Abstract
Modern recommender systems increasingly rely on Knowledge Graphs (KGs) to enhance accuracy and interpretability. However, traditional models often suffer from geometric mismatch—using Euclidean space to represent hierarchical data—and signal confusion, where collaborative user-item signals are indistinguishably mixed with item attributes. This paper introduces the Hierarchy-Aware Knowledge Gated Network (HAKG), a novel framework that utilizes hyperbolic geometry to capture the hierarchical nature of interactions and KGs. By employing a dual-embedding design and a gated fusion mechanism, HAKG effectively separates and integrates collaborative signals with knowledge associations, achieving state-of-the-art performance across multiple benchmarks.
Introduction
Knowledge Graphs (KGs) serve as structured repositories of real-world facts, providing rich context for items in recommendation tasks. Despite their utility, existing KG-based recommendation models face two primary challenges. First, they typically embed data in Euclidean space, which is ill-suited for the power-law distributions and hierarchical structures inherent in user-item interactions and KG relations. Second, they often fail to distinguish between an item's intrinsic attributes (e.g., brand, category) and its collaborative signals (e.g., user preference patterns), leading to suboptimal representations.
The Geometric Challenge: Why Hyperbolic Space?
Most recommendation models operate in flat Euclidean space. However, data in recommender systems is rarely flat. User-item interactions often follow a power-law distribution where a small number of items receive the majority of interactions, creating a tree-like hierarchy. Similarly, KGs are naturally hierarchical, containing "is-a" or "belongs-to" relationships.
Euclidean space struggles to represent these structures without significant distortion unless the dimensionality is extremely high. In contrast, hyperbolic geometry—specifically the Poincaré ball model—expands exponentially as one moves away from the origin. This property provides the necessary "room" to embed hierarchical data with minimal distortion, naturally placing popular items near the origin and pushing the "long tail" of unpopular items toward the boundary.
The HAKG Framework
The HAKG framework addresses these challenges through three core innovations:
1. Hyperbolic Relation-Transitive Aggregation
To update item representations, HAKG aggregates information from neighbors within the KG. Because standard vector operations (like calculating a mean) are not directly applicable in hyperbolic space, the model maps data to a "tangent space"—a local Euclidean approximation. After performing aggregation in this tangent space, the data is mapped back to the hyperbolic manifold, preserving the geometric integrity of the embeddings.
2. Modeling Hierarchies with Angle Constraints
To explicitly model hierarchical relations, HAKG utilizes hyperbolic entailment cones. If an item possesses a specific attribute (such as a brand or category), its embedding is constrained to fall within a specific angular "cone" associated with that attribute. This geometric constraint ensures that the "belongs-to" relationship is mathematically enforced, improving the model's ability to capture categorical hierarchies.
3. Dual-Embedding and Gated Fusion
HAKG acknowledges that an item serves as a bridge between world knowledge and user behavior. Every item is assigned two distinct embeddings:
- Knowledge Embedding: Captures intrinsic properties from the KG.
- Collaborative Embedding: Captures patterns derived from user-item interactions.
A gated fusion unit, inspired by Gated Recurrent Units (GRUs), adaptively determines the weight of each signal for a specific user. This allows the model to discern whether a user is attracted to an item because of its specific attributes or because it is currently trending among similar users.
Experimental Evaluation
The model was evaluated on three diverse real-world datasets: Alibaba-iFashion (high hierarchy), Yelp2018 (local business attributes), and Last-FM (music recommendations). All datasets utilized a "10-core" setting to ensure data quality.
Performance Metrics
HAKG consistently outperformed state-of-the-art baselines, including KGIN, CKAN, and Hyper-Know. The relative improvements in NDCG@20 (Normalized Discounted Cumulative Gain) were significant:
- Alibaba-iFashion: 15.43% improvement
- Yelp2018: 8.21% improvement
- Last-FM: 9.79% improvement
Ablation and Visualization
Ablation studies confirmed that both the hyperbolic geometry and the gated fusion mechanism are critical to performance; removing either led to a substantial drop in accuracy. Visualization of the Poincaré ball confirmed that popular items cluster near the origin, while less popular items are distributed toward the boundary, validating the model's ability to handle power-law distributions.
Discussion and Insights
The success of HAKG suggests that the choice of geometric space acts as a powerful inductive bias. By moving to hyperbolic space, the model naturally accounts for the "long tail" of items. Furthermore, the distinction between hierarchical and non-hierarchical relations allows for more nuanced representational power.
Conclusion and Future Work
HAKG demonstrates the efficacy of hierarchy-aware hyperbolic embeddings in recommendation. While the model provides a significant performance boost, future work will focus on improving numerical stability near the Poincaré boundary and exploring automatic hierarchy detection to infer hierarchical relations when explicit metadata is unavailable.
Summary updated