Evaluating Network Embedding Models for Machine Learning Tasks

摘要

Network embedding is a representation learning paradigm that seeks to learn a compact low-dimensional distributed vector representation for each vertex in the network; this learned low-dimensional vector representation can thus be used for different machine learning tasks. Over the years, so many network embedding models have been worked upon based on several approaches. In this paper, we study vector embeddings of 10 different representation learning models, with the sole aim of carrying out two machine learning tasks on these learned representations - unsupervised community clustering and link prediction analysis. The goal is to compare the output of these tasks using the 10 models, and draw inference based on the obtained results. We analyze the results using 4 link prediction baseline heuristic measures for the link prediction analysis; and a combination of silhouette score analysis and dissimilarity metric index for the community analysis.