CirE: Circular Embeddings of Knowledge Graphs

摘要

The embedding representation technology provides convenience for machine learning on knowledge graphs (KG), which encodes entities and relations into continuous vector spaces and then constructs (entity,relation,entity) triples. However, KG embedding models are sensitive to infrequent and uncertain objects. Furthermore, there is a contradiction between learning ability and learning cost. To this end, we propose circular embeddings (CirE) to learn representations of entire KG, which can accurately model various objects, save storage space, speed up calculation, and is easy to train and scalable to very large datasets. We have the following contributions: (1) We improve the accuracy of learning various objects by combining holographic projection and dynamic learning. (2) We reduce parameters and storage by adopting the circulant matrix as the projection matrix from the entity space to the relation space. (3) We reduce training time through adaptive parameters update algorithm which dynamically changes learning time for various objects. (4) We speed up the computation and enhance scalability by fast Fourier transform (FFT). Extensive experiments show that CirE outperforms state-of-the-art baselines in link prediction and entity classification, justifying the efficiency and the scalability of CirE.