Parfaite: PageRank-Matrix Factorization for Interpretable Graph Embeddings

There have been increasing efforts in recent years to develop so-called graph embeddings, which among other things allow to employ standard machine learning techniques to solve urgent real-world problems. However, developing interpretable graph embeddings has received much less attention. In our work, we develop Parfaite, an algorithm for finding an interpretable and effective graph embedding, based on the factorization of the PageRank matrix of the input graph. We evaluate the interpretability of our method against popular graph embedding techniques, such as node2vec, showing that Parfaite boasts significantly higher interpretability scores. Another contribution of our work is the release of a novel dataset constructed from all pages of the French version of Wikipedia, which we release for reproducibility and benchmarking.