Project

Hierarchical Information in Knowledge Graphs

Date

    2015/11 - 2017/03

Sponsored by

    National Grand Fundamental Research 973 Program of China No. 2014CB340 401

Advisor

    Yantao Jia, Yuanzhuo Wang

Problem

    50% triples in knowledge graphs are organized in hierarchical structures, which also contain rich inference patterns to predict links but do not be fully used. A main feature of hierarchical structure is multi-layer, which traditional methods fails to make use of.

My Work 1

  • Proposed a hierarchy-constrained link prediction method by defining single-step specific margin, called hTransA, on the basis of the translation-based knowledge graph embedding methods. It can adaptively determine the optimal single-step specific margin to separate positive and negative triples.
  • Provided one way to detect the hierarchical structures in knowledge graphs by employing the properties of hierarchical relations.
  • Experimented hTransA on two datasets, and it decreased the mean rank of the state-of-the-art method at that time, TransA, by more than 20%.
  • The paper "Hierarchy-Based Link Prediction in Knowledge Graphs" was published as a poster by WWW 2016 (Lead Author).

My Work 2

  • Came up with the idea to divide the hierarchical structures into two categories, i.e., single-step hierarchical structures and multi-step hierarchical structures.
  • Proposed multi-step specific margin, and the link prediction method based on detecting single-step and multi-step hierarchical structures, called hTransM. It can adaptively determine the optimal margin to separate positive and negative triples.
  • Proved the convergence of hTransM by demonstrating its uniform stability and provide the upper bound of the error of the proposed model.
  • Experimented hTransM over 3 datasets of 2 sub-tasks, i.e., entity prediction task and relation prediction task, which demonstated that hTransM can achieve better prediction performance.
  • Studyed the variation of the optimal margin value along with the optimization process, and compared with the methods withnot considering hierarchical information. It validated the superiority of hierarchy-constrained margin.
  • The paper "Link Prediction in Knowledge Graphs: A Hierarchy-Constrained Approach" was submitted to IEEE Transaction on Big Data (Lead Author) and under review now.