Article recommendation method and system based on big language model enhanced graph representation learning

Abstract

The invention belongs to the technical field of article recommendation, and provides an article recommendation method and system based on big language model enhanced graph representation learning, and the method comprises the steps: obtaining a user-article graph recommendation data set; extracting semantic knowledge features of the obtained graph recommendation data set; according to the extracted semantic knowledge features and an article recommendation model, completing article recommendation learning; wherein the article recommendation model adopts a big language model to enhance a graph representation learning model, embeds semantic similarity through the big language model, adopts an adaptive graph structure learning mechanism to identify a semantic boundary so as to determine structure information, and performs bidirectional knowledge interaction transmission on semantic and structure information; interactive noise is filtered and relieved in combination with information bottleneck regularization, optimization of an article recommendation model is carried out with the purpose of minimizing a comprehensive multi-objective loss function, and article recommendation is completed by calculating preference scores of articles.

Description

Technical Field

[0001] This invention belongs to the field of item recommendation technology, specifically relating to an item recommendation method and system based on large language model-enhanced graph representation learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In an increasingly vast information environment, recommender systems have become a key component of modern digital platforms, connecting users with relevant content. Graph-based methods capture the complex relationships between users and items through collaborative signals, demonstrating significant effectiveness in recommender systems. The integration of graph neural networks (GNNs) with recommender frameworks can bring significant performance improvements. Architectures such as NGCF and Light GCN have established new benchmarks by effectively utilizing the inherent structural patterns in user-item interaction data.

[0004] However, current GNN-based recommender systems face significant limitations restricting their practical effectiveness. Recommendation models primarily operate based on ID information and structural patterns, neglecting the rich textual semantics associated with users and items. This results in representations that fail to fully capture the nuanced semantic features behind user preferences. Furthermore, real-world interaction data inevitably contains noise from accidental clicks, information bias, or popularity-driven behavior. Traditional GNN architectures treat these signals indiscriminately, giving equal importance to all interactions and lacking mechanisms to identify potentially misleading signals. This vulnerability becomes particularly pronounced in sparse data scenarios, where noise can severely distort the learned representations. While self-supervised learning methods attempt to address these issues through contrastive learning strategies, they are fundamentally limited by their reliance on interaction data without incorporating external semantic knowledge.

[0005] Large Language Models (LLMs) can address the semantic limitations of GNN-based recommenders. For example, state-of-the-art LLMs such as GPT-4 and LLaMA demonstrate exceptional capabilities in semantic understanding and knowledge representation, making them valuable supplements to recommender systems. Integrating LLMs into recommender frameworks generates rich textual profiles for users and items through cue engineering methods. However, existing LLM-integrated recommender methods incur significant computational costs that hinder real-time deployment on a large scale, are susceptible to illusion problems, and may generate recommendations for non-existent items.

[0006] In addition, current integrated advancement methods face the following two fundamental challenges: (1) Indicates alignment challenge There is a significant gap between the structural representations learned by GNNs and the semantic knowledge captured by LLMs. While GNN-based methods excel at modeling collaborative signals through interaction patterns, they lack a deep understanding of the semantic attributes behind user preferences and item features. Conversely, LLM-based methods capture rich semantic details but may overlook important structural patterns in interactive data. The limited effectiveness of methods relying on direct contrastive learning between fundamentally different representation spaces cannot bridge these significant gaps.

[0007] (2) Noise mitigation challenges Real-world recommendation scenarios invariably contain noise in user-item interactions, including false positives (interactions driven by popularity bias rather than genuine interest) and false negatives (potentially relevant items that the user has not yet discovered). Current LLM-based methods primarily focus on semantic aspects and fail to address the noise problem by leveraging structural properties that could help distinguish genuine preferences from noisy interactions. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes an item recommendation method and system based on Large Language Model Enhanced Graph Representation Learning (LLM-AGR). The method employs an LLM-AGR recommendation model, collaboratively integrating semantic knowledge derived from LLM with graph-based structural information. Simultaneously, adaptive structure learning is used to mitigate noise, further improving the quality of item learning representation recommendations.

[0009] According to some embodiments, the first solution of the present invention provides an item recommendation method based on large language model enhanced graph representation learning, which adopts the following technical solution: An item recommendation method based on large language model-enhanced graph representation learning includes: Obtain the user-item graph recommendation dataset; Extract semantic knowledge features from the acquired graph recommendation dataset; Based on the extracted semantic knowledge features and the item recommendation model, the item recommendation learning is completed. The item recommendation model adopts a large language model-enhanced graph representation learning model, embeds semantic similarity through the large language model, uses an adaptive graph structure learning mechanism to identify semantic boundaries to determine structural information, performs bidirectional knowledge interaction and transfer between semantic and structural information, combines information bottleneck regularization filtering to alleviate interaction noise, optimizes the item recommendation model with the goal of minimizing the comprehensive multi-objective loss function, and completes the item recommendation by calculating the item preference score.

[0010] As a further technical limitation, in the process of extracting semantic knowledge features from the acquired graph recommendation dataset, textual description information from the user-item recommendation dataset is obtained based on a large language model. The large language model embedding function is used to embed user-item profiles into the acquired textual description information. The semantic space of the semantic knowledge derived from the embedded large language model is then aligned to obtain semantic knowledge features.

[0011] As a further technical limitation, the process of identifying semantic boundaries using an adaptive graph structure learning mechanism includes at least edge importance estimation and denoised graph construction; wherein, the importance score between user node embeddings and item node embeddings is calculated through a parameterized graph learning network, and the edge importance is evaluated by combining the obtained importance score; combined with the obtained importance score, a denoised adjacency matrix is ​​constructed, and interactive denoising is performed using the constructed denoised adjacency matrix to complete the construction of the denoised graph and obtain semantic boundary structure information.

[0012] As a further technical limitation, the bidirectional knowledge exchange and transfer includes at least semantic knowledge distillation and structural knowledge distillation, which are used to achieve comprehensive information exchange between different data modalities.

[0013] As a further technical limitation, the comprehensive multi-objective loss function includes at least a Bayesian personalized ranking loss function for recommendation, an overall semantic distillation loss function, an overall structural knowledge distillation loss function, a reconstruction loss function, an independence criterion function, and a denoising loss function.

[0014] As a further technical limitation, in the process of item recommendation, a graph structure and learned embedding are used to generate personalized recommendations for items. By calculating the preference scores of all candidate items for each user, the items are ranked according to the obtained preference scores, and the items are recommended based on the obtained item ranking.

[0015] According to some embodiments, a second aspect of the present invention provides an item recommendation system based on large language model-enhanced graph representation learning, employing the following technical solution: An item recommendation system based on large language model-enhanced graph representation learning includes: The acquisition module is configured to acquire a user-item graph recommendation dataset; The extraction module is configured to extract semantic knowledge features from the acquired graph recommendation dataset; The recommendation module is configured to learn and recommend items based on extracted semantic knowledge features and an item recommendation model. The item recommendation model employs a large language model-enhanced graph representation learning model, embedding semantic similarity within the large language model, using an adaptive graph structure learning mechanism to identify semantic boundaries and determine structural information, and performing bidirectional knowledge exchange between semantic and structural information. Information bottleneck regularization filtering is combined to mitigate interaction noise, and the item recommendation model is optimized with the goal of minimizing the comprehensive multi-objective loss function. Item recommendations are completed by calculating item preference scores.

[0016] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium, employing the following technical solution: A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the item recommendation method based on large language model-enhanced graph representation learning as described in the first aspect of the present invention.

[0017] According to some embodiments, the fourth aspect of the present invention provides an electronic device, which adopts the following technical solution: An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the item recommendation method based on large language model augmented graph representation learning as described in the first aspect of the present invention.

[0018] According to some embodiments, the fifth aspect of the present invention provides a computer program product, which adopts the following technical solution: A computer program product includes software code, wherein the program in the software code performs the steps of the item recommendation method based on large language model augmented graph representation learning as described in the first aspect of the present invention.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses a key challenge in recommender systems by effectively aligning and fusing semantic knowledge and collaborative signals while mitigating noisy interactions. By integrating LLM-derived semantic knowledge with adaptive graph structure learning, LLM-AGR creates a seamless bridge between semantic understanding and structural patterns in user-item interactions. Enriching semantic representations through LLM enables more comprehensive user-item preference reasoning beyond simple interaction patterns. A dual knowledge distillation mechanism bridges the gap between semantic understanding and collaborative signals. Information bottleneck (IB) regularization effectively aligns denoised knowledge with recommendation goals, balancing information preservation and noise reduction. Based on LLM-AGR, the collaborative integration of LLM-derived semantic knowledge and graph-based structural information, along with adaptive structure learning to mitigate noise, improves the quality of item learning representation recommendations. Attached Figure Description

[0020] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0021] Figure 1 This is a flowchart of the item recommendation method based on large language model enhanced graph representation learning in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the overall framework structure of LLM-AGR in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the ablation study results on the Amazon-book and Yelp datasets in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram comparing the performance under different noise injection levels in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of cold start performance analysis in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the sensitivity analysis of key hyperparameters of LLM-AGR in Embodiment 1 of the present invention; Figure 7 This is a structural block diagram of the item recommendation system based on large language model enhanced graph representation learning in Embodiment 2 of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 Embodiment 1 of this invention introduces an item recommendation method based on large language model-enhanced graph representation learning.

[0028] like Figure 1 The illustrated item recommendation method based on large language model-enhanced graph representation learning includes: Obtain the user-item graph recommendation dataset; Extract semantic knowledge features from the acquired graph recommendation dataset; Based on the extracted semantic knowledge features and the item recommendation model, the item recommendation learning is completed. The item recommendation model adopts a large language model-enhanced graph representation learning model, embeds semantic similarity through the large language model, uses an adaptive graph structure learning mechanism to identify semantic boundaries to determine structural information, performs bidirectional knowledge interaction and transfer between semantic and structural information, combines information bottleneck regularization filtering to alleviate interaction noise, optimizes the item recommendation model with the goal of minimizing the comprehensive multi-objective loss function, and completes the item recommendation by calculating the item preference score.

[0029] It's worth noting that graph neural networks (GNNs) have demonstrated significant effectiveness in recommender systems by modeling high-order collaborative relationships. Firstly, embedding propagation pioneered the integration of GNNs and collaborative filtering. LightGCN, through feature removal transformations and nonlinear activations, simplifies the process to a lightweight architecture that often outperforms complex models in recommender scenarios. Secondly, self-supervised learning can enhance robustness and address data sparsity. A unified graph augmentation operator is employed, introducing contrastive learning without augmentation, resulting in significant improvements in sparse recommender scenarios. Decoupling representations captures multiple user intents, while domain-specific extensions integrate additional information sources. Knowledge graphs are also combined to enhance semantic modeling. Despite these advancements, GNN recommenders primarily rely on structural patterns while neglecting rich textual semantics and struggle with noise recognition, especially in sparse data scenarios.

[0030] This embodiment uses a large language model for item recommendation, which can be divided into two main paradigms based on LLM integration: (1) LLM as a predictor As a predictor, LLM directly generates recommendations through context learning, cue tuning, and instruction tuning; it effectively connects LLM and recommendation tasks by fine-tuning using recommendation data; however, these methods typically incur high computational costs, limiting large-scale practical deployment.

[0031] (2) LLM as an extractor As an extractor, LLM provides semantic profiles for downstream recommendation tasks; it unifies diverse recommendation data into natural language sequences through adaptive personalized prompts; and in the context of GNN, it enhances user-item interactions and multimodal data by aligning textual representations from LLM with structural representations from GNN.

[0032] Despite these advances, current LLM-based methods exhibit significant limitations. Effectively bridging semantic gaps across different representation spaces remains a major challenge, as the direct alignment of textual and structural information is inherently complex. Furthermore, these methods demonstrate insufficient ability to mitigate the effects of noise in interactive data, focusing primarily on semantic aspects without leveraging structural properties to identify and filter noisy interactions.

[0033] The LLM-AGR framework in this embodiment addresses these limitations through bidirectional knowledge distillation between semantic and structural domains, while employing adaptive graph structure learning to filter noisy interactions for more effective recommendations.

[0034] set up Represents a user set, Represents a set of items, in which and These represent the number of users and items, respectively. (Binary matrix) This represents user-item interaction, where Indicates user With items Interaction between them Otherwise, this embodiment models these interactions as a bipartite graph. ,in Represents the edge set; that is .

[0035] Using semantic embeddings derived from large language models, users and items are represented respectively. and ,in It is the semantic embedding dimension; learning representations that effectively capture collaborative signals and semantic information to predict unobserved user-item interactions; that is... .

[0036] Graph Neural Networks (GNNs) have become a powerful method for recommendation by leveraging the graph structure of user-item interactions, enabling them to perform message-passing operations to capture cooperative signals. ; in Indicates the first Layered connection user-item embedding matrix, It is an adjacency matrix (or its normalized variant). Indicates learnable parameters, It is a message passing function specific to each GNN architecture.

[0037] The final representation typically combines embeddings from different layers: ; in, It is a function that combines hierarchical embeddings and can be a weighted sum, a join, or other aggregation method.

[0038] Existing GNN-based recommendation models face two key limitations: relying solely on ID-based collaborative signals while ignoring semantic information, and lacking appropriate mechanisms to handle implicit feedback noise, which may significantly reduce recommendation performance.

[0039] The information bottleneck principle provides a theoretical framework for balancing information preservation and compression in representation learning. When learning input... The representation of To save information about the target When information is discarded along with irrelevant details, the target for: ; in, Indicates mutual information, Control the compression-correlation tradeoff.

[0040] To effectively implement this principle, this embodiment uses the Hilbert-Schmidt Independence Criterion (HSIC) as an approximation of mutual information: ; in, and It uses a Gaussian kernel from and The calculated kernel matrix, It is defined as The centralized matrix, Represents the trace of a matrix.

[0041] This embodiment uses, as follows: Figure 2The LLM-AGR overall framework shown includes semantic knowledge integration, adaptive graph structure learning, bidirectional knowledge distillation, and information bottleneck regularization; specifically: (1) Semantic knowledge integration This embodiment extracts high-quality semantic representations from noisy text data related to users and items. Real-world datasets like Amazon Book contain item descriptions with many irrelevant attributes, while user reviews often include subjective opinions, sentiment expressions, and irrelevant content that may not accurately reflect core preferences. This embodiment addresses these challenges by leveraging the powerful semantic understanding capabilities of LLM and a carefully designed prompting strategy to extract meaningful preference information despite these inherent noise sources.

[0042] Inspired by the powerful summarizing capabilities of LLM, this implementation enhances collaborative filtering by leveraging rich textual information such as user reviews and item descriptions. To effectively mitigate the LLM illusion and improve response quality, this example provides each user with... and items Design specific prompts Obtain high-quality configuration files from LLM, i.e.: ; in, This indicates a prompt for generating user / item profiles. As input for the subsequent embedding stage.

[0043] Semantic information about users and items is converted into embeddings using LLM embedding functions: ; These embeddings capture rich textual information that complements the structural patterns in the interaction graph. For example, in movie recommendation, LLM can extract genre preferences, director preferences, and thematic interests from review texts that may not be explicitly captured individually in the interaction data; semantic embeddings are pre-computed and maintained as fixed tensors during training, serving as anchors for knowledge distillation.

[0044] To effectively integrate semantic embeddings and cooperative representations, this embodiment uses a multilayer perceptron to project them into a cooperative space. For semantic embeddings... (from) or ) entities (representing user) or items Projection is defined as: ; in, , , , These are the learnable parameters of the MLP; It is a dimension of semantic space. It is the dimension of the hidden layer. It is the dimension of the collaborative space; this projection network, implemented as a shared component between users and items, facilitates effective knowledge transfer between semantics and the collaborative space; the resulting... The semantic embedding representing the projection can be directly integrated into the graph-based recommendation framework in this embodiment, enabling the semantic knowledge derived from LLM to be incorporated into the collaborative filtering process.

[0045] (2) Adaptive graph structure learning This embodiment employs adaptive graph structure learning to address the interaction noise problem present in real-world recommendation scenarios. For example, users often click on items due to misleading thumbnails, popularity bias, or accidental interactions, creating false positive signals. The adaptive structure learning mechanism used in this embodiment identifies and filters these unreliable edges while preserving basic collaborative patterns.

[0046] For each edge in the interaction graph This embodiment is based on node embedding of connections. and (representing user) and items The initial collaborative embedding uses a parametric graph learning network to compute importance scores, allowing the model to learn complex relationships between users and items for edge importance evaluation, i.e., the importance score is: ; in, Indicates a connection. and These are learnable parameters. It is the sigmoid function.

[0047] To maintain differentiability and enable random exploration during training, this embodiment employs Gumbel-Softmax reparameterization, namely: ; in, From Gumbel distribution Mid-sampling, This is the temperature parameter that controls the sharpness of the distribution (in this embodiment, it is set to 0.2 by default). It is a bias term with a default value of 0.5, used for balancing edge preservation and pruning during the initial training phase.

[0048] This embodiment uses the learned edge importance scores to construct a denoised adjacency matrix by modulating the original adjacency values. ,Right now: ; in, It is the original adjacency matrix. Includes the calculated edge importance score. This represents element-wise multiplication. The denoised adjacency matrix. By assigning higher weights to reliable edges and lower weights to potentially noisy edges, the interaction graph is effectively denoised while preserving its sparse structure for efficient computation.

[0049] (3) Two-way knowledge distillation This embodiment promotes bidirectional knowledge transfer between semantic and structural information through bidirectional knowledge distillation; unlike traditional recommendation models that focus on collaborative signals or content features, it achieves comprehensive information exchange between different data modalities.

[0050] By comparing the InfoNCE learning method, knowledge is transferred from semantic embedding to co-embedding. Let... Indicates training triplet batches, users With the right items Interact but not with negative items Interaction.

[0051] For each entity (representing user) or items ),Right now ; in, It is a physical entity Collaborative embedding, It is a projective semantic embedding from an aligned MLP. Indicates and A collection of entities of the same type (i.e., users correspond to users, and items correspond to items). It is a temperature parameter that controls the sharpness of the distribution.

[0052] The overall semantic distillation loss aggregates these components on the sampled triples: .

[0053] As a supplement to semantic knowledge distillation, this embodiment can realize structural knowledge distillation, transferring information from augmented graph structure to refined graph structure; structural representation derived from LLM augmented graph ( ) and the collaborative representation from the fine-scale map ( Maintain separate embedding spaces and facilitate knowledge transfer between them.

[0054] Following the same InfoNCE framework as semantic distillation, for each entity ,Right now: ; in, Entities derived from augmented graphs The structural embedding, It is the temperature parameter for structural distillation.

[0055] The overall structural knowledge distillation loss is: ; This bidirectional knowledge transfer effectively bridges semantic and structural spaces, enhancing representation quality by allowing information flow between different views of user-item interaction data.

[0056] (4) Information bottleneck regularization The information bottleneck principle is used to balance information preservation and noise reduction. In real-world recommendation scenarios, distinguishing between information patterns and noise is crucial for model performance.

[0057] Given the embedding from the original graph and embeddings from fine-grained maps (For entities) They optimize their mutual information to preserve the basic pattern while filtering out noise.

[0058] This embodiment uses the Hilbert-Schmidt Independence Criterion (HSIC) to measure independence: ; in, and These are the embedding matrices from the original and refined maps, respectively. It is the number of entities in the batch. and It is the kernel matrix calculated from the corresponding embedding. It is a centralized matrix. Represents the trace of a matrix.

[0059] This embodiment uses a Gaussian kernel to calculate the kernel matrix: ; in, It is the kernel width parameter. This kernel-based approach allows for the capture of complex, nonlinear relationships between different spatial embeddings.

[0060] The formulaic HSIC loss for user and item embedding in this embodiment is: ; in, and These are user embedding matrices from the original and refined graphs, respectively, and similarly... and The corresponding items are embedded with the same symbols as the pre-knowledge section. Minimizing this loss encourages fine embeddings to remain independent of the original noisy embeddings, effectively acting as an information bottleneck to filter redundant information, while preserving the basic structural patterns through other target components.

[0061] This embodiment uses LLM-AGR to optimize recommendations and obtain item recommendations for the user; all model parameters are optimized by minimizing a comprehensive multi-objective loss function that balances recommendation accuracy, representation quality, and noise reduction. ; in, This is the Bayesian personalized ranking loss used for recommendations: ; in, Indicates user and items Predicted preference score It is an L2 regularization term, which prevents all trainable parameters in the model from being affected. Overfitting.

[0062] Reconstruction loss Defined as: ; in, Representing entities Predictive semantic embedding, It is a true projection semantic embedding. It is a temperature parameter. This loss term enhances the robustness of the representation by encouraging the model to accurately reconstruct semantic information from co-embeddings.

[0063] hyperparameters , , , and The relative importance of each loss component is controlled and carefully tuned to balance different learning objectives. Through extensive experiments on various datasets, it is found that, relative to... set up and Higher values ​​produce optimal performance. Optimize to maintain a balanced contribution to the overall objectives.

[0064] During inference, this embodiment employs a sophisticated graph structure and learned embeddings to generate personalized recommendations. For each user... Calculate the preference scores for all candidate items, and you will get: ; in, It is a collection of all items. User The system collects items that have already been interacted with; then, it sorts these items based on preference scores and recommends them to the user. Items.

[0065] It should be noted that during inference, this embodiment disables the dropout mechanism and uses the full fine graph structure to ensure stable and consistent recommendation results; a balance is provided between enhanced semantic understanding from LLM knowledge and fine collaborative signals from the adaptive graph learning mechanism, leveraging semantic similarity captured through LLM embedding and structural patterns identified by the adaptive graph learning mechanism.

[0066] This embodiment uses three real-world datasets to verify the performance of LLM-AGR. The statistics of the datasets used are summarized in Table 1.

[0067] Table 1. Statistics of the Experimental Dataset

[0068] As shown in Table 1, Amazon-Book contains user book purchase, rating, and review records from Amazon; Yelp ] It covers user reviews and ratings of local businesses; Amazon-movie provides user movie viewing history, ratings, and review data from Amazon. Data processing follows previous work: interactions with ratings below 1 in Amazon-Book, below 2 in Yelp, and below 3 in Amazon-movie are filtered out, followed by k-core filtering. The Amazon-Book and Yelp datasets are split into training, validation, and test sets in a 3:1:1 ratio, while the Amazon-Movie dataset is split in an 8:1:1 ratio.

[0069] This embodiment evaluates the effectiveness of LLM-AGR by integrating it with a state-of-the-art graph-based recommender; it comprehensively compares its performance with traditional graph-based methods, knowledge augmentation methods, and recent LLM-based recommender systems.

[0070] Basic model: LightGCN: A simplified GCN-based model that removes redundant feature transformations and non-linear activations, significantly improving training efficiency and generalization ability for recommendation tasks.

[0071] SGL: A self-supervised graph learning method that systematically explores the basic graph structure to enhance representation learning through unified augmentation operators including node dropout, edge dropout, and random walks.

[0072] SimGCL: A simple and effective graph-free augmented contrastive learning method that injects uniform noise into the embedding and smooths out the uniformity to achieve superior performance with minimal computational overhead.

[0073] BIGCF: A framework for exploring the individuality and collectivity of the intentions behind collaborative filtering interactions.

[0074] Knowledge augmentation and LLM-based methods: KAR: Incorporates external knowledge of user preferences and factual knowledge about items, which is compressed into vectors to improve recommendations.

[0075] Semantic Only: An ablation variant that uses only LLM-derived semantic embeddings without combining graph structure learning, which helps to separate the contribution of semantic knowledge.

[0076] RLMRec-Con, RLMRec-Gen: Recent state-of-the-art LLM-enhanced recommendation methods, including LLM ensemble methods for fair comparison of recommendation systems.

[0077] To ensure fair comparison, this embodiment uses the sampling method and dataset format of the open-source framework SSLRec. In the experimental setup, the embedding dimension of the model and the transformed LLM embeddings was uniformly set to 32. All models, including the baseline and base models, were trained using the Adam optimizer, with embeddings initialized using Xavier, a fixed batch size of 4096, and a learning rate of 1e-3.

[0078] This embodiment sets the knowledge distillation coefficient. Information bottleneck coefficient Structural knowledge weight str_weight LLM knowledge integration coefficient Furthermore, early stopping is applied based on the model's performance on the validation set to prevent overfitting. For text embedding generation, OpenAI's gpt-3.5-turbo is used for configuration file generation, and text-embeddings-3-large is used for semantic representation to ensure the accuracy of the expression and the consistency of the semantic representation.

[0079] To verify the superior performance of the LLM-AGR framework used in this embodiment, the LLM-AGR and the state-of-the-art recommendation model, as shown in Table 2, were comprehensively evaluated on three public datasets. The results of all experiments represent the average of five independent runs to ensure statistical reliability. In Table 2, the best results are highlighted in bold with an asterisk (*). "Improv." indicates the relative improvement of LLM-AGR relative to the best baseline.

[0080] Table 2 Performance comparison of different methods on three datasets

[0081] LLM-AGR outperforms all baseline methods across various architectural paradigms. When integrated with LightGCN, the method in this embodiment achieves significant improvements on Amazon-book (highest recall @5 8.22%), Yelp (highest NDCG @5 8.69%), and Amazon-movie (highest NDCG @5 16.41%); these improvements persist in more complex models, with significant gains observed in SGL, SimGCL, and even the advanced BIGCF framework. These general improvements demonstrate the effectiveness of this embodiment as a dual-objective approach to addressing the challenges of semantic integration and noise mitigation.

[0082] The significant performance gap between the semantic-only baseline and LLM-AGR empirically validates the core assumption of this embodiment: effective recommendation requires the synergistic integration of semantic understanding and collaborative signals. On Amazon-book, the recall of the semantic-only method @20 is only 0.0199, while LLM-AGR reaches 0.1567, an almost eight-fold improvement. The bidirectional knowledge distillation mechanism in this embodiment achieves this integration by facilitating the interactive information transfer between LLM-derived semantic embeddings and graph-based collaborative signals.

[0083] A consistent pattern emerges across all datasets and architectures: LLM-AGR yields significantly greater improvements at lower cutoff thresholds, where user attention is focused in real-world applications. On Amazon-movie, which uses LightGCN as its backbone, the improvement in recall@5 (14.63%) significantly outweighs that of recall@20 (4.95%). Similarly, NDCG@5 shows a gain of 16.41%, compared to 11.22% for NDCG@20. This pattern confirms LLM-AGR's strength in identifying and prioritizing the most relevant items—a crucial capability for real-world recommendation scenarios.

[0084] To evaluate the contribution of each core component in LLM-AGR, this embodiment conducts an ablation study on four graph-based recommendation architectures (LightGCN, SimGCL, SGL, and BIGCF) on the Amazon-book and Yelp datasets. The results are as follows: Figure 3 As shown; among them, the created variant is: Base: The original GNN base recommender without the enhancements proposed in this embodiment; w / o SE: Remove semantic embeddings from LLM and use only co-signals; w / o KD: Remove the two-way knowledge distillation mechanism; w / o AGS+IB: Removes adaptive graph structure learning and information bottleneck regularization; LLM-AGR: A complete framework containing all proposed components; The complete LLM-AGR framework consistently outperforms all ablation variants, validating the effectiveness of the method in this embodiment. This advantage is most pronounced when using the Light GCN-based architecture, showing the greatest relative improvement in recall and NDCG metrics.

[0085] The contributions of each component differ in importance. Removing the adaptive graph structure learning with its information bottleneck (w / o AGS+IB) results in a moderate performance degradation, but still provides consistent improvements across architectures, particularly noticeable in BIGCF-based models. A more significant performance drop occurs when the knowledge distillation mechanism (w / o KD) is removed, highlighting the crucial role of bidirectional knowledge transfer in bridging semantic understanding and collaborative filtering. Semantic embeddings from LLM provide the most significant contribution, as evidenced by the considerable gap between the base / w / o SE variant and other variants, confirming that the semantic knowledge derived from LLM significantly enhances recommendation quality beyond what can be achieved with interaction data alone.

[0086] To assess robustness to noisy interactions, this embodiment systematically introduces controlled levels of noise into the interaction data by randomly changing a certain percentage of user-item interactions (from positive to negative or vice versa). Figure 4 This illustrates how different models degrade in performance with increasing noise levels.

[0087] like Figure 4 As shown, all models experience performance degradation with increasing noise levels, but LLM-AGR maintains significantly better performance across all noise ratios from 5% to 25%. At higher noise levels, LLM-AGR significantly outperforms the baseline model in recall@20 and NDCG@20 metrics, with the performance gap widening as noise increases. This highlights the superior noise resilience of LLM-AGR, which can be attributed to two key mechanisms: (1) Adaptive graph structure learning Edge weight computation effectively identifies and reduces the weights of potentially noisy interactions, creating a cleaner structural representation. Information bottleneck regularization further helps eliminate noise while preserving essential information.

[0088] (2) Semantic knowledge integration LLM-derived semantic embeddings provide additional signals orthogonal to the interaction data, allowing the model to rely more on semantic understanding when interaction patterns become unreliable due to noise.

[0089] The cold start problem—recommending items to users with little or no historical interaction—remains a significant challenge for collaborative filtering systems. To evaluate the performance of LLM-AGR in a cold start scenario, this embodiment performs a comprehensive analysis, examining recommendation quality as a function of user interaction counts, such as... Figure 5 As shown.

[0090] For users with minimal interaction (groups [0, 13]), LLM-AGR achieves significantly better NDCG@20 performance compared to LightGCN, SGL, SimGCL, BIGCF, and RLMRec. As shown in the improvement chart, LLM-AGR shows nearly a 20% improvement over LightGCN for the least interactive users and significantly outperforms all other baseline models.

[0091] LLM-AGR maintained superior performance across all user interaction groups, from the least active ([0, 13]) to those with moderate ([13, ​​28], [28, 74]) and extensive interaction history ([74, 444]).

[0092] like Figure 5 As shown in the violin-shaped plot at the bottom, LLM-AGR exhibits the smallest performance gap compared to the baseline method between cold-start users (group [0,13]) and regular users (group [74, 444]).

[0093] This embodiment analyzes the computational complexity of LLM-AGR to evaluate its theoretical efficiency. Let... , , , and These represent the number of users, the number of items, the number of interactions, the embedding dimension, and the number of GNN layers, respectively.

[0094] The time complexity of LLM-AGR comprises several components. Graph convolution operations require... Time constitutes the fundamental complexity inherited from LightGCN. The bidirectional knowledge distillation mechanism introduces additional complexity through MLP transformations between semantic and structural domains. Operation. For adaptive graph structure learning, edge weight calculation requires... The operation is complex because each interaction is processed through a neural network. Information bottleneck regularization based on HSIC computation increases. Complexity, in This represents the batch size; therefore, the total training time complexity is... Considering Typically very small (set to 2-3 in this embodiment) and in practical recommended scenarios The dominant term becomes This results in an effective asymptotic complexity matching that of LightGCN, demonstrating that the semantic integration and adaptive structure learning enhancements in this embodiment maintain computational efficiency.

[0095] To comprehensively evaluate the robustness of LLM-AGR and gain a deeper understanding of the impact mechanism of key hyperparameters on model performance, this paper conducts a comprehensive sensitivity analysis of six key hyperparameters on the Amazon-book dataset, using LightGCN as the base model. Figure 6 As shown, the system analyzes the specific impact of each parameter on recommendation performance (Recall@20 and NDCG@20).

[0096] The kernel width parameter σ controls the sensitivity of semantic similarity calculation. Optimal performance is achieved when σ=0.4. Values ​​that are too small lead to oversensitivity, while values ​​that are too large make the similarity calculation too broad, affecting the accuracy of semantic matching.

[0097] The LLM knowledge integration coefficient α balances the fusion ratio of LLM semantic knowledge and collaborative signals. Optimal performance is achieved when α=0.2. Values ​​that are too high (>0.4) cause the semantic signal to overwhelm the collaborative mode, while values ​​that are too low (<0.1) fail to fully utilize the rich semantic knowledge of LLM.

[0098] Information bottleneck coefficient β: It performs best when β=5.0, effectively balancing the dual needs of information preservation and compression. Lower values ​​lead to insufficient compression, while higher values ​​lead to excessive loss of key information and useful signals.

[0099] Structural knowledge weight str_weight: Peak performance is achieved when str_weight=1.0. Values ​​below this do not fully utilize graph structural information, while values ​​above this overemphasize structural patterns and affect semantic understanding capabilities.

[0100] The preference knowledge weight pref_weight performs best when pref_weight=0.02. Performance drops significantly when pref_weight exceeds 0.04, demonstrating the importance of accurately balancing preference knowledge with other signal sources.

[0101] The Gumbel-Softmax temperature τ_G controls the degree of randomness in the adaptive graph structure learning. It is optimal when τ_G = 0.2. Too low a value leads to overly deterministic selection, while too high a value introduces too much noise interference.

[0102] Experimental results show that LLM-AGR maintains good stability near each optimal value, fully demonstrating the strong robustness of the model architecture to moderate hyperparameter changes.

[0103] This embodiment utilizes a novel framework to enhance the denoising process of graph-based recommender systems using Large Language Models (LLM). By effectively aligning and fusing semantic knowledge with collaborative signals, it simultaneously mitigates noisy interactions, addressing a key challenge in recommender systems. It integrates LLM-derived semantic knowledge with adaptive graph structure learning, creating a seamless bridge between semantic understanding and structural patterns in user-item interactions through LLM-AGR. By enriching semantic representations with LLM, it enables more comprehensive user-item preference reasoning beyond simple interaction patterns. A dual knowledge distillation mechanism bridges the gap between semantic understanding and collaborative signals, while Information Bottleneck (IB) regularization effectively aligns denoising knowledge with the recommendation objective, balancing information preservation and noise reduction.

[0104] Example 2 Embodiment 2 of this invention introduces an item recommendation system based on large language model-enhanced graph representation learning.

[0105] like Figure 7 The illustrated item recommendation system based on large language model-enhanced graph representation learning includes: The acquisition module is configured to acquire a user-item graph recommendation dataset; The extraction module is configured to extract semantic knowledge features from the acquired graph recommendation dataset; The recommendation module is configured to learn and recommend items based on extracted semantic knowledge features and an item recommendation model. The item recommendation model employs a large language model-enhanced graph representation learning model, embedding semantic similarity within the large language model, using an adaptive graph structure learning mechanism to identify semantic boundaries and determine structural information, and performing bidirectional knowledge exchange between semantic and structural information. Information bottleneck regularization filtering is combined to mitigate interaction noise, and the item recommendation model is optimized with the goal of minimizing the comprehensive multi-objective loss function. Item recommendations are completed by calculating item preference scores.

[0106] The detailed steps are the same as those of the item recommendation method based on large language model enhanced graph representation learning provided in Example 1, and will not be repeated here.

[0107] Example 3 Embodiment 3 of the present invention provides a computer-readable storage medium.

[0108] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the item recommendation method based on large language model augmented graph representation learning as described in Embodiment 1 of the present invention.

[0109] The detailed steps are the same as those of the item recommendation method based on large language model enhanced graph representation learning provided in Example 1, and will not be repeated here.

[0110] Example 4 Embodiment 4 of the present invention provides an electronic device.

[0111] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the item recommendation method based on large language model enhanced graph representation learning as described in Embodiment 1 of the present invention.

[0112] The detailed steps are the same as those of the item recommendation method based on large language model enhanced graph representation learning provided in Example 1, and will not be repeated here.

[0113] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0114] A computer program product includes software code, wherein the program in the software code performs the steps of the item recommendation method based on large language model enhanced graph representation learning as described in Embodiment 1 of the present invention.

[0115] The detailed steps are the same as those of the item recommendation method based on large language model enhanced graph representation learning provided in Example 1, and will not be repeated here.

[0116] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0121] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0122] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. An item recommendation method based on large language model-enhanced graph representation learning, characterized in that, include: Obtain the user-item graph recommendation dataset; Extract semantic knowledge features from the acquired graph recommendation dataset; Based on the extracted semantic knowledge features and the item recommendation model, the item recommendation learning is completed. The item recommendation model adopts a large language model-enhanced graph representation learning model, embeds semantic similarity through the large language model, uses an adaptive graph structure learning mechanism to identify semantic boundaries to determine structural information, performs bidirectional knowledge interaction and transfer between semantic and structural information, combines information bottleneck regularization filtering to alleviate interaction noise, optimizes the item recommendation model with the goal of minimizing the comprehensive multi-objective loss function, and completes the item recommendation by calculating the item preference score.

2. The item recommendation method based on large language model augmented graph representation learning as described in claim 1, characterized in that, In the process of extracting semantic knowledge features from the acquired graph recommendation dataset, textual description information from the user-item recommendation dataset is obtained based on a large language model. User-item profiles are embedded into the acquired textual description information using a large language model embedding function. The semantic space of the embedded semantic knowledge derived from the large language model is then aligned to obtain semantic knowledge features.

3. The item recommendation method based on large language model augmented graph representation learning as described in claim 1, characterized in that, The process of identifying semantic boundaries using an adaptive graph structure learning mechanism includes at least edge importance estimation and denoised graph construction. Specifically, an importance score between user node embeddings and item node embeddings is calculated using a parameterized graph learning network, and the edge importance is evaluated by combining the obtained importance scores. A denoised adjacency matrix is ​​constructed based on the obtained importance scores, and interactive denoising is performed using the constructed denoised adjacency matrix to complete the construction of the denoised graph and obtain semantic boundary structure information.

4. The item recommendation method based on large language model augmented graph representation learning as described in claim 1, characterized in that, The bidirectional knowledge exchange and transfer includes at least semantic knowledge distillation and structural knowledge distillation, which are used to achieve comprehensive information exchange between different data modalities.

5. The item recommendation method based on large language model augmented graph representation learning as described in claim 1, characterized in that, The comprehensive multi-objective loss function includes at least the Bayesian personalized ranking loss function for recommendation, the overall semantic distillation loss function, the overall structural knowledge distillation loss function, the reconstruction loss function, the independence criterion function, and the denoising loss function.

6. The item recommendation method based on large language model augmented graph representation learning as described in claim 1, characterized in that, In the process of item recommendation, a graph structure and learned embedding are used to generate personalized recommendations for items. The preference scores of all candidate items for each user are calculated, the items are ranked according to the obtained preference scores, and the items are recommended based on the obtained item ranking.

7. An item recommendation system based on large language model-enhanced graph representation learning, characterized in that, include: The acquisition module is configured to acquire a user-item graph recommendation dataset; The extraction module is configured to extract semantic knowledge features from the acquired graph recommendation dataset; The recommendation module is configured to learn and recommend items based on extracted semantic knowledge features and an item recommendation model. The item recommendation model employs a large language model-enhanced graph representation learning model, embedding semantic similarity within the large language model, using an adaptive graph structure learning mechanism to identify semantic boundaries and determine structural information, and performing bidirectional knowledge exchange between semantic and structural information. Information bottleneck regularization filtering is combined to mitigate interaction noise, and the item recommendation model is optimized with the goal of minimizing the comprehensive multi-objective loss function. Item recommendations are completed by calculating item preference scores.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the item recommendation method based on large language model augmented graph representation learning as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the item recommendation method based on large language model augmented graph representation learning as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the item recommendation method based on large language model augmented graph representation learning as described in any one of claims 1-6.

Images