Cross View Link Prediction by Exploiting Rich Graph Structure Information via Kernel Embedding

Link Prediction plays an important role in various machine learning applications during the last two decades. Despite the success of existing link prediction approaches, their performance is limited to the common assumption that the whole information of graph structure is available. In the machine learning community, partially observable attributed networks have incomplete information about attributes or links. Moreover, the use of appropriate similarity is a crucial issue for preserving the complex graph structure. We propose a novel cross-view link prediction technique to alleviate these challenging problems by utilizing nonlinear node relations via kernel embedding. This study aims to reveal the intrinsic graph structure by implicitly using a similarity estimation of the inner product between the transformed graph representations in the high dimensional feature space. In practice, this similarity is explicitly computed in the original space through Gaussian Kernel. Experimental results on seven publicly available datasets demonstrate the superiority of the proposed method compared to the earlier link prediction method in terms of precision and recall evaluations

Keywords:

Graph Embedding , Link Prediction , Gaussian Kernel