Abstract
Link prediction aims at predicting future links or inferring missing
links based on observed partial networks or graphs. Active link
prediction has recently been shown to be highly effective at reducing
the amount of data needed to train a link predictor by leveraging active
learning. Existing research on active link prediction concentrated
solely on the information gain caused by Shannon entropy to actively
select samples. Few studies consider information gain brought by rule
uncer- tainty in the process of actively selecting samples. This will
cause the existing active link prediction research to be unable to use
the rich semantic information contained in the knowl- edge graph rules,
which is not conducive to further reducing the amount of data needed to
train a link predictor. In this paper, we address the aforementioned
challenge by propos- ing a Rule Guided Active Link Prediction (RGAL)
approach. Through the joint modeling of information uncertainty based on
Shannon entropy and information uncertainty based on rules, our proposed
method can select the most informative triples while taking into account
the latent semantics informa- tion contained in the triples. Extensive
experiments on real- world datasets verify the effectiveness of RGAL
compared with state-of-the-art methods.