An edge enhancement graph neural network model with node discrimination for knowledge graph representation learning

Tao Wang, Bo ShenComplex 2025 · DOI

TL;DR

The NDEE-GNN model improves Knowledge Graph (KG) embeddings by addressing two common oversights in Graph Neural Networks: the failure to distinguish between the semantic roles of source versus destination nodes, and the lack of context-aware edge representations. By implementing a "Node Discrimination" aggregation strategy and an "Edge Enhancement" mechanism, the model achieves state-of-the-art results in link prediction across several benchmark datasets.

What Problem Does This Paper Address?

Knowledge Graphs (KGs) represent information as triples—specifically (head entity, relation, tail entity). While Graph Neural Networks (GNNs) are powerful tools for learning these representations, the authors identify two major gaps in current approaches:

  1. Indiscriminate Aggregation: Most GNNs treat all neighboring nodes the same way during information fusion. However, in a directed KG, being a "source" (head) is semantically different from being a "destination" (tail). Ignoring this distinction leads to feature redundancy and a failure to capture the directional logic of relations.
  2. Static Edge Representations: Edges (relations) are often initialized randomly or via simple linear transforms and remain relatively static. They don't "learn" from the specific entities they connect, missing out on the rich structural context provided by the graph's topology.

Methods

The authors propose the Edge Enhancement GNN model with Node Discrimination (NDEE-GNN), which introduces two novel components:

  • Node Discrimination IA (Information Aggregation): Instead of one-size-fits-all neighbor fusion, the model uses separate attention-based mechanisms for source nodes and destination nodes. This allows the model to weigh the impact of a neighbor differently based on its role in a triple.
  • Edge Enhancement IA: This mechanism allows edge representations to evolve by aggregating information from their adjacent source and destination nodes. This ensures that a relation like "educated_at" carries different structural nuances depending on the types of entities it connects.
  • GNN Encoder & Decoder: These components are stacked in a multi-layer GNN. For the final "link prediction" task (predicting a missing head or tail), the model uses ConvE (a 2D convolutional neural network) as the decoder to score the plausibility of triples.

Datasets

The model was evaluated on six widely used real-world benchmarks:

  • FB15k & FB15k-237: Subsets of Freebase containing general-world facts.
  • WN18 & WN18RR: Subsets of WordNet, a lexical database of English words.
  • UMLS: A specialized medical semantic dataset.
  • Kinship: A dataset representing family relationships.

Key Insights and Ideas

  • The "Source" Advantage: One of the most interesting findings is that aggregating information solely from source nodes often yields better performance than aggregating from destination nodes or even a combination of both. This is particularly true for "N-to-1" relations, where knowing the head entity provides more discriminative power for predicting the tail.
  • Contextual Edges: By allowing edges to "absorb" information from the nodes they connect, the model creates more expressive relation embeddings. This helps the model handle complex mapping properties like many-to-many (N-to-N) connections more effectively.
  • Efficiency through Selection: The authors found that selectively aggregating node types (e.g., focusing on source nodes) can reduce training time by up to one-third (from 24 hours down to 16 hours on FB15k-237) while actually improving accuracy.

Key Results

NDEE-GNN consistently outperformed or matched top-tier baseline models (like RotatE, PairRE, and CompGCN):

  • FB15k: Achieved a Mean Reciprocal Rank (MRR) of 0.813, surpassing the previous best of 0.811.
  • WN18RR: Reached an MRR of 0.495, edging out QuatRE (0.493).
  • UMLS & Kinship: The model showed significant dominance on these smaller, more structured datasets, achieving an MRR of 0.942 on UMLS and 0.862 on Kinship.
  • Ablation Studies: Removing the "Edge Enhancement" or "Node Discrimination" components led to a measurable drop in performance, confirming that both innovations are necessary for the model's success.

Limitations and Future Work

  • Over-smoothing: Like many GNN-based models, NDEE-GNN is susceptible to over-smoothing if too many layers are used. The authors found that 1–3 layers are usually the "sweet spot" before node representations become too similar to be useful.
  • Mean Rank (MR) Sensitivity: While the model excels in MRR and Hits@N, it doesn't always achieve the best Mean Rank. The authors attribute this to the nature of GNNs, which can be sensitive to outliers in the prediction ranking list.
  • Future Directions: The authors suggest integrating Large Language Models (LLMs) to provide better initial semantic embeddings for entities and relations, and exploring bidirectional information aggregation to further strengthen structural associations.

Summary updated

Abstract

The vectorized representation of a knowledge graph is essential for effectively utilizing its implicit knowledge. Graph neural networks (GNNs) are particularly adept at learning graph representations due to their ability to handle graph topologies. However, GNN-based approaches face two main challenges: first, they fail to differentiate between the types of adjacent nodes during the information aggregation process; second, the edge representations lack relational semantic information and fail to capture the characteristics of adjacent nodes. Conventional methods typically treat source and destination nodes as identical, ignoring the distinct information that arises from different node types. This results in a failure to accurately capture the various semantic features, leading to feature redundancy. Additionally, many existing methods derive edge representations through random initialization or linear transformations, which do not adequately reflect relational semantics and adjacent node information, resulting in ineffective edge representations.To address these limitations, we propose the Edge Enhancement GNN model with Node Discrimination (NDEE-GNN). This model establishes node discrimination information aggregation mechanisms for source and destination nodes, allowing for a deeper investigation into the impact of various adjacent node types. It also employs a specially designed information aggregation mechanism for each edge, incorporating relation and adjacent node features. Experimental results across multiple real-world datasets demonstrate that by discriminating node types and enhancing edge representations, NDEE-GNN can accurately capture and represent complex associations between entities and relations, significantly improving link prediction performance and outpacing other baseline models.
BibTeX
@inproceedings{wang2025edge,
  author = {Tao Wang and Bo Shen},
  title = {{An edge enhancement graph neural network model with node discrimination for knowledge graph representation learning}},
  booktitle = {Complex},
  year = {2025},
  doi = {10.1007/s40747-025-01860-6},
}
Figure