Anti-information cocoon house recommendation method and system fusing large model semantics and hypergraph structure

Abstract

The invention discloses an anti-information cocoon house recommendation method and system fusing large model semantics and a hypergraph structure. The method comprises the steps that article attribute information and interaction information between a user and an article are obtained through large language model reasoning; constructing an original relation graph; article semantic embedding features and user semantic embedding features are extracted; modeling is carried out on the user behavior sequence through a transform branch, and local fine-grained node representation is captured; constructing a heterogeneous hypergraph structure through hypergraph branches, and capturing high-order node representation of a global topological structure; aligning the local fine-grained node representation and the high-order node representation of the global topological structure through comparative learning; splicing the aligned node representations to obtain a fusion feature; and generating an article personalized recommendation list based on the fused features. According to the method, through LLM-driven semantic enhancement hypergraph construction, double-branch collaborative representation learning and double-vision comparison alignment, the problems of poor semantics, incomplete structure and insufficient diversity of a traditional hypergraph model are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information technology, specifically to an anti-information cocoon recommendation method and system that integrates large model semantics and hypergraph structure. Background Technology

[0002] As the volume of information on online platforms continues to expand, recommendation systems, while helping users efficiently discover valuable content, inevitably face the structural risk of "information cocoons." The more the system reinforces the exposure of similar content based on historical behavior, the narrower the user's visible information boundaries become, which, in the long run, weakens the breadth of perspective and cognitive flexibility. To counteract this problem mechanistically, models need to proactively expand the coverage of sources and viewpoints while ensuring relevance, explicitly introducing diverse signals across groups, positions, and topics.

[0003] Hypergraphs offer a new structured tool for countering information cocoons by explicitly modeling many-to-many group relationships. Connecting multiple related nodes to the same "hyperedge" can naturally characterize group behavior and multi-perspective co-occurrence relationships in real-world scenarios, such as "a collection of content from different sources and perspectives on the same topic" or "a user community with differentiated interest profiles."

[0004] However, existing hypergraph recommendation models are still limited in their anti-cocoon effectiveness by two types of bottlenecks. The first is the limitation of text and source understanding. Most methods treat users and items as anonymous ID nodes, heavily relying on interaction topology and ignoring the rich semantic and authoritative cues contained in item descriptions, user reviews, and external knowledge. This makes it difficult to build reliable multi-perspective connections when data is sparse or novelty requirements are high. The second is the issue of the completeness and credibility of hypergraph construction. Extracting hyperedges solely based on historical behavior often fails to cover potential but unexpressed interests and opinions; while connections based on shallow features are prone to introducing noise or even misleading associations. Poorly structured hypergraphs directly weaken the efficiency of diverse information propagation in the network, hindering the formation of robust diffusion channels for "de-cocooning."

[0005] The development of Large Language Models (LLMs) and the Transformer architecture has offered new possibilities for solving the aforementioned problems. LLMs possess world-class knowledge and powerful natural language understanding and reasoning capabilities, enabling them to generate embedding representations rich in deep semantics for items and users. The Transformer architecture, with its powerful self-attention mechanism, excels at capturing flexible contextual dependencies. However, existing methods still have significant shortcomings in organically combining these advanced technologies with hypergraph recommendation frameworks: either they over-rely on manually designed prompts while ignoring differences in graph structure, or they are insufficient in the deep fusion of semantic and structural information, limiting their effectiveness in counteracting information cocoons in real-world scenarios. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides an anti-information cocoon recommendation method and system that integrates large model semantics and hypergraph structure. This solves the problems of existing technologies ignoring graph structure differences, incomplete hypergraph construction, excessive reliance on manual design, and low recommendation efficiency.

[0007] To achieve the above-mentioned objectives, this invention provides an anti-information cocoon recommendation method that integrates large model semantics and hypergraph structure, comprising: Item attribute information and user interaction information with items are obtained through a large language model, including the text description of the items and user comments. The original relationship graph is constructed based on the item attribute information and the user's interaction information with the item; Extract item semantic embedding features and user semantic embedding features based on the text description of the item and user reviews; Modeling user behavior sequences through transformer branches captures local fine-grained node representations; Heterogeneous hypergraph structures are constructed by hypergraph branching, and high-order node representations of the global topology in the heterogeneous hypergraph structure are captured. By comparing and learning, we can align the local fine-grained node representations with the high-order node representations of the global topology. The local fine-grained node representation of each aligned node is concatenated with the high-order node representation of the global topology to obtain the fusion features, including user fusion features and item fusion features. A personalized list of recommended items is generated based on fused features.

[0008] Secondly, this invention also provides an anti-information cocoon recommendation system that integrates large model semantics and hypergraph structure, comprising: The preprocessing module is used to obtain item attribute information and user interaction information between the item and the user, including the text description of the item and user comments. It is also used to construct the original relationship graph based on the item attribute information and user interaction information between the user and the item. The semantic representation initialization module is used to extract item semantic embedding features and user semantic embedding features based on the item's text description and user comments; The asymmetric bi-branch module is used to model user behavior sequences through transformer branches, capture local fine-grained node representations, and also to construct semantically enhanced hypergraph structures through hypergraph branches, capturing high-order node representations of global topology in heterogeneous hypergraph structures. The alignment and prediction module is used to concatenate the local fine-grained node representation of each aligned node with the high-order node representation of the global topology to obtain fused features; it is also used to generate a personalized recommendation list of items based on the fused features.

[0009] The beneficial effects of this invention are as follows: 1. By constructing three types of hyperedges—including user interaction, item semantic clusters, and user communities—this invention forms a structural foundation capable of naturally introducing cross-topic and cross-group content. Compared to traditional hypergraph methods that rely solely on historical behavior to construct hyperedges, the semantically enhanced hypergraph of this invention can discover potential semantic relationships and proactively expand the boundaries of user interests. Experiments demonstrate that, while ensuring relevance, the recommendation list of this invention covers a wider range of content domains, breaking the homogenization dilemma caused by the over-reliance on popular items in traditional recommendation systems.

[0010] 2. By designing an asymmetric dual-branch architecture, this invention simultaneously captures local fine-grained semantic context and global high-order structural associations. Compared to recommendation models with a single architecture, this collaborative representation learning mechanism can understand user interests at different granularities, focusing on both specific preference details and grasping the macroscopic interest graph. Visualization analysis shows that the node representations learned by this invention are more evenly distributed and have clear boundaries in the embedding space, indicating that it has stronger feature discrimination capabilities.

[0011] 3. By employing contrastive learning to align local and global representations, this invention spontaneously enhances recommendation diversity through representation consistency optimization without explicit diversity labeling. Compared to methods requiring manual labeling of diversity signals, this invention is more practical and scalable, automatically balancing exploration and utilization in real-world recommendation scenarios to provide users with relevant and diverse content.

[0012] 4. Through reasonable modular design and parameter sharing, the computational overhead of this invention is kept within a reasonable range. Experimental data show that, with only a small increase in parameters and computation time, this invention achieves an average improvement of 5.50% and 3.35% in Recall@20 and NDCG@20 metrics on multiple public datasets compared to the optimal baseline, respectively, achieving the best balance between performance and efficiency. Attached Figure Description

[0013] Figure 1 A flowchart of an anti-information cocoon recommendation method that integrates large model semantics and hypergraph structure is provided for an embodiment. Figure 2 A flowchart of the workflow of an anti-information cocoon recommendation system that integrates large model semantics and hypergraph structure; Figure 3 Visualize the learning capabilities of each model. Detailed Implementation

[0014] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0015] like Figure 1 , Figure 2 As shown, in one embodiment of the present invention, an anti-information cocoon recommendation method that integrates large model semantics and hypergraph structure includes the following steps: S1. Obtain item attribute information and user interaction information between the item and the item through large language model reasoning, including the item's text description and user comments.

[0016] The specific information comes from the official description and all public comments left by users.

[0017] S2. Construct the original relationship graph based on item attribute information and user-item interaction information.

[0018] S3. Extract the semantic embedding features of the item and the semantic embedding features of the user based on the text description of the item and user reviews.

[0019] Specific methods include: Build multiple independent text instances for each item Each text instance includes the item's title. ,category and a specific text content Its expression is: .

[0020] Inputting text instances into a large language model generates corresponding summary-inference text, denoted as _____. This text aims to answer: "What is this item? And what kind of user would like it?"

[0021] For the user side, a logically symmetrical approach is adopted. Each user's core textual information originates from their handwritten comments. Based on each user's historical comments and corresponding item information, an independent preference instance is constructed for each user's comment. Each preference instance includes comment content. and their corresponding item titles and categories Its expression is: .

[0022] This prompt guides the LLM to act as a user persona analyst, generating a preference-inference text, denoted as... This text aims to answer the question: "What kind of user is this, based on this review of a specific item? What are their preferences?" For example, if a user reviews the album above, saying, "This album brought me peace during my morning prayers; the arrangements are simple yet deeply moving," the model might infer: "This user values ​​the connection between content and their spiritual life, seeking experiences that bring inner peace and profound insights. They cherish emotional resonance and the beauty of simplicity more than complex techniques." Utilizing LLM's own embedding API (represented as functions) The summary-inference text and preference-inference text are encoded into fixed-dimensional semantic vectors and preference vectors, respectively.

[0023] For the summary-inference text of item i, the semantic vector is obtained: ; For user u's preference-inferred text, we obtain the preference vector: .

[0024] Average pooling is performed on the semantic vector and preference vector respectively to obtain the item semantic embedding features and the user semantic embedding features: In the formula, For the quantity of items, For the number of users.

[0025] Through a unified initialization process, each item and user in the graph is given a high-quality semantic embedding. These vectors centrally embody deep features extracted from text and behavior, providing valuable external knowledge beyond IDs for the model's subsequent interaction learning.

[0026] S4. Model user behavior sequences through transformer branches to capture local fine-grained node representations.

[0027] High-fidelity modeling is performed on the local semantic environment of each node (in this invention, entities in the original relational graph or heterogeneous hypergraph structure are referred to as nodes, such as users and items). To this end, the Transformer architecture is introduced, whose core advantage lies in its powerful self-attention mechanism. This mechanism enables the model to dynamically and adaptively evaluate the importance of different nodes within the local neighborhood, thereby transcending fixed graph structures and capturing finer-grained semantic relationships.

[0028] The specific method is as follows: Obtain the embedded features of user-item interactions; The standard Transformer architecture has a computational complexity of O(N) for its self-attention mechanism when processing sequential data. 2 ), where N is the sequence length. When directly applied to large-scale graph structures, if all nodes are considered as a sequence, its quadratic complexity is O(|V|). 2 This would result in unbearable computational and memory overhead, which is completely unfeasible in recommendation scenarios involving millions of nodes and items.

[0029] To create a unified representation that can simultaneously reflect both types of information for similarity calculation, a hybrid representation is constructed for each node v. This hybrid representation is formed by concatenating its interaction embeddings and semantic embeddings. Its expression is:

[0030] In the formula, This represents a hybrid representation of node v. Represents the interactive embedding feature matrix Middle node The row vector; Represents the semantic embedding feature matrix Middle node The row vectors, and the semantic embedding feature matrix includes user semantic embedding features and item semantic embedding features; Indicates a splicing operation; Based on hybrid representation, cosine similarity is used as the core metric to select the most relevant neighboring nodes for each node, including: The expression for calculating the association strength between any two nodes is:

[0031] In the formula, Represents a node and nodes The strength of the correlation between them The norm is denoted by T, and the superscript T denotes the transpose of the matrix. The neighborhood of each central node is selected based on the strength of the association between nodes, and the expression is as follows:

[0032] In the formula, Represents nodes The K most similar neighbor nodes, argtop-K represents taking the K largest values. Indicates the node The set of all nodes except itself.

[0033] Using this principle-based sampling method, a fixed-size (K) neighbor set is generated for each node u. This set is not only computationally controllable, but more importantly, it represents a highly relevant semantic environment. These carefully selected neighbor nodes will serve as input to the subsequent Transformer aggregation module, providing it with the most crucial local structural and semantic information, thereby enabling efficient and accurate representation learning.

[0034] Aggregate the neighboring nodes of each center point to generate a local fine-grained node representation.

[0035] After identifying the K most relevant neighbor sets for each central node u, the information of these neighbors is aggregated using a Transformer. A core consideration is that the importance of different nodes within the neighborhood should be directly related to their similarity to the central node; that is, neighbors with higher similarity should be given greater weight in the aggregation. Standard self-attention mechanisms learn weights themselves, but they cannot directly utilize the relative position information of neighbors in the "similarity ranking". To inject relative position information into the model, rotational position encoding is used.

[0036] Specific methods include: Based on the association strength between the central node and its neighboring nodes, the neighboring nodes of each central node are sorted in descending order to obtain an ordered local structure sequence, the expression of which is:

[0037] In the formula, For the ordered local structure sequence of the central node u, Let represent the Kth neighbor node of the central node u. This sorting method, driven by both semantics and cooperative signals, ensures that nodes more closely associated with the target node are given priority, establishing a meaningful sequence for subsequent aggregation phases.

[0038] Subsequently, the position of the central node u is defined as 0, and its ordered neighbors are... The position is j. A transformer with rotational position encoding is used to encode the obtained ordered local structure sequence. The weighted sum of the value vectors output by the attention head is as follows: The query vector, key vector, and value vector of each attention head are projected, and their expression is:

[0039] In the formula, This represents the query vector for the h-th attention head. This represents the hybrid representation of node u. This represents the query weight matrix for the h-th attention head. This represents the key vector of the h-th attention head. This represents the mixed representation of the j-th neighbor node. This represents the key weight matrix of the h-th attention head. The weight matrix represents the value of the h-th attention head; When calculating the attention score, RoPE is applied to the query and key vectors to encode their relative position information, expressed as:

[0040] In the formula, This represents the attention score of the h-th attention head. This represents the rotation position encoding matrix applied at the center node. This represents the rotation position encoding matrix applied at node j. The dimension representing the key; The attention weights are calculated using the following expression:

[0041] In the formula, This represents the attention weight of the h-th attention head. express Activation function; The weighted sum of the value vectors is expressed as follows:

[0042] In the formula, This represents the weighted sum of the value vectors output by the h-th attention head; By concatenating the outputs of all attention heads and performing linear layer processing, a local fine-grained node representation is obtained, the expression of which is:

[0043] In the formula, This represents the local fine-grained node representation of node u. This indicates a splicing operation. This represents the weighted sum of the value vectors output by the Hth attention head. This indicates a linear layer.

[0044] S5. Construct heterogeneous hypergraph structures through hypergraph branching, and capture high-order node representations of the global topology in the heterogeneous hypergraph structure, including: A heterogeneous hypergraph is constructed based on the original relationship graph, where the vertex set of the heterogeneous hypergraph is consistent with the vertex set of the semantic enhancement hypergraph of the large model. The heterogeneous hypergraph includes user-item interaction hyperedges, item semantic hyperedges, and user semantic hyperedges. Each hyperedge can connect two or more vertices, thus naturally capturing group relationships between multiple nodes. The heterogeneous hypergraph contains three types of hyperedges, each capturing different structural and semantic associations.

[0045] These hyperedges aim to model a user's real-world behavioral history holistically. They treat each user's complete interaction history as an indivisible whole, rather than a series of independent interaction events. Specifically, for each user u, a hyperedge is constructed consisting of the user node u and the set of all items it has interacted with. The formal definition of the user-item interaction hyperedge is as follows:

[0046] In the formula, This represents the user-item interaction hyperedge. It represents a set of user nodes; this hyperedge "binds" users' personal preferences together, effectively treating users' historical behavior as a high-order, holistic signal.

[0047] To capture deep, content-driven relationships between items (e.g., functional substitutability or stylistic complementarity), a hyperedge is constructed for each item i, consisting of item node i and its most relevant element in the semantic space. The set of neighbor nodes constitutes the semantic hyperedge of an item; the formal definition of the semantic hyperedge of an item is as follows:

[0048]

[0049] In the formula, This indicates that item i is the most relevant in the semantic space. A set of neighboring nodes, Indicates taking the first The maximum value, The semantic embedding features of item i are represented. The semantic embedding features of neighboring item node j are represented. Indicates the semantic hyperedge of an item. This represents a set of item nodes. Such hyperedges can reveal content-based product collections, such as "a series of classical music albums suitable for meditation" or "a set of Nordic-style furniture with similar design language".

[0050] To identify user communities with shared tastes and similar preferences, a hyperedge is constructed for each user u, consisting of user u and its most relevant nodes in the semantic space. It consists of _ neighbor nodes; the formal definition of the user semantic hyperedge is:

[0051]

[0052] In the formula, This indicates that user node u is the most relevant in the semantic space. 1 neighboring node, The semantic embedding features of user node u are represented. This represents the semantic embedding features of the neighboring user node v. Represents a user semantic hyperedge.

[0053] Ultimately, this heterogeneous hypergraph H, which integrates real-world interactions, object semantic clusters, and user communities, constitutes an extremely information-rich multi-relational topology. It will serve as input to the subsequent Heterogeneous Hypergraph Neural Network (HGNN) to learn and propagate higher-order dependencies from a global perspective.

[0054] Hypergraph Neural Networks (HGNNs) utilize the hyperedge connections in heterogeneous hypergraphs to aggregate multiple neighbor information and capture high-order node representations of the global topology. The core of HGNNs lies in a two-stage message passing process: first, information convergence from vertex to hyperedge, and then information distribution from hyperedge to vertex.

[0055] S6. Align local fine-grained node representations with high-order node representations of the global topology through comparative learning.

[0056] The steps described above modeled nodes from two different perspectives: local fine-tuning and global consensus. However, there is a lack of a clear and direct supervision signal to ensure that the learned representations are semantically consistent and aligned. Two branches may each learn valid representations, but there may be deviations or inconsistencies between them, which poses challenges to the subsequent fusion module.

[0057] To address this issue and introduce stronger self-supervised signals to the model, a two-view contrastive alignment module was designed. Its main purpose is to align the embedded representations from the two branches before they enter the final fusion layer. By adding additional supervisory information through contrastive learning, the local representation of the T-Branch (transformer branch) becomes aware of the macroscopic characteristics of the global topology, and conversely, the global representation of the H-Branch (hypergraph branch) learns the subtle differences in the local context.

[0058] The local representations obtained in T-Branch and the global representations obtained in H-Branch are treated as two different but complementary views about the same node. A cross-view contrastive learning paradigm is used to supervise the representation learning of these two views.

[0059] Its contrastive learning loss is specifically as follows:

[0060] In the formula, This represents the contrastive learning loss value. This indicates the number of nodes, where j and k are both node indices. This represents an exponential function with base e. This represents the local fine-grained node representation of node v. A higher-order node representation of the global topology of node v. This indicates temperature hyperparameters. This represents the local fine-grained node representation of node j. A higher-order node representation of the global topology of node k.

[0061] The goal of this contrastive loss function is to enhance the model's ability to distinguish positive sample pairs by amplifying the relative similarity of positive sample pairs. The loss function is designed to maximize the similarity of positive sample pairs while considering the similarity of negative sample pairs, thereby optimizing the learning effect of the embedding.

[0062] S7. The local fine-grained node representation of each aligned node and the high-order node representation of the global topology are concatenated to obtain the fusion features, including user fusion features and item fusion features. S8. Generate a personalized recommendation list of items based on fusion features.

[0063] In this invention, the above steps are implemented using an anti-information cocoon recommendation system that integrates large model semantics and hypergraph structure, including a preprocessing module, a semantic representation initialization module, an asymmetric bi-branch module, and a contrast alignment and prediction module.

[0064] The preprocessing module is used to obtain item attribute information and user interaction information between the item and the user, including the text description of the item and user comments. It is also used to construct the original relationship graph based on the item attribute information and user interaction information between the user and the item. The semantic representation initialization module is used to extract item semantic embedding features and user semantic embedding features based on the item's text description and user comments; The asymmetric dual-branch module is used to model user behavior sequences through transformer branches, capturing local fine-grained node representations, and also to construct semantically enhanced hypergraph structures through hypergraph branches, capturing high-order node representations of the global topology. The alignment and prediction module is used to concatenate the local fine-grained node representation of each aligned node with the high-order node representation of the global topology to obtain fused features; it is also used to generate a personalized recommendation list of items based on the fused features.

[0065] After joint optimization, this system is used to generate a personalized item recommendation list. The specific method of joint optimization is as follows: Calculate the inner product of the item fusion features and the user fusion features to obtain the user's predicted preference score for the item, expressed as follows:

[0066] In the formula, Let be the predicted preference score for user u on item i. For the fusion features of user node u, This is a local fine-grained node representation of user node u. A high-order node representation of the global topology of user node u. For the fusion features of item i, Let i be a local fine-grained node representation of item node i. The higher-order node representation of the global topology of item node i; The expression for reconstructing historical data using BPR loss is:

[0067] In the formula, This is the BPR loss value. For a set of nodes, For the sigmoid function, The predicted preference score for user u on item j; The system parameters are optimized by training the system in a manner that minimizes the joint loss. The specific expression for the joint loss is:

[0068] In the formula, For the joint loss value, To balance the importance of recommendation and alignment tasks, To control the hyperparameters of regularization strength, This represents all learnable parameters in the system.

[0069] In this way, not only is the accuracy of recommendation ranking optimized, but the consistency of its internal representation is also optimized within a self-supervised framework, thereby obtaining more robust and generalizable node embeddings.

[0070] To verify the beneficial effects of this invention, comparative experiments were conducted on three widely used large-scale benchmark datasets (Amazon Book, Amazon Music, and Amazon Movies). The proposed method was compared with representative baseline models, including: NGCF: a classic GNN model that explicitly injects collaborative signals into the embedding process; Light GCN: a strong baseline in current GNN collaborative filtering models, improving performance by simplifying GCN design; DHCF: a representative work that purely utilizes hypergraph structures for collaborative filtering; KGAT: a classic method for knowledge graph-enhanced recommendation, used to compare with LLM-based knowledge injection methods; LLM-Rec: a recent work using LLM for sequential recommendation; SHT: a self-supervised model combining hypergraphs and Transformers; and TransGNN: a model that interleaved and fused GNNs and Transformers. The comparative experimental results are shown in Table 1.

[0071] Table 1

[0072] Table 1 shows HyperTRLLM, representing the method proposed in this invention. The experimental results clearly demonstrate that this invention achieves state-of-the-art performance across all datasets and evaluation metrics. Specifically, compared to the best-performing baseline model, this invention achieves average performance improvements of 5.50% and 3.35% on Recall@20 and NDCG@20, respectively. This significant advantage is primarily attributed to the invention's effective utilization of a large language model to distill high-value semantic knowledge from text and its successful alignment of this semantic information with user behavior data through a sophisticated two-branch architecture. The Transformer branch meticulously characterizes the micro-level local context, while the hypergraph branch models the macro-level global relationships; their synergistic effect ultimately generates higher-quality user and item embeddings.

[0073] In contrast, various baseline models exhibit their respective limitations: GNN-based methods (NGCF, LightGCN) excel at capturing higher-order connectivity, but their fixed graph structure makes it difficult to fully utilize rich semantic information and limits their perception of the importance of dynamic neighborhoods. Hypergraph-based methods (DHCF), while capable of modeling group associations, are constrained in their ability to convey user preferences from abstract graph structures when faced with complex textual semantics. Auxiliary information-based methods (KGAT), while introducing external knowledge, often struggle to construct and utilize knowledge graphs as flexibly and deeply as LLMs to understand complex semantic relationships between items. LLM-based methods (LLM-Rec) fully utilize the world knowledge of LLMs, but are lacking in structured user behavior modeling and fail to fully leverage collaborative signals. Hybrid architecture methods (SHT, TransGNN), while attempting to combine graphs and Transformers, fail to reach the depth of semantic and structural information fusion achieved in this invention through summarization-reasoning and two-branch collaborative distillation.

[0074] Meanwhile, to further explore the contributions of each core component in the system provided by this invention, a series of ablation experiments were conducted, as shown in Table 2. The ablation experiments included three key variants: w / o (LLM replaces the large language model with the BERT encoder), w / o H-Branch (removing the Hypergraph branch), and w / o T-Branch (removing the Transformer branch).

[0075] Table 2

[0076] Experimental results show that HyperTRLLM with / o LLM variants exhibits the largest performance drop, decisively demonstrating the core value of the summarization-reasoning paradigm. It reveals that the contribution of LLM goes far beyond providing a powerful text embedding; it lies in its ability to proactively distill and construct structured knowledge highly aligned with recommendation tasks.

[0077] The performance of HyperTRLLM w / o H-Branch variant also showed a significant decline. This result clearly demonstrates that relying solely on T-Branch for fine-grained modeling of local neighborhoods encounters a performance bottleneck, and the model fails to capture complex relationships spanning multiple hops. These findings suggest that the global high-order dependencies captured by H-Branch are a key component of the model's overall performance.

[0078] Correspondingly, the performance of the HyperTRLLM w / o T-Branch variant is also significantly lower than the full model. This phenomenon reveals the inherent limitations of relying solely on global structure for information aggregation. While HGNN can effectively identify macroscopic structures, it struggles to distinguish the fine-grained relative importance of neighbors within hyperedges. In contrast, T-Branch, through semantic ordering and rotational position encoding, can inject high-fidelity local context. This result confirms that optimal representation learning is achieved through a combination of macroscopic global associations and microscopic local fine-tuning.

[0079] Furthermore, to intuitively evaluate the actual impact of the proposed dual-view contrast alignment on representation learning quality, a visualization experiment was conducted. Following visualization conventions in the field of representation learning, singular value decomposition was used to project the high-dimensional embedding vectors learned by each model for item nodes on the Amazon-Book dataset into a two-dimensional space. By observing the distribution pattern of the embedding points, insights can be gained into the model's ability to capture relationships between items and the characteristics of its learned representations. Figure 3 As shown, by comparing HyperTRLLM with its variant HyperTRLLM-noALIGN (which removes the contrast alignment module), the crucial role of the contrastive learning mechanism can be clearly seen. The embedding points generated by the complete HyperTRLLM model are more evenly and fully distributed in the two-dimensional space, indicating that its learned representation has better isotropy and can make fuller use of the entire embedding space, thus giving each node a more discriminative and unique representation. Conversely, the embedding points of HyperTRLLM-noALIGN exhibit a certain degree of clustering and collapse, indicating that without the regularization constraint of contrast alignment, the discriminativeness of the model's learned representation decreases, and the representations of different nodes are not sufficiently dispersed in the space.

[0080] The visualization results of LightGCN show a narrow, banded distribution, suggesting that the model may be oversmoothed to some extent, leading to convergence of node representations in certain dimensions. In contrast, KGAT's embedding distribution is relatively more balanced, showing that its ability to capture item relationships is enhanced after incorporating knowledge graph information. However, compared to HyperTRLLM, its distribution uniformity and coverage still lag behind. DHCF's deep models tend to overfit noisy data. In summary, HyperLLM demonstrates the best representation learning performance, showing good results in learning user and item embeddings. Other models such as HyperLLM-noALIGN and TransGNN also show good clustering results, but slightly inferior to HyperLLM. LightGCN and KGNN may show bias in some cases, but overall they still perform well.

[0081] In summary, this invention overcomes the limitations of traditional recommendation methods that treat users and items as anonymous IDs. Compared to collaborative filtering methods that rely solely on interaction history, this invention can deeply understand the semantics of content and the essence of user preferences, performing particularly well in cold start and data-sparse scenarios. For example, for long-tail items like "Introduction to Philosophy" with only one interaction in the training set, the traditional method Light GCN, lacking semantic understanding, can only recommend popular items, while this invention can accurately recommend highly semantically relevant professional books such as "History of Philosophical Thought" and "Foundations of Philosophy."

Claims

1. A method for recommending anti-information cocoons that integrates large model semantics and hypergraph structure, characterized in that, include: Item attribute information and user interaction information with items are obtained through large language model inference, including the text description of items and user comments; The original relationship graph is constructed based on the item attribute information and the user's interaction information with the item; Extract item semantic embedding features and user semantic embedding features based on the text description of the item and user reviews; Modeling user behavior sequences through transformer branches captures local fine-grained node representations; Heterogeneous hypergraph structures are constructed by hypergraph branching, and high-order node representations of the global topology in the heterogeneous hypergraph structure are captured. By comparing and learning, we can align the local fine-grained node representations with the high-order node representations of the global topology. The local fine-grained node representation of each aligned node is concatenated with the high-order node representation of the global topology to obtain the fusion features, including user fusion features and item fusion features. A personalized list of recommended items is generated based on fused features.

2. The method according to claim 1, characterized in that, Specific methods for extracting semantic embedding features of items and semantic embedding features of users based on the text description of items and user reviews include: For each item, multiple independent text instances are constructed, each text instance including the item's title, category, and a specific text content; Input text instances into a large language model to generate corresponding summary-inference text; For each user’s historical comments and their corresponding item information, an independent preference instance is constructed for each of the user’s comments. Each preference instance includes the comment content and its corresponding item title and category. Input preference instances into a large language model to generate corresponding preference-inference text; The summary-inference text and preference-inference text are encoded into fixed-dimensional semantic vectors and preference vectors, respectively; Average pooling is performed on the semantic vector and preference vector respectively to obtain the semantic embedding features of items and the semantic embedding features of users.

3. The method according to claim 1, characterized in that, The specific method for modeling user behavior sequences and capturing local fine-grained node representations using transformer branching is as follows: Obtain the embedded features of user-item interactions; By concatenating the interaction embedding features and semantic embedding features of each node, a hybrid representation is obtained, the expression of which is: In the formula, This represents a hybrid representation of node v. Represents the interactive embedding feature matrix Middle node The row vector; Represents the semantic embedding feature matrix Middle node The row vectors, and the semantic embedding feature matrix includes user semantic embedding features and item semantic embedding features; Indicates a splicing operation; Based on the hybrid representation, cosine similarity is used as the core metric function to select the most relevant neighbor node for each node; Aggregate the neighboring nodes of each center point to generate a local fine-grained node representation.

4. The method according to claim 3, characterized in that, Based on hybrid representation, pre-similarity is used as the core metric to select the most relevant neighboring nodes for each node, including: The expression for calculating the association strength between any two nodes is: In the formula, Represents a node and nodes The strength of the correlation between them The norm is denoted by T, and the superscript T denotes the transpose of the matrix. The neighborhood of each central node is selected based on the strength of the association between nodes, and the expression is as follows: In the formula, Represents nodes The K most similar neighbor nodes, argtop-K represents taking the K largest values. Indicates the node The set of all nodes except itself.

5. The method according to claim 4, characterized in that, Aggregate the neighboring nodes of each center point to generate a local fine-grained node representation, including: Based on the association strength between the central node and its neighboring nodes, the neighboring nodes of each central node are sorted in descending order to obtain an ordered local structure sequence, the expression of which is: In the formula, For the ordered local structure sequence of the central node u, This represents the Kth neighbor node of the central node u; The obtained ordered local structure sequence is encoded using a transformer with rotational position encoding, and the weighted sum of the value vectors output by the attention head is used as follows: The query vector, key vector, and value vector of each attention head are projected, and their expression is: In the formula, This represents the query vector for the h-th attention head. This represents the hybrid representation of node u. This represents the query weight matrix for the h-th attention head. This represents the key vector of the h-th attention head. This represents the mixed representation of the j-th neighbor node. This represents the key weight matrix of the h-th attention head. The weight matrix represents the value of the h-th attention head; The attention score is calculated using the following expression: In the formula, This represents the attention score of the h-th attention head. This represents the rotation position encoding matrix applied at the center node. This represents the rotation position encoding matrix applied at node j. The dimension representing the key; The attention weights are calculated using the following expression: In the formula, This represents the attention weight of the h-th attention head. express Activation function; The weighted sum of the value vectors is expressed as follows: In the formula, This represents the weighted sum of the value vectors output by the h-th attention head; By concatenating the outputs of all attention heads and performing linear layer processing, a local fine-grained node representation is obtained, the expression of which is: In the formula, This represents the local fine-grained node representation of node u. This indicates a splicing operation. This represents the weighted sum of the value vectors output by the Hth attention head. This indicates a linear layer.

6. The method according to claim 1, characterized in that, Heterogeneous hypergraph structures are constructed through hypergraph branching, capturing high-order node representations of the global topology within these structures, including: A heterogeneous hypergraph is constructed based on the original relation graph, wherein the vertex set of the heterogeneous hypergraph is consistent with the vertex set of the semantic enhancement hypergraph of the large model. The heterogeneous hypergraph includes user-item interaction hyperedges, item semantic hyperedges and user semantic hyperedges. By using a hypergraph neural network to aggregate neighbor information through the hyperedge connections in a heterogeneous hypergraph, a high-order node representation of the global topology can be captured.

7. The method according to claim 6, characterized in that, The formal definition of the user-item interaction hyperedge is as follows: In the formula, The user-item interaction hyperedge is represented by the user node u and the set of all items it has interacted with. constitute; Represents a set of user nodes; The formal definition of the semantic hyperedge of an item is as follows: In the formula, This indicates that item i is the most relevant in the semantic space. A set of neighboring nodes, Indicates taking the first The maximum value, The semantic embedding features of item i are represented. The semantic embedding features of neighboring item node j are represented. The semantic hyperedge represents an item, consisting of item node i and its most relevant element in the semantic space. It consists of a set of neighboring nodes; Represents a set of item nodes; The formal definition of a user semantic hyperedge is: In the formula, This indicates that user node u is the most relevant in the semantic space. One neighboring node, The semantic embedding features of user node u are represented. This represents the semantic embedding features of the neighboring user node v. Representing a user semantic hyperedge, consisting of user node u and its most relevant node in the semantic space. It consists of several neighboring nodes.

8. The method according to claim 4, characterized in that, By aligning the local fine-grained node representations with the high-order node representations of the global topology through contrastive learning, the specific contrastive learning loss is as follows: In the formula, This represents the contrastive learning loss value. This indicates the number of nodes, where j and k are both node indices. This represents an exponential function with base e. This represents the local fine-grained node representation of node v. A higher-order node representation of the global topology of node v. This indicates temperature hyperparameters. This represents the local fine-grained node representation of node j. A higher-order node representation of the global topology of node k.

9. A system for implementing the anti-information cocoon recommendation method that integrates large model semantics and hypergraph structure as described in any one of claims 1 to 8, characterized in that, include: The preprocessing module is used to obtain item attribute information and user interaction information between the item and the user, including the text description of the item and user comments. It is also used to construct the original relationship graph based on the item attribute information and user interaction information between the user and the item. The semantic representation initialization module is used to extract item semantic embedding features and user semantic embedding features based on the item's text description and user comments; The asymmetric dual-branch module is used to model user behavior sequences through transformer branches, capturing local fine-grained node representations, and also to construct semantically enhanced hypergraph structures through hypergraph branches, capturing high-order node representations of the global topology. The alignment and prediction module is used to concatenate the local fine-grained node representation of each aligned node with the high-order node representation of the global topology to obtain fused features. It is also used to generate personalized recommendation lists of items based on fused features.

10. The system according to claim 9, characterized in that, After joint optimization, this system is used to generate a personalized item recommendation list. The specific method is as follows: Calculate the inner product of the item fusion features and the user fusion features to obtain the user's predicted preference score for the item, expressed as follows: In the formula, Let be the predicted preference score for user u on item i. For the fusion features of user node u, This is a local fine-grained node representation of user node u. A high-order node representation of the global topology of user node u. For the fusion features of item i, Let i be a local fine-grained node representation of item node i. The higher-order node representation of the global topology of item node i; The expression for reconstructing historical data using BPR loss is: In the formula, This is the BPR loss value. For a set of nodes, For the sigmoid function, The predicted preference score for user u on item j; The system parameters are optimized by training the system in a manner that minimizes the joint loss. The specific expression for the joint loss is: In the formula, For the joint loss value, To balance the importance of recommendation and alignment tasks, To control the hyperparameters of regularization strength, This represents all learnable parameters in the system.

Images