Abstract
Collaborative Filtering (CF) based on graph neural networks (GNNs) has yielded immense success for recommendation systems by capturing high-order dependencies from implicit feedback. Recently, the outstanding text comprehension ability of the Large Language Models (LLMs) has shown promising potential to provide auxiliary semantics for collaborative representation. However, when aligning textual information with collaborative signals, inconsistent semantics between user-item and item-item text pairs may lead to the degradation of the alignment model, thus hindering the recommender system from effectively utilizing heterogeneous information. In this paper, we propose a novel method: Semantic Enhanced Heterogeneous Hypergraph Network (SEHHN), which enhances the representations of CF correlations with semantics, thereby avoiding alignment degradation. To better model the collaborative signals, we design a graph autoencoder that captures the bidirectional relationship between user preferences and item features in review semantics. Furthermore, we develop an LLM-based item classifier to adaptively exploit potential correlations of items via the co-occurrences of item features. Finally, we design a heterogeneous hypergraph network to achieve efficient alignment and propagation of heterogeneous information, thereby alleviating the impact of semantic inconsistency on CFs. Extensive experiments on three real-world datasets demonstrate that our proposed SEHHN outperforms existing SOTA methods and validates the effectiveness of each component.
Introduction
Collaborative Filtering (CF) remains the cornerstone of modern recommendation systems, primarily leveraging user-item interaction history to predict future preferences. While Graph Neural Networks (GNNs) have significantly advanced CF by modeling high-order connectivity, they often struggle to capture the nuanced "why" behind interactions—information typically buried in unstructured text like reviews and product descriptions.
The integration of Large Language Models (LLMs) offers a solution by providing rich semantic embeddings. However, a critical challenge arises: semantic inconsistency. As illustrated in Figure 1, a user's review might focus on a specific pain point (e.g., poor battery life), while the item's official description highlights a different strength (e.g., a high-quality camera). Forcing an alignment between these disparate semantic signals can lead to "alignment degradation," where the model fails to capture the true collaborative behavior.
Methodology: The SEHHN Framework
To address these challenges, we introduce the Semantic Enhanced Heterogeneous Hypergraph Network (SEHHN). This framework integrates textual semantics into the collaborative filtering process through three primary components.
1. LLM-based Item Classification
Rather than processing raw, noisy item descriptions directly, SEHHN employs an LLM (such as GPT-3.5) as a zero-shot classifier. By bucketing items into discrete, semantically meaningful categories, the model extracts high-level commonalities that drive item co-occurrence, effectively acting as a form of semantic regularization.
2. Graph Autoencoder (GAE) for Review Semantics
Reviews serve as the bridge between subjective user preferences and objective item features. We design a specialized GAE to project high-dimensional LLM review embeddings into a lower-dimensional collaborative space. The encoder compresses the semantic information, while the decoder reconstructs it using user and item embeddings. This bidirectional relationship ensures that the learned semantic representations are grounded in actual user-item interactions.
3. Heterogeneous Hypergraph Construction
Standard graphs are limited to pairwise relations. SEHHN utilizes a heterogeneous hypergraph to model complex, multi-node dependencies. We define four distinct hyperedge types:
- Interaction Hyperedges: Connect a user, an item, and their specific review.
- Observation Hyperedges: Link an item to all its associated reviews.
- Preference Hyperedges: Link a user to all reviews they have authored.
- Feature Hyperedges: Connect items to their LLM-inferred categories.
Experimental Evaluation
We evaluated SEHHN against several state-of-the-art baselines, including LightGCN and RLMRec, across three real-world datasets: Book (Amazon), Yelp, and Steam.
Key Results
- Superior Performance: SEHHN consistently outperformed all baselines. For instance, on the Book dataset, SEHHN achieved a Recall@20 of 0.1677, surpassing the previous best (RLMRec) at 0.1608.
- Ablation Studies: Removing either review semantics (SEHHN-R) or category semantics (SEHHN-C) resulted in performance drops, confirming that both subjective and objective semantic tracks are vital.
- Robustness: The framework demonstrated consistent gains regardless of the underlying LLM used for embeddings (e.g., GPT, BGE-large, Bloom), highlighting the robustness of the hypergraph architecture.
Conclusion
SEHHN effectively mitigates the issue of semantic inconsistency in recommendation systems by leveraging a heterogeneous hypergraph and a review-focused graph autoencoder. By grounding LLM-generated semantics in collaborative signals, SEHHN provides a more accurate and robust framework for modern CF. Future work will focus on optimizing the computational overhead of LLM processing and exploring dynamic hypergraph updates for real-time recommendation scenarios.
Presented at AAAI-25.
Summary updated