Hypernetwork Knowledge Graph Embeddings

21 Aug 2018  ·  Ivana Balažević, Carl Allen, Timothy M. Hospedales ·

Knowledge graphs are graphical representations of large databases of facts, which typically suffer from incompleteness. Inferring missing relations (links) between entities (nodes) is the task of link prediction. A recent state-of-the-art approach to link prediction, ConvE, implements a convolutional neural network to extract features from concatenated subject and relation vectors. Whilst results are impressive, the method is unintuitive and poorly understood. We propose a hypernetwork architecture that generates simplified relation-specific convolutional filters that (i) outperforms ConvE and all previous approaches across standard datasets; and (ii) can be framed as tensor factorization and thus set within a well established family of factorization models for link prediction. We thus demonstrate that convolution simply offers a convenient computational means of introducing sparsity and parameter tying to find an effective trade-off between non-linear expressiveness and the number of parameters to learn.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Link Prediction FB15k HypER MRR 0.790 # 11
Hits@10 0.885 # 11
Hits@3 0.829 # 7
Hits@1 0.734 # 7
Link Prediction FB15k-237 HypER MRR 0.341 # 36
Hits@10 0.520 # 38
Hits@3 0.376 # 28
Hits@1 0.252 # 27
Link Prediction WN18 HypER MRR 0.951 # 5
Hits@10 0.958 # 10
Hits@3 0.955 # 2
Hits@1 0.947 # 3
Link Prediction WN18RR HypER MRR 0.465 # 44
Hits@10 0.522 # 55
Hits@3 0.477 # 37
Hits@1 0.436 # 34
MR 5796 # 27

Methods