Evaluation of LID-Aware Graph Embedding Methods for Node Clustering

Data generated by everyday applications may appear in different forms. Various important and frequently used machine learning and data mining techniques have been designed assuming the tabular data form. To apply those techniques to graph structured data, it necessary to form graph embeddings. The crucial moment in creating a graph embedding is to choose the best embedding technique that preserves all the vital information when converting a graph into its tabular representation. Determining the best approach requires some form of evaluation of the internal qualities of potential embeddings and their utility in concrete applications. In this paper, we present a comparative evaluation of graph embeddings when used to cluster graph nodes in the embedded space. The examined graph embedding methods are node2vec and two recently proposed extensions of this algorithm based on local intrinsic dimensionality. The results of both intrinsic and external clustering evaluation on real-world graphs indicate that LID-aware extensions improve node clustering, especially when detecting a small number of clusters.