UPGAT: Uncertainty-Aware Pseudo-neighbor Augmented Knowledge Graph Attention Network.

The uncertain knowledge graph (UKG) generalizes the representation of entity-relation facts with a certain confidence score. Existing methods for UKG embedding view it as a regression problem and model different relation facts independently. We aim to generalize the graph attention network and use it to capture the local structural information. Yet, the uncertainty brings in excessive irrelevant neighbor relations and complicates the modeling of multi-hop relations. In response, we propose UPGAT, an uncertainty-aware graph attention mechanism to capture the probabilistic subgraph features while alleviating the irrelevant neighbor problem; introduce the pseudo-neighbor augmentation to extend the attention range to multi-hop. Experiments show that UPGAT outperforms the existing methods. Specifically, it has more than 50% Weighted Mean Rank improvement over the existing approaches on the NL27K dataset.