MHRWalk

Abstract

Network representation learning is a task aimed at embedding network typology into low-dimensional vector space. One advantage is that tedious feature engineering work is avoided as the network representation learning model runs in an unsupervised manner, the learned latent vectors are general enough to be applied in a wide variety of network tasks. Another benefit is that the network representation learning model better captures the homophily characteristic and encodes it into latent vectors, which provide the performance improvement in the multi-label vertex classification task.

Various network representation learning models have been proposed[1, 2, 3]. Although these models are empirically effective, we find that they suffer from the "stop words" issue and the "short-sighted" problem. The "stop words" issue indicates that meaningless vertices are involved in the computation procedure, causing the model to be inefficient. The "short-sighted" problem is that the models fail in capturing the information of long-distance proximities between vertices, resulting in inadequate vertex representation are learned.

In this thesis, we present a novel network representation learning model named MHRWalk to learn latent vectors for a network. It mitigates the "stop words" issue and resolves the "short-sighted" problem by introducing the Metropolis-Hastings random walk sampler to this task. Moreover, we also show how latent representations of individual vertices learned using MHRWalk can be extent to sub-networks.

Experiments on real-world network datasets demonstrate the effectiveness and efficiency of the proposed model compared the state-of-the-art vertex representation models. We also demonstrate that the learned latent sub-network representation outperforms all baseline approaches in the sequence matching task. Beyond this, we visualize several sub-networks in a two-dimensional plane with the help of MHRWalk, which projects the sub-network topology into a low-dimensional and continuous-valued vector.