Exploring the representational power of graph autoencoder

While representation learning has yielded a great success on many graph tasks, there is little understanding behind the structures that are being captured by these embeddings. For example, we wonder if topological features, such as Triangle Count, Degree of node and other centrality measures concretely encoded in Furthermore, ask presence embeddings necessary for better performance downstream clustering classification. To address questions, conduct an extensive empirical study over three classes unsupervised embedding models seven different variants Graph Autoencoders. Our results show five features: Degree, Local Clustering Score, Betweenness Centrality, Eigenvector Count preserved first layer autoencoder employs SUM aggregation rule, under condition model preserves second-order proximity. We supplement further evidence features revealing hierarchy distribution aforementioned model. also with properties can outperform certain especially when relevant to task at hand. Finally, evaluate suitability our findings through test case related social influence prediction.