Few-Shot Representation Learning for Knowledge Graph with Variational Auto-encoder Data Augmentation

Few-shot in knowledge graph (KG) is a widespread and common phenomenon. It means only seldom relations in knowledge graphs have enough related triples, which brings difficulties to representation learning due to the lack of training samples. To overcome these limitations in scenarios where training triples are sparse, we introduce a novel Few-shot knowledge graph Representation Learning method based on entity contextual data augmentation with Variational Auto-Encoder (VAE) named FRL-VAE. The method obtains the representation of entities in triples by aggregating contextual information firstly. Then, VAE is used to extract latent features in entities and contextual information and generate representations of new samples highly similar to the original ones for training. We conduct link prediction experiments on two public datasets, illustrating the leading performance, effectiveness, and stability of FRL-VAE in few-shot scenario. Furthermore, we empirically underscore the significance of the data augmentation module and highlight its promising applications in knowledge graph representation learning.