Enhanced Feature Learning with Its Applications to Link Prediction and Defense in Heterogeneous Information Networks

Abstract

Since the heterogeneous information networks contain rich information it is worth discussing how to extract useful information from the networks We hope that we can design a method to preserve both structural of heterogeneous network and correlation between nodes Our task is to predict the users’ social relationships (UU-LP) and the links between users and items (UI-LP) If we can predict the relationships between users or user and item precisely it means that the method can capture the correlation between the nodes We propose two methods metamotif2vec and diversewalk2vec to learn a low-dimentional feature representation for each node in heterogeneous information networks The metamotif2vec model formalizes a structural random walk which can consider the relationships between much more different types of nodes at the same time On the other hand the diversewalk2vec model designs a diversified random walk to capture the correlation automatically without defining the form of random walk in advance Experiments conducted on large-scale Twitter check-ins dataset and Douban book dataset exhibit that metamotif2vec and diversewalk2vec can average achieve 7 1% and 5 2% improvement over the state-of-the-art heterogeneous network representation learning method metapath2vec in both tasks of UU-LP and UI-LP respectively While the Internet makes human life more convenient it also raises privacy risks Therefore we propose some defense mechanisms for disturbing data and also conduct experiments for link prediction to evaluate their effectiveness The results show that the defense mechanisms can reduce the possibility of leakage of users' personal privacy

Date of Award	2019
Original language	English
Supervisor	Cheng-Te Li (Supervisor)