AHyTE: An Adaptive Hyperplane-based Method for Time-aware Knowledge Graph Embedding

Abstract

Knowledge graph (KG) embedding can learn the representations of entities and relations from a KG in low-dimensional continuous vector space, which is the foundation of knowledge acquisition and reasoning. Time-aware KG embedding is a hot research branch of KG embedding, which introduces time information while considering the knowledge expression ability to make knowledge dynamic. Among them, t-TransE model is the most classic. However, t-TransE has the following two flaws: one is that it only regards time information as a constraint and the embedding learned by this way is not explicitly temporally aware; the other is that it uses traditional Euclidean distance as a metric, and all feature dimensions are treated identically, hence, the model may not accurately reflect the priority of features. Currently, there has not been a single method resolves the flaws simultaneously, so we propose an adaptive hyperplane-based time-aware knowledge graph embedding model, namely, AHyTE. AHyTE associates each time stamp with the corresponding hyperplane, and introduces a diagonal weight matrix to assign weights to each feature dimension. Through link prediction experiments on the time data set extracted from the real world, we demonstrate the remarkable improvement of AHyTE.