Message Intercommunication for Inductive Relation Reasoning

TL;DR

This work proposes a novel GraIL-based inductive relation reasoning model, termed MINES, by introducing a Message Intercommunication mechanism on the Neighbor-Enhanced Subgraph, designed to capture the omitted hidden mutual information.

Abstract

Inductive relation reasoning for knowledge graphs, aiming to infer missing links between brand-new entities, has drawn increasing attention. The models developed based on Graph Inductive Learning, called GraIL-based models, have shown promising potential for this task. However, the uni-directional message-passing mechanism hinders such models from exploiting hidden mutual relations between entities in directed graphs. Besides, the enclosing subgraph extraction in most GraIL-based models restricts the model from extracting enough discriminative information for reasoning. Consequently, the expressive ability of these models is limited. To address the problems, we propose a novel GraIL-based inductive relation reasoning model, termed MINES, by introducing a Message Intercommunication mechanism on the Neighbor-Enhanced Subgraph. Concretely, the message intercommunication mechanism is designed to capture the omitted hidden mutual information. It introduces bi-directed information interactions between connected entities by inserting an undirected/bi-directed GCN layer between uni-directed RGCN layers. Moreover, inspired by the success of involving more neighbors in other graph-based tasks, we extend the neighborhood area beyond the enclosing subgraph to enhance the information collection for inductive relation reasoning. Extensive experiments on twelve inductive benchmark datasets demonstrate that our MINES outperforms existing state-of-the-art models, and show the effectiveness of our intercommunication mechanism and reasoning on the neighbor-enhanced subgraph.

Figures (8)

• Figure 1: Illustration of transductive and inductive relation reasoning. In the transductive scenario, entities in test graphs are all seen in the model during training. While as for the inductive scenario, unseen entities may exist in test graphs.
• Figure 2: Limitations of existing GraIL-based models. The differences between ideal scenarios (right figures) and scenarios in previous models (left figures) are colored in orange.
• Figure 3: The framework of the proposed MINES. The framework includes four main steps: neighbor-enhanced subgraph extraction, entity labeling $\&$ embedding initialization, message intercommunication, and triplet scoring. Our main contribution lies in the first and the third step. Precisely, in the first step, we extract the neighbor-enhanced subgraph by adding neighboring entities and edges (colored in orange) to the enclosing subgraph (colored in black). In the previous work, only the black entities and relations are included in the subgraph for reasoning; In the third step, we feed subgraphs into the novel message intercommunication module to learn representations. In this step, GCN is integrated with RGCN to achieve better information interactions. Note that the 1-hop subgraph is taken as an example for the illustration.
• Figure 4: Comparison of the parallel and sequential intercommunication frameworks.
• Figure 5: Comparison of the parameter numbers of 3-layer MINES and prototype GraIL for training on benchmark datasets.