive experience in real social scenarios, effectively improving recommendation accuracy and user satisfaction.

[0015] By collecting explicit and implicit feedback from target users to the recommendation list, mapping these feedbacks into feedback feature vectors, and inputting them into a social cognitive state machine driven by a large language model, the system updates the user's dynamic state vector using state transition equations. Based on this, the system refines the user's comprehensive description, enabling it to keenly capture shifts in user interests and changes in the strength of social intent. This allows for adaptive adjustments to user profiles, ensuring that recommendations consistently align with the user's current needs. Furthermore, the state machine, driven by a large language model, understands the semantic meaning behind the feedback, making state updates interpretable. This avoids noise accumulation and outdated preference residue problems caused by simple numerical summation, giving the recommendation system continuous self-optimization capabilities. Attached Figure Description

[0016] 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 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.

[0017] Figure 1 This is a flowchart illustrating a social recommendation method based on large model enhancement and graph neural networks, provided as an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a social recommendation device based on large model enhancement and graph neural networks, provided as an embodiment of the present invention.

[0019] The realization of the invention's objective, its functional characteristics, and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the 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 scope of protection of the present invention.

[0021] The present invention will be described in detail below with reference to the embodiments.

[0022] Reference Figure 1 The diagram illustrates a flowchart of a social recommendation method based on large model enhancement and graph neural networks according to an embodiment of the present invention. Specifically, the method includes: S11. Collect unstructured social corpora generated by users on social platforms and perform de-identification processing to obtain semantic feature representations. Input the semantic feature representations into a large language model to generate a comprehensive user description for the corresponding user.

[0023] In step S11, the unstructured social corpus generated by users on the social platform is collected and de-identified to obtain semantic feature representations. These semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user, including: S11-1, Perform concept-level parsing on the unstructured social corpus to generate a high-dimensional semantic concept graph. The high-dimensional semantic concept graph is represented by a graph structure. The nodes of the high-dimensional semantic concept graph include user nodes, interest concept nodes, topic nodes, social need nodes, communication style nodes, and topic avoidance nodes. The edges of the high-dimensional semantic concept graph include user-interest edges, user-topic edges, user-social need edges, user-communication style edges, interest-topic co-occurrence edges, and user-topic avoidance edges. S11-2, Apply Laplace noise perturbation to the entity slots, location slots, and event slots in the high-dimensional semantic concept graph according to the local differential privacy budget to obtain the semantic feature representation; S11-3, The semantic feature representation is input into the large language model, and the large language model is driven by the preset prompt word template to extract the profile information of the corresponding user in a targeted manner. The profile information is then compressed to generate the comprehensive description text of the user within a preset word range. The profile information includes demographic attributes, fine-grained interest preferences and representative entities, social needs type, personality and communication style, values ​​and taboo words.

[0024] In this embodiment, unstructured social corpora accumulated by users in social applications are first collected. These corpora include, but are not limited to, group chat snippets, private chat summaries, dynamic posts, comment replies, and personal profiles. To protect user privacy while utilizing a large language model for semantic extraction, a distributed privacy barrier semantic extraction mechanism is deployed on the terminal side or edge agent before the corpus leaves the user's local database or enters the large language model service. Specifically, the edge agent first performs concept-level parsing on the original dialogue text to generate a high-dimensional semantic concept graph, which is structured as a graph G. sem =(V sem E sem ,X sem W sem ) is represented as follows, where V sem A semantic node set, including at least user nodes, interest concept nodes, topic nodes, social need nodes, communication style nodes, and topic avoidance nodes; E sem The set of semantic edges includes at least user-interest edges, user-topic edges, user-social needs edges, user-communication style edges, interest-topic co-occurrence edges, and user-topic avoidance edges; X sem W is a node feature matrix used to record node type, node text, confidence score, source time window, and semantic vector;sem This is the set of edge weights, used to record the confidence or intensity of occurrence of corresponding semantic relations. The dimension of the high-dimensional semantic concept graph is determined by the node feature matrix X. sem And the set of edge weights W sem The node semantic vector dimension d is determined jointly. sem The dimensions can be 384, 768, or 1024, depending on the output dimension of the text encoder used. In addition to the semantic vector, each node also includes structured fields such as node type, confidence level, timestamp, and source type.

[0025] If the deployment environment does not require an explicit graph, the graph can also be compressed into a user semantic vector v. sem =[v profile ,v interest ,v need ,v style ,v avoid The vectors represent demographic attributes, interests, social needs, communication styles, and topics to avoid. Based on this, the edge agent uses a local differential privacy budget. Laplace noise perturbation is applied to the entity slots, location slots, and event slots in the high-dimensional semantic concept graph, i.e. Wherein, Gsem represents the semantic concept graph before de-identification or its equivalent semantic feature representation. Laplace(Δ / ) represents the de-identified semantic representation after perturbation. ) indicates that the scale parameter is Δ / The Laplacian noise, Δ represents the maximum change in the semantic feature extraction function across adjacent input corpora, i.e., the sensitivity. This represents a locally differential privacy budget. The smaller the perturbation, the stronger the privacy protection and the greater the perturbation. The perturbed semantic representation no longer exposes the original chat details, but still retains core social features related to recommendation, such as interests, social needs, communication styles, and topics to avoid, as semantic feature representations.

[0026] The semantic feature representation is input into a large language model (the qwen3.6-27b model can be used in the specific implementation). The hierarchical prompt word template drives the large language model to perform targeted extraction of user corpus. The example prompt word is: "You are a social profile analysis expert. Based on the following user [ID]'s dialogue and dynamic text in the past 30 days, please extract information according to the following template and generate a comprehensive description of no more than 200 words: 1) Demographics: ...; 2) Interests and preferences (including fine-grained entities): ...; 3) Social needs: ...; 4) Communication style: ...; 5) Topics to avoid: ...

[0027] Text: [Specific content]”.

[0028] By using preset prompt word templates to drive a large language model, targeted extraction of user profile information is achieved. This profile information includes five dimensions: demographic attributes, fine-grained interests and representative entities, social needs, personality and communication style, and values ​​and taboo words. The profile information is then compressed to generate a structured comprehensive user description text within a preset word limit (e.g., no more than 200 characters). This achieves explicit decoupling of unstructured text de-identification and high-dimensional semantic extraction at the architectural level, generating a comprehensive user description that can be used for subsequent representation learning, candidate matching, and interpretive output.

[0029] S12, construct a heterogeneous social graph with users as nodes, friends and interactive behaviors between users as edges, encode the comprehensive description of users into semantic vectors as the initial features of nodes in the graph neural network, perform message passing and iterative updates on the heterogeneous social graph through the graph neural network, and output the high-order topology vector of each user.

[0030] Furthermore, in step S12, the construction of a heterogeneous social graph using users as nodes, friend relationships between users, and interactive behaviors as edges, and encoding the comprehensive user description into a semantic vector as the initial feature of the nodes in the graph neural network, includes: S12-1, construct the heterogeneous social graph using users as nodes and friend relationships, group co-occurrence relationships, and interactive behaviors including likes, comments, private chats, and live chats as edges; wherein, the edges of friend relationships are established and assigned initial weights when one-way or two-way following occurs; the edges of group co-occurrence relationships are established when two users are in the same group and meet the activity threshold, and the initial weights are determined by the number of common groups and their activity levels; the edges of interactive behaviors are established and assigned different initial weights when the corresponding behavior occurs for the first time, and the base weights of private chats and live chats are higher than those of likes; all edge weights are dynamically updated as the frequency of interaction increases, the time interval lengthens, and the quality of interaction changes; S12-2, The comprehensive user description text is mapped into a fixed-dimensional dense vector through a semantic vector encoding model, and used as the initial node feature of the corresponding user node, which is then input into the graph neural network.

[0031] Furthermore, all edge weights are dynamically updated as the frequency of interaction increases, the time interval lengthens, and the quality of interaction changes, including: use Perform calculations to obtain the updated edge weights. In the formula, r represents the edge type, and η r This represents the basic weight of edge type r. This represents the cumulative number of interactions between user i and user j under edge type r up to time t. μ represents the time interval from the most recent interaction of the corresponding edge type r. r Indicates the time decay coefficient. Norm represents the interaction quality coefficient, which includes one or more of the following: private chat response rate, average chat rounds, duration of live chat, or number of valid characters in comments. r This indicates normalization over all neighborhood edge weights of the same edge type.

[0032] Furthermore, during the training phase, the graph neural network aligns the semantic vectors and higher-order topological vectors of the same user in a common space through a semantic-topological bidirectional spatial alignment constraint. This semantic-topological bidirectional spatial alignment constraint is implemented using a contrastive learning loss function, which is expressed as follows: Where i represents the user index in the current batch, j represents other user indices in the current batch used to constitute negative samples, and z sem,i z represents the semantic vector obtained by the text encoder from the comprehensive user description of user i. topo,i z represents the higher-order topological vector output by the graph neural network for the same user i. topo,j denoted as the higher-order topological vector of user j, sim(·) represents the similarity function (such as cosine similarity), and τ represents the temperature coefficient; The total training loss of the graph neural network is represented as L. total =L link +λ1*L align +λ2*L aux ;in, L link This indicates that the link predicts the cross-entropy loss for binary classification. In the formula, z i and z j y represents the graph neural network output vector for users i and j. ij= 1 indicates that there is a real friendship between the two users or that there is a valid interaction edge that reaches a threshold, y ij =0 indicates a pair of users with no or weak connections obtained by negative sampling, and σ represents the Sigmoid function; L aux Indicates the loss of auxiliary supervision. In the formula, c represents the category of group or interest community, and y ic Indicates whether user i belongs to category c, p ic λ1 and λ2 represent the probability that the model predicts user i belongs to category c, and λ1 and λ2 represent the corresponding weight coefficients, respectively.

[0033] In this implementation, a heterogeneous social graph is constructed using users as nodes and their friendship relationships and interactive behaviors as edges. Specifically, the heterogeneous social graph is constructed using friendship relationships, group co-occurrence relationships, and interactive behaviors including likes, comments, private chats, and live chat as edges. The edges for friendship relationships are established and assigned initial weights when users follow each other in one-way or two-way; for two-way friendships, the weight is multiplied by the number of mutual relationships. The edges for group co-occurrence relationships are established when two users are in the same group and meet an activity threshold; the initial weights are determined by the number of shared groups and the activity level of those shared groups. The edges for interactive behaviors are established and assigned different initial weights when the corresponding behavior occurs for the first time; the base weights for private chats and live chats are higher than those for likes. All edge weights are dynamically updated as interactions continue to occur. The specific update method is as follows: Thus, edge weights are initialized by relationships or initial actions, subsequently increasing in frequency and decreasing in time intervals, and are modulated by interaction quality. Simultaneously, the comprehensive user description text generated in the preceding steps is mapped into a fixed-dimensional dense vector using a semantic vector encoding model. This vector serves as the initial node feature input to the graph neural network for the corresponding user node, directly mapping semantic features extracted from the large model, such as interests, social needs, communication styles, and avoided topics, to the initial semantic prior for graph representation learning. Based on this, a multi-layer graph neural network architecture combining Graph Attention Network (GAT) and GraphSAGE is used to iteratively update node representations: GAT assigns learnable weights to different neighbors through an attention mechanism, enabling the model to distinguish between strong and weak relationships, and the contributions of homogeneous and heterogeneous neighbors; GraphSAGE's neighbor sampling mechanism ensures scalable training in a social graph with hundreds of millions of nodes.

[0034] During the offline training phase, the system incorporates a semantic-topological bidirectional spatial alignment constraint, enabling the implicit personality traits, social intentions, and communication tendencies extracted by the large model to be jointly optimized with the social hierarchy structure in the graph network within a low-dimensional dense space. For user i, let z be the semantic vector obtained by the large model's comprehensive description of the encoded text. sem,i The higher-order topological vector obtained after iteration by the multi-layer graph neural network is z. topo,i A contrastive learning task is constructed using the two classes of representations of the same user as positive sample pairs and the representations of other users within the batch as negative sample pairs. The contrastive learning loss function is... This loss brings the semantic and topological vectors of the same user closer together, while distinguishing the semantic and topological vectors of different users. The total training loss of the graph neural network is L. total =L link +λ1*L align +λ2*L aux Through the above training process, each user ultimately obtains a dense embedding vector that integrates their own semantics with higher-order social structures, namely a higher-order topological vector, which serves as the basis for similarity calculation in the subsequent recall stage.

[0035] S13: Receive the recommendation request initiated by the target user and obtain the target user's current social context. Based on the target user's high-order topology vector, recall the set of candidate users that meet the preset hard constraints from the full user pool.

[0036] Furthermore, in step S13, the process of recalling a set of candidate users that meet preset hard constraints from all users based on the high-order topology vector of the target user includes: S13-1, construct an approximate nearest neighbor index structure using the high-order topological vectors of all users as index data, and use the high-order topological vectors of the target user as query vectors to search in the approximate nearest neighbor index structure to obtain the top K users with the highest similarity to the high-order topological vectors of the target user as the initial candidate set. S13-2, the initial candidate set is filtered according to the preset hard constraints to obtain a filtered candidate subset, wherein the preset hard constraints include one or more of blacklist filtering, gender preference filtering, age group filtering and geographical range filtering. S13-3, the filtered candidate subset is reorganized based on the maximum marginal relevance diversity sampling strategy of each user's community tags to obtain the candidate user set.

[0037] In this embodiment, a recommendation request initiated by a target user is received, and the target user's current social context description is obtained (e.g., looking for basketball partners in Chaoyang Park on the weekend, 3v3 amateur level). Then, the vector recall stage begins. An approximate nearest neighbor index structure is constructed using the high-order topological vectors of all users as index data. The target user's high-order topological vector is used as the query vector in this approximate nearest neighbor index structure for retrieval. The similarity between the target user's high-order topological vector and each candidate vector is calculated, and the top K users with the highest similarity to the target user's high-order topological vector are retrieved as the initial candidate set. The initial candidate set is filtered according to preset hard constraints, resulting in a filtered candidate subset. These preset hard constraints include one or more of blacklist filtering, gender preference filtering, age group filtering, and geographic range filtering to ensure the compliance of the recall results. Based on this, the filtered candidate subset is reorganized using a maximum marginal relevance diversity sampling strategy based on each user's community tags (derived from the user's comprehensive description of interests and representative entities). This improves the community diversity of the recall results while maintaining relevance, preventing the recommendation list from being overly concentrated on a single type of user. Through the three-layer recall mechanism of vector retrieval → hard constraint filtering → diversity sampling, a candidate user set is finally obtained, which serves as the input for the subsequent fine ranking stage, achieving a balance between retrieval efficiency, compliance and security, and result diversity.

[0038] S14: Input the comprehensive user description of the target user, the current social context, and the comprehensive user descriptions of the candidate users in the candidate user set into the large language model to obtain the semantic matching score.

[0039] In this embodiment, the input to the large model (which acts as the core semantic discrimination unit for semantic refinement, such as the qwen3.6-27b model) includes a comprehensive description of the target user, the target user's current social context, a comprehensive description of the candidate user, and the results of hard constraint filtering; the output is a semantic matching score, a recommendation reason, and a description of the risk of mismatch. For example, the following prompt word template can be used to drive the large model to perform semantic matching discrimination: "Target user profile: [Description A]; Current social context: [Looking for basketball partners in Chaoyang Park this weekend, 3v3 amateur level]; Candidate user profile: [Description B]. Please output a matching score of 0-10 and provide a recommendation reason of no more than 30 characters. If the candidate is clearly mismatched, the score should be below 3 points and the reason should be explained." After performing semantic reasoning and context judgment based on the above input, the large model outputs a semantic matching score. sem Its value ranges from 0 to 10, with higher values ​​indicating a higher degree of semantic matching in the current context. This semantic matching score serves as an input for the subsequent final matching score fusion and ranking, but it does not independently determine the final recommendation result. Simultaneously, the natural language recommendation reasons output can be used to display the recommendation list, thereby enhancing the interpretability of the recommendation results.

[0040] S15: Based on the comprehensive user descriptions of the target user and candidate users, a shadow agent is derived to perform virtual dialogue simulation, and an interaction simulation score is obtained. The structural similarity score is calculated based on the high-order topological vectors of the target user and candidate users.

[0041] Furthermore, in step S15, the virtual dialogue simulation performed by the derived shadow agent based on the comprehensive user descriptions of the target user and candidate users to obtain an interaction simulation score, and the structural similarity score calculated based on the high-order topological vectors of the target user and candidate users, including: S15-1, a first shadow agent is derived based on the comprehensive user description of the target user and a second shadow agent is derived based on the comprehensive user description of the candidate user, wherein the shadow agent is a temporary role-playing instance constructed by the large language model according to the cue word constraint; S15-2, Under the constraints of the current social context, drive the first shadow agent and the second shadow agent to conduct a preset number of virtual dialogues to obtain a virtual dialogue summary; S15-3, input the virtual dialogue summary, the comprehensive user description of the target user, the comprehensive user description of the candidate user, and the current social context into the evaluation model to obtain the topic resonance degree, communication smoothness, and icebreaking probability; S15-4, using Scoreagent =γ1*R topic +γ2*R smooth +γ3*P break The interactive simulation score is calculated, where R... topic R represents the degree of resonance of a topic. smooth P indicates the smoothness of communication. break γ1, γ2 and γ3 represent the icebreaking probability, and γ1, γ2 and γ3 are the corresponding weighting coefficients, satisfying γ1+γ2+γ3=1; S15-5, the cosine similarity is used to calculate the higher-order topological vectors of the target user and the candidate user to obtain the structural similarity score.

[0042] In this embodiment, two subtasks are executed in parallel: shadow agent virtual dialogue deduction and structural similarity calculation. In the shadow agent virtual dialogue deduction, shadow agent A is derived based on the comprehensive user description of the target user, and shadow agent B is derived based on the comprehensive user description of the candidate user. The shadow agents are temporary role-playing instances constructed by the large language model based on cue word constraints. Their roles are determined by their respective interests, social needs, communication styles, avoided topics, and the current social context. The lifespan of a shadow agent is limited to a single candidate evaluation request and is not stored long-term. Then, under the constraints of the current social context, shadow agents A and B are driven to perform a preset number of rounds of virtual dialogue deduction to obtain a virtual dialogue summary.

[0043] Subsequently, the virtual dialogue summary, the comprehensive user description of the target user, the comprehensive user description of the candidate user, the current social context, and the structural similarity features corresponding to the structural similarity score are jointly input into the evaluation model (e.g., using the qwen3.6-27b model). The evaluation model then evaluates the topic resonance degree R of the virtual dialogue. topic Communication fluency R smooth And the icebreaking probability P break Each component is quantitatively scored, resulting in three separate scores. Topic resonance measures the degree of match between the two parties' interests and the current social context; communication fluency measures the consistency of responses, the level of conflict, and the ability to continue the conversation in the virtual dialogue; and icebreaker probability measures the likelihood of the initial contact escalating into effective interaction. Then, the scores are used... agent =γ1*R topic +γ2*R smooth +γ3*P break The interactive simulation score is obtained by performing a weighted summation calculation. agent In terms of structural similarity calculation, cosine similarity is used to calculate the higher-order topological vectors of the target user and the candidate user to obtain the structural similarity score. graphThrough the aforementioned shadow agent virtual dialogue inference and structural similarity calculation, the dynamic interpersonal attraction beyond static text matching is captured, extending the semantic understanding capability of the large model to the interactive inference capability. At the same time, the social structural similarity between users is measured from the graph structure level, providing multi-dimensional matching signals for subsequent fusion ranking.

[0044] S16, the semantic matching score, interaction simulation score and structural similarity score are combined to calculate the final matching score of the corresponding candidate user. The final matching scores of all candidate users are sorted and a recommendation list is generated and output to the target user.

[0045] The final matching score for the corresponding candidate user is calculated by integrating the semantic matching score, interaction simulation score, and structural similarity score, including: Use Score final =ω1*Score sem +ω2*Score agent +ω3*Score graph The final matching score is obtained by calculation, where Score sem Score represents the semantic matching score. agent Score represents the interactive simulation score. graph Let ω1, ω2, and ω3 represent the structural similarity scores, and let ω1+ω2+ω3=1.

[0046] In this embodiment, the semantic matching score, interaction simulation score, and structural similarity score obtained in the preceding steps are weighted and fused to obtain the final matching score for the corresponding candidate user. final Score final =ω1*Score sem +ω2*Score agent +ω3*Score graph Score sem This represents the semantic matching score output by the semantic ranking of the large language model, with a value ranging from 0 to 10. A higher value indicates a higher degree of semantic matching in the current social context. agent The score represents the interaction simulation score obtained by fusing the evaluation model scores after the shadow agent's virtual dialogue inference. graphThis represents the structural similarity score calculated based on the higher-order topological vectors of the target user and candidate users (preferably obtained by normalization after cosine similarity calculation); ω1, ω2, and ω3 are the corresponding weights, satisfying ω1+ω2+ω3=1, and their specific values ​​can be determined through experiments or grid search based on the business scenario. This fusion ranking mechanism complements the semantic reasoning ability of the large model, the ability of shadow agent simulation to evaluate dynamic interpersonal attraction, and the encoding ability of graph neural networks for higher-order social structures, jointly supporting high-precision re-ranking.

[0047] After calculating the final matching score for each candidate user in the candidate user set, all candidate users are sorted from high to low according to their final matching scores. Candidate users with higher final matching scores are ranked higher in the recommendation list. Finally, a recommendation list is generated and output to the target user. The recommendation reasons generated synchronously by the large model can also be attached for the display of recommendation results, thereby realizing context-aware and explainable social partner recommendations.

[0048] In another embodiment, the method further includes: S17, collect explicit and implicit feedback from target users to the recommendation list results, map the explicit and implicit feedback into feedback feature vectors, and input the feedback feature vectors into a social cognitive state machine driven by a large language model, so as to utilize... Update the dynamic state vector representing the current social state of the target user, where s t Let s represent the user's state vector at time t. t 1 represents the state vector of the previous time step, f t A represents the feedback feature vector (obtained by mapping the latest explicit and implicit feedback). t B represents the state preservation matrix (used to control the proportion of historical states retained). t This represents the feedback injection matrix (used to control the intensity of the impact of new feedback on each state slot); the updated dynamic state vector is used to correct the comprehensive user description of the target user.

[0049] In this embodiment, after the recommendation list is output to the target user, a feedback-driven dynamic update of user preferences is performed. Specifically, this involves collecting explicit and implicit feedback from the target user regarding the recommendation list results. Explicit feedback includes agreeing or rejecting friend requests, while implicit feedback includes changes in private chat frequency, chat rounds, and interaction depth. The newly arrived explicit and implicit feedback are then converted into state features f. t Then f t Input a social cognitive state machine driven by a large language model to leverage Update the dynamic state vector s representing the target user's current social state. tThe state vector inherits historical states while incorporating the latest feedback. This social cognitive state machine uses a structured state vector to represent a user's social openness, immediate topic desire, long-term network stability, relationship exploration tendency, and sensitivity to negative feedback. For scenarios such as cooling interactions, decreased private chat frequency, and shortened chat rounds, the system introduces a time decay mechanism to dynamically forget the weights of the corresponding social intentions, specifically w. k (t) = w k(t-1) ·exp(-β k ·δ t ) + u k (t), where w k (t) represents the effective weight of the k-th preference slot at time t, β k δ represents the forgetting coefficient of the k-th slot. t Indicates the time interval since the last relevant feedback, u k (t) represents the incremental signal injected into the k-th slot by the latest feedback. The preference weight decays over time and is re-enhanced when new feedback arrives. The large model reads a formatted JSON state tree or an instant social state vector online, rather than the original behavior flow itself, and uses this to complete the semantic understanding of the user's instant social needs and the next round of recommendation judgment. Finally, the updated dynamic state vector is used to correct the comprehensive user description of the target user, specifically correcting the instant social needs dimension in the comprehensive user description, thereby achieving agile response to short-term context switching and smooth evolution of long-term preferences, enabling the recommendation system to have a closed-loop capability of continuous self-optimization.

[0050] In another embodiment, the method further includes: The high-order topology vector output by the graph neural network, the semantic matching score output by the large language model, and the interaction simulation score output by the shadow agent simulation link are used as joint teacher signals. Based on the joint teacher signals and the distillation loss function, a network architecture including a text encoder, a structural feature encoder, a feature splicing layer, and a multilayer perceptron scoring head is trained to obtain a lightweight student model. The text encoder encodes the comprehensive user description of the target user, the current social context, and the comprehensive user description of the candidate user, outputting a text feature vector. The structural feature encoder encodes the high-order topological vectors of the target user and the candidate user, outputting a structural feature vector. The feature concatenation layer concatenates and fuses the text feature vector, the structural feature vector, and the hard constraint features encoded by preset hard constraint conditions to generate a fused feature. The multilayer perceptron scoring head outputs the matching probability of the candidate user based on the fused feature. The distillation loss function is expressed as: In the formula, h slm h represents the candidate representation of the student model output. gnn p represents the higher-order topological vector output by the graph neural network. teacher p represents the teacher probability distribution obtained after fusing the joint teacher signals. slm denoted by , where y represents the probability distribution of the student model output, y represents the true label obtained from historical interaction or manual annotation, KL(·) represents the KL divergence, CE(·) represents the cross-entropy loss, and α1, α2, and α3 represent the weight coefficients of the structural distillation loss, soft label distillation loss, and true label supervision loss, respectively. During online inference, the target user's comprehensive user description, the current social context, the candidate user's comprehensive user description, the target user's high-order topological vector, the candidate user's high-order topological vector, and the hard constraint features encoded by preset hard constraint conditions are input into the student model. The entropy value of the student model's output distribution is calculated, and C(x)=1 is used. H(p slm The entropy value is normalized using (x) / log(M) to obtain the confidence score, where H(p) slm (x) represents the entropy value of the student model output distribution, and M represents the number of candidate categories or the number of discrete scoring intervals; wherein, the confidence score is used to measure the degree of certainty of the student model for the current matching sample, the higher the confidence score, the higher the certainty of the student model, and the lower the confidence score, the higher the uncertainty. If the confidence score is not lower than the preset routing threshold, the output of the student model is used as the matching score of the candidate user; otherwise, the current sample is routed to the large language model to re-obtain the semantic matching score or to the shadow agent simulation link to re-obtain the interaction simulation score, and the re-obtained semantic matching score or interaction simulation score replaces the corresponding output of the student model, and the final matching score is recalculated; wherein, the preset routing threshold is determined based on the accuracy-efficiency tradeoff of the validation set.

[0051] In this embodiment, the method further includes multi-teacher distillation and dynamic routing steps. During the offline training phase, the high-order topology vector output by the graph neural network, the semantic matching score output by the large language model, and the interaction simulation score output by the shadow agent simulation link are used as joint teacher signals. The network architecture is trained based on the joint teacher signals and the distillation loss function to obtain a lightweight student model (such as the qwen3.5-9b model). This network architecture includes a text encoder, a structural feature encoder, a feature concatenation layer, and a multilayer perceptron scoring head. The distillation loss function is specifically represented as follows: Through the aforementioned distillation training, the student model simultaneously learns the structural similarity ability of graph neural networks and the semantic judgment ability of large model / simulation links. During online inference, the target user's comprehensive user description, the current social context, the candidate user's comprehensive user description, the target user's high-order topological vector, the candidate user's high-order topological vector, and the hard constraint features encoded by preset hard constraint conditions are input into the student model. The entropy value of the student model's output distribution is calculated, specifically... And using C(x)=1 H(p slm The entropy value is normalized using (x) / log(M) to obtain the confidence score C(x). Here, x represents a single target user-candidate user matching sample, and p... m (x) represents the probability of the student model outputting the m-th candidate category or score interval, and logM is used to normalize the maximum entropy.

[0052] The confidence score measures the student model's certainty about the current matching sample. A higher confidence score indicates higher certainty, while a lower confidence score indicates higher uncertainty. If the confidence score C(x) is not lower than the preset routing threshold α, the student model's output is used as the candidate user's matching score. If the confidence score C(x) is lower than the preset routing threshold α, the current sample is routed to the large language model to re-acquire a semantic matching score, or routed to the shadow agent simulation link to re-acquire an interaction simulation score. The re-acquired semantic matching score or interaction simulation score replaces the corresponding output of the student model, and the final matching score is recalculated. The preset routing threshold α is determined based on the accuracy-efficiency tradeoff of the validation set. Through this dynamic routing mechanism, the student model prioritizes mainstream traffic prediction tasks online to reduce latency and computational costs, while long-tail complex samples are triggered by the confidence score to fall back to the large model's fine ranking or the shadow agent simulation link. This maintains the large model's ranking capability while keeping inference costs controllable, ensuring matching quality for cold-start users and complex social intent scenarios.

[0053] Reference Figure 2 The diagram shown is a schematic representation of a social recommendation device based on large model enhancement and graph neural networks according to an embodiment of the present invention.

[0054] In this embodiment, the device 20 includes: The comprehensive description generation unit 21 is used to collect unstructured social corpora generated by users on social platforms and perform de-identification processing to obtain semantic feature representations. The semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user. The topology vector generation unit 22 is used to construct a heterogeneous social graph with users as nodes, friends and interactive behaviors between users as edges, and encodes the comprehensive description of users into semantic vectors as the initial features of nodes in the graph neural network. The graph neural network is used to perform message passing and iterative updates on the heterogeneous social graph, and outputs the high-order topology vectors of each user. Vector recall unit 23 is used to receive recommendation requests initiated by target users and obtain the current social context of target users, and recall a set of candidate users that meet the preset hard constraints from the full set of users based on the high-order topological vector of target users. Semantic matching unit 24 is used to input the comprehensive user description of the target user, the current social context, and the comprehensive user description of the candidate users in the candidate user set into the large language model to obtain a semantic matching score; The computing unit 25 is used to derive a shadow agent based on the comprehensive user description of the target user and the candidate user to perform virtual dialogue inference, obtain the interaction simulation score, and calculate the structural similarity score based on the high-order topological vector of the target user and the candidate user. The fusion and sorting unit 26 is used to fuse the semantic matching score, interaction simulation score and structural similarity score to calculate the final matching score of the corresponding candidate user, sort all candidate users according to their final matching scores, and generate a recommendation list result to output to the target user.

[0055] Each unit module of the device 20 can execute the corresponding steps in the above method embodiment, so the details of each unit module will not be elaborated here. Please refer to the description of the corresponding steps above for details.

[0056] This invention also provides a social recommendation device based on large model augmentation and graph neural networks. The device includes the social recommendation apparatus based on large model augmentation and graph neural networks as described above. The social recommendation apparatus based on large model augmentation and graph neural networks can employ... Figure 2 The structure of the embodiment, correspondingly, can be executed Figure 1 The technical solutions of the method embodiments shown are similar in implementation principle and technical effect. For details, please refer to the relevant records in the above embodiments, which will not be repeated here.

[0057] The device includes: a server, personal computer, smartphone or tablet computer, or other device with image / text processing and computing functions. The device may include components such as memory, processor, input unit, display unit, and power supply.

[0058] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc.; the data storage area can store data created according to the use of the device, etc. In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory can also include a memory controller to provide access to the memory for the processor and input units.

[0059] The input unit can be used to receive input numerical, character, or image information, and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control. Specifically, in addition to a camera, the input unit of this embodiment may also include a touch-sensitive surface (e.g., a touch screen) and other input devices.

[0060] The display unit can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The display unit may include a display panel, optionally configured as an LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar display panel. Furthermore, a touch-sensitive surface may cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to the processor to determine the type of touch event. Subsequently, the processor provides corresponding visual output on the display panel based on the type of touch event.

[0061] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The illustrated social recommendation method is based on large model augmentation and graph neural networks. The computer-readable storage medium can be a read-only memory, a hard disk, or an optical disk, etc.

[0062] This invention also provides a computer program product, including a computer program / instructions, which are loaded and executed by a processor to implement... Figure 1 This paper presents a social recommendation method based on large model enhancement and graph neural networks.

[0063] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the device embodiments, equipment embodiments, and storage medium embodiments, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions in the method embodiments.

[0064] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A social recommendation method based on large model augmentation and graph neural networks, characterized in that, The method includes: Unstructured social corpora generated by users on social platforms are collected and de-identified to obtain semantic feature representations. These semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user. A heterogeneous social graph is constructed with users as nodes, friends and interactive behaviors between users as edges. The comprehensive description of users is encoded into semantic vectors as the initial features of nodes in the graph neural network. Message passing and iterative updates are performed on the heterogeneous social graph through the graph neural network, and the high-order topology vectors of each user are output. Receive recommendation requests initiated by target users and obtain the target user's current social context. Based on the target user's high-order topological vector, recall a set of candidate users that meet the preset hard constraints from all users. Input the comprehensive user description of the target user, the current social context, and the comprehensive user description of the candidate users in the candidate user set into the large language model to obtain the semantic matching score; The virtual dialogue simulation is performed by deriving a shadow agent based on the comprehensive user description of the target user and the candidate user, and the interaction simulation score is obtained. The structural similarity score is calculated based on the high-order topological vectors of the target user and the candidate user. The semantic matching score, interaction simulation score, and structural similarity score are combined to calculate the final matching score for the corresponding candidate user. The final matching scores of all candidate users are then sorted to generate a recommendation list, which is then output to the target user.

2. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The unstructured social corpus generated by users on the social platform is collected and de-identified to obtain semantic feature representations. These semantic feature representations are then input into a large language model to generate a comprehensive user description for each user, including: The unstructured social corpus is subjected to concept-level parsing to generate a high-dimensional semantic concept graph. The high-dimensional semantic concept graph is represented by a graph structure. The nodes of the high-dimensional semantic concept graph include user nodes, interest concept nodes, topic nodes, social need nodes, communication style nodes, and topic avoidance nodes. The edges of the high-dimensional semantic concept graph include user-interest edges, user-topic edges, user-social need edges, user-communication style edges, interest-topic co-occurrence edges, and user-topic avoidance edges. Laplace noise perturbation is applied to the entity slots, location slots, and event slots in the high-dimensional semantic concept graph according to the local differential privacy budget to obtain the semantic feature representation; The semantic feature representation is input into a large language model. The large language model is driven by a preset prompt word template to extract the profile information of the corresponding user. The profile information is then compressed to generate a comprehensive description text of the user within a preset word range. The profile information includes demographic attributes, fine-grained interest preferences and representative entities, social needs type, personality and communication style, values ​​and taboo words.

3. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The process of constructing a heterogeneous social graph using users as nodes, user friendships, and interactive behaviors as edges, and encoding the comprehensive user descriptions into semantic vectors as initial features for the nodes of the graph neural network, includes: The heterogeneous social graph is constructed using users as nodes and friend relationships, group co-occurrence relationships, and interactive behaviors including likes, comments, private chats, and live chats as edges. Friend relationship edges are established and assigned initial weights when users follow each other in one-way or two-way. Group co-occurrence relationship edges are established when two users are in the same group and meet an activity threshold, with initial weights determined by the number of shared groups and their activity levels. Interactive behavior edges are established and assigned different initial weights when the corresponding behavior occurs for the first time, with private chats and live chats having higher base weights than likes. All edge weights are dynamically updated as interaction frequency increases, time intervals lengthen, and interaction quality changes. The comprehensive user description text is mapped into a fixed-dimensional dense vector using a semantic vector encoding model, and this vector is used as the initial node feature of the corresponding user node and input into the graph neural network.

4. The social recommendation method based on large model augmentation and graph neural networks according to claim 3, characterized in that, All edge weights are dynamically updated as the frequency of interaction increases, the time interval lengthens, and the quality of interaction changes, including: use Perform calculations to obtain the updated edge weights. In the formula, r represents the edge type, and η r This represents the basic weight of edge type r. This represents the cumulative number of interactions between user i and user j under edge type r up to time t. μ represents the time interval from the most recent interaction of the corresponding edge type r. r Indicates the time decay coefficient. Norm represents the interaction quality coefficient, which includes one or more of the following: private chat response rate, average chat rounds, duration of live chat, or number of valid characters in comments. r This indicates normalization over all neighborhood edge weights of the same edge type.

5. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The graph neural network, during the training phase, aligns the semantic vectors and higher-order topological vectors of the same user in a common space through a semantic-topological bidirectional spatial alignment constraint. This semantic-topological bidirectional spatial alignment constraint is implemented using a contrastive learning loss function, which is expressed as follows: Where i represents the user index in the current batch, j represents other user indices in the current batch used to constitute negative samples, and z sem,i z represents the semantic vector obtained by the text encoder from the comprehensive user description of user i. topo,i z represents the higher-order topological vector output by the graph neural network for the same user i. topo,j Let represent the higher-order topological vector of user j, sim(·) represent the similarity function, and τ represent the temperature coefficient; The total training loss of the graph neural network is represented as L. total =L link +λ1*L align +λ2*L aux ;in, L link This indicates that the link predicts the cross-entropy loss for binary classification. In the formula, z i and z j y represents the graph neural network output vector for users i and j. ij= 1 indicates that there is a real friendship between the two users or that there is a valid interaction edge that reaches a threshold, y ij =0 indicates a pair of users with no or weak connections obtained by negative sampling, and σ represents the Sigmoid function; L aux Indicates the loss of auxiliary supervision, In the formula, c represents the category of group or interest community, and y ic Indicates whether user i belongs to category c, p ic λ1 and λ2 represent the probability that the model predicts user i belongs to category c, and λ1 and λ2 represent the corresponding weight coefficients, respectively.

6. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The high-order topology vector based on the target user recalls a set of candidate users that meet preset hard constraints from all users, including: An approximate nearest neighbor index structure is constructed using the high-order topological vectors of all users as index data. The high-order topological vector of the target user is used as the query vector to search in the approximate nearest neighbor index structure, and the top K users with the highest similarity to the high-order topological vector of the target user are obtained as the initial candidate set. The initial candidate set is filtered according to the preset hard constraints to obtain a filtered candidate subset, wherein the preset hard constraints include one or more of blacklist filtering, gender preference filtering, age group filtering, and geographical range filtering. The filtered candidate subset is reorganized using a maximum marginal relevance diversity sampling strategy based on each user's community tags to obtain the candidate user set.

7. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The virtual dialogue simulation, derived from the comprehensive user descriptions of the target user and candidate users, yields an interaction simulation score, and a structural similarity score is calculated based on the high-order topological vectors of the target user and candidate users, including: A first shadow agent is derived from the comprehensive user description of the target user, and a second shadow agent is derived from the comprehensive user description of the candidate user. The shadow agent is a temporary role-playing instance constructed by the large language model according to the cue word constraint. Under the constraints of the current social context, the first shadow agent and the second shadow agent are driven to conduct a preset number of virtual dialogues to obtain a virtual dialogue summary; The virtual dialogue summary, the comprehensive user description of the target user, the comprehensive user description of the candidate user, and the current social context are input into the evaluation model to obtain the topic resonance degree, communication smoothness, and icebreaking probability. Use Score agent =γ1*R topic +γ2*R smooth +γ3*P break The interactive simulation score is calculated, where R... topic R represents the degree of resonance of a topic. smooth P indicates the smoothness of communication. break γ1, γ2 and γ3 represent the icebreaking probability, and γ1, γ2 and γ3 are the corresponding weighting coefficients, satisfying γ1+γ2+γ3=1; The structural similarity score is obtained by calculating the higher-order topological vectors of the target user and the candidate user using cosine similarity.

8. The social recommendation method based on large model augmentation and graph neural networks according to claim 1, characterized in that, The method further includes: The system collects explicit and implicit feedback from target users to the recommended list results, maps this feedback into feedback feature vectors, and inputs these feature vectors into a social cognitive state machine driven by a large language model to utilize... Update the dynamic state vector representing the current social state of the target user, where s t Let s represent the user's state vector at time t. t 1 represents the state vector of the previous time step, f t Let A represent the feedback feature vector. t B represents the state preservation matrix. t Represents the feedback injection matrix; The updated dynamic state vector is used to revise the comprehensive user description of the target user.

9. A social recommendation device based on large model augmentation and graph neural networks, characterized in that, The device includes: The comprehensive description generation unit is used to collect unstructured social corpora generated by users on social platforms and perform de-identification processing to obtain semantic feature representations. The semantic feature representations are then input into a large language model to generate a comprehensive user description for the corresponding user. The topology vector generation unit is used to construct a heterogeneous social graph with users as nodes, friends and interactive behaviors between users as edges. The comprehensive description of the users is encoded into semantic vectors as the initial features of the nodes of the graph neural network. The graph neural network is used to perform message passing and iterative updates on the heterogeneous social graph, and outputs the high-order topology vectors of each user. The vector recall unit is used to receive recommendation requests initiated by target users and obtain the current social context of target users. Based on the high-order topological vector of target users, it recalls a set of candidate users that meet the preset hard constraints from the full user base. The semantic matching unit is used to input the comprehensive user description of the target user, the current social context, and the comprehensive user descriptions of the candidate users in the candidate user set into the large language model to obtain the semantic matching score. The computing unit is used to derive a shadow agent based on the comprehensive user description of the target user and the candidate user to perform virtual dialogue inference, obtain the interaction simulation score, and calculate the structural similarity score based on the high-order topological vector of the target user and the candidate user. The fusion and sorting unit is used to fuse the semantic matching score, interaction simulation score, and structural similarity score to calculate the final matching score of the corresponding candidate user. The final matching scores of all candidate users are sorted to generate a recommendation list and output it to the target user.

10. A social recommendation device based on large model augmentation and graph neural networks, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory, wherein the computer program, when executed by the processor, implements the steps of a social recommendation method based on large model augmentation and graph neural networks as described in any one of claims 1 to 8.

Images