Core2Vec: A Core-Preserving Feature Learning Framework for Networks
Soumya Sarkar, Aditya Mukund Bhagwat, Animesh Mukherjee · 2018
Recent advances in the field of network representation learning are mostly attributed to the application of the skip-gram model in the context of graphs. State-of-the-art analogues of skip-gram model in graphs define a notion of neighbourhood and aim to find the vector representation for a node, which maximizes the likelihood of preserving this neighborhood. In this paper, we take a drastic departure from the existing notion of neighbourhood of a node by utilizing the idea of coreness. More specifically, we utilize the well-established idea that nodes with similar core numbers play equivalent roles in the network and hence induce a novel and an organic notion of neighbourhood. Based on this idea, we propose core2vec, a new algorithmic framework for learning low dimensional continuous feature mapping for a node. Consequently, the nodes having similar core numbers are relatively closer in the vector space that we learn. We further demonstrate the effectiveness of core2vec by comparing word similarity scores obtained by our method where the node representations are drawn from standard word association graphs11In linguistics, such networks built from various linguistic units are known to have a core-periphery structure (see [3] and the references therein), against scores computed by other state-of-the-art network representation techniques like node2vec, Deep-Walk and LINE. Our results always outperform these existing methods, in some cases achieving improvements as high as 46 % on certain ground-truth word similarity datasets. We make all codes used in this paper available in the public domain: https://github.com/Sam131112/Core2vec_test.