VN Network: Embedding Newly Emerging Entities with Virtual Neighbors

TL;DR

This work tackles inductive knowledge graph completion for newly emerging entities by proposing VN network, a framework that creates virtual neighbors through rule-guided inferences and long-distance symmetric path rules. It combines a rule-based virtual neighbor predictor with a structure- and query-aware GNN encoder and a DistMult-style decoder, training via a joint objective that leverages both hard facts and soft, rule-consistent soft labels. The approach iteratively refines embeddings and rule predictions, addressing neighbor sparsity and enabling long-range patterns to inform unseen entities. Empirical results on WordNet11, FB15K, and YAGO37 show state-of-the-art performance in triple classification and link prediction, with particular robustness as the fraction of unseen entities grows, demonstrating practical utility for evolving knowledge graphs.

Abstract

Embedding entities and relations into continuous vector spaces has attracted a surge of interest in recent years. Most embedding methods assume that all test entities are available during training, which makes it time-consuming to retrain embeddings for newly emerging entities. To address this issue, recent works apply the graph neural network on the existing neighbors of the unseen entities. In this paper, we propose a novel framework, namely Virtual Neighbor (VN) network, to address three key challenges. Firstly, to reduce the neighbor sparsity problem, we introduce the concept of the virtual neighbors inferred by rules. And we assign soft labels to these neighbors by solving a rule-constrained problem, rather than simply regarding them as unquestionably true. Secondly, many existing methods only use one-hop or two-hop neighbors for aggregation and ignore the distant information that may be helpful. Instead, we identify both logic and symmetric path rules to capture complex patterns. Finally, instead of one-time injection of rules, we employ an iterative learning scheme between the embedding method and virtual neighbor prediction to capture the interactions within. Experimental results on two knowledge graph completion tasks demonstrate that our VN network significantly outperforms state-of-the-art baselines. Furthermore, results on Subject/Object-R show that our proposed VN network is highly robust to the neighbor sparsity problem.

Figures (5)

• Figure 1: The average ratio of neighbors to predicted facts with and without virtual neighbors.
• Figure 2: The example of the unseen entities problem about original KG.
• Figure 3: The framework of the VN network. VN network contains three components: the rule-based virtual neighbor prediction, the GNN-based encoder to capture structure information and embed entities, the decoder to calculate the probability of edges and then refine the model with soft labels.
• Figure 4: The example of the the symmetric path rule. P is person. R is role. F is film. And T is type. Note that for symmetric paths, we only care about the type and direction of the edges, and the letters of the nodes in the graph are just to illustrate the underlying meaning of them.
• Figure 5: Link prediction results on Subject/Object-R.