Finding Latent Relationships from Auxiliary Information for Tensor Completion

摘要

Discovering multi-object relationships is fundamental to knowledge discovery. Tensor is a highly suitable representation for multi-object relationships. However, most tensors are very sparse because the multi-object relationships that can be observed directly are usually limited, especially when the order of tensors increases. Tensor completion is therefore a meaningful task in many applications. Most of the existing approaches for tensor completion assume that tensors are of low rank and then complete them relying on the limited number of observations only. We propose an approach called GCD-AA (Graph Community Detection-Apriori Algorithm) that uses the relationships contained in auxiliary information to improve the performance of tensor completion. GCD-AA first utilizes graph community detection algorithm to detect sets of highly related items in low-order auxiliary information, then it makes use of Apriori algorithm to capture the latent relationships in high-order tensor. The extracted high-order relationships are strong and reliable, so they can be used for tensor completion. To validate the performance of GCD-AA against other tensor completion approaches, we conduct experiments on a publicly available dataset FB15k-237. The experimental results demonstrate that GCD-AA achieves better performance on tensor completion than those comparing approaches.