Are Graph Embeddings the Panacea? An Empirical Survey from the Data Fitness Perspective

Abstract

Graph representation learning has emerged as a machine learning go-to technique, outperforming traditional tabular view of data across many domains. Current surveys on graph representation learning predominantly have an algorithmic focus with the primary goal of explaining foundational principles and comparing performances, yet the natural and practical question “Are graph embeddings the panacea?” has been so far neglected. In this paper, we propose to examine graph embedding algorithms from a data fitness perspective by offering a methodical analysis that aligns network characteristics of data with appropriate embedding algorithms. The overarching objective is to provide researchers and practitioners with comprehensive and methodical investigations, enabling them to confidently answer pivotal questions confronting node classification problems: 1) Is there a potential benefit of applying graph representation learning? 2) Is structural information alone sufficient? 3) Which embedding technique would best suit my dataset? Through 1400 experiments across 35 datasets, we have evaluated four network embedding algorithms – three popular GNN-based algorithms (GraphSage, GCN, GAE) and node2vec – over traditional classification methods, namely SVM, KNN, and Random Forest (RF). Our results indicate that the cohesiveness of the network, the representation of relation information, and the number of classes in a classification problem play significant roles in algorithm selection.