Inductive link prediction based on directed subgraphs reasoning
TL;DR
DLP is more effective than state-of-the-art baselines on inductive link prediction and an integration of an LSTM and Transformer architecture is presented to improve the node representations and acquire the order of dependencies and the whole context.
Abstract
Inductive link prediction is the problem of inferring about new entities and relations present in dynamic knowledge graphs. Traditional approaches typically depend on Personalized PageRank to select subgraphs but do not consider the directional nature of graphs and cannot take care of asymmetrical links. Also, traditional GNN-based communication of closed subgraphs has limited functionality when dealing with complex relationships between nodes. To solve these issues we suggest the use of directed subgraph reasoning model. The addition of in/out-degrees information is used as a method to mine local directed subgraphs that preserve the structural directionality and minimize noise by employing Directed Personalized PageRank. Moreover, an integration of an LSTM and Transformer architecture is presented to improve the node representations and acquire the order of dependencies and the whole context. The experimental results with public benchmark datasets have shown that DLP is more effective than state-of-the-art baselines on inductive link prediction.