Graph Auto-Encoders for Learning Edge Representations

Virgile Rennard, Giannis Nikolentzos, Michalis Vazirgiannis

Abstract: Graphs evolved as very effective representations of different types of data including social networks, biological data or textual documents. In the past years, significant efforts have been devoted to methods that learn vector representations of nodes or of entire graphs. But edges, representing interactions between nodes, have attracted less attention. Surprisingly, there are only a few studies that focus on generating edge representations or deal with edge-related tasks such as the problem of edge classification. In this paper, we propose a new model (in the form of an auto-encoder) to learn edge embeddings in (un)directed graphs. The encoder corresponds to a graph neural network followed by an aggregation function, while a multi-layer perceptron serves as our decoder. We empirically evaluate our approach in two different tasks, namely edge classification and link prediction. In the first task, the proposed model outperforms the baselines, while in the second task, it achieves results that are comparable to the state-of-the-art.