NetGloVe: Learning Node Representations for Community Detection

Sudarshan Srinivasa Ramanujam

Abstract: In this paper, we propose NetGloVe, a node representation learning method inspired by GloVe (Global Vectors for Word Representation) [3], a word embeddings technique in NLP. GloVe uses a log bilinear model to derive vector representations of words, taking into consideration both the word co-occurrence statistics as well as the words context. GloVe is comparable, if not superior, to the Skip-gram model, which considers only the words local context to derive the word representation. Our goal here is to extend GloVe to the context of graphs, by finding a suitable analogy for the word-word co-occurrence matrix used by GloVe. Our intuition is that, nodes that belong to the same communities would have similar embeddings, and thus would be clustered together. That way, we have employed the inverse of the shortest path distance between individual pairs of nodes in the graph, to populate GloVe’s co-occurrence matrix – as we wanted to weight higher nodes that are at close proximity.