Understanding the Benefits of Forgetting When Learning on Dynamic Graphs.

In order to solve graph-related tasks such as node classification, recommendation or community detection, most machine learning algorithms are based on node representations, also called embeddings, that allow to capture in the best way possible the properties of these graphs. More recently, learning node embeddings for dynamic graphs attracted significant interest due to the rich temporal information that they provide about the appearance of edges and nodes in the graph over time. In this paper, we aim to understand the effect of taking into account the static and dynamic nature of graph when learning node representations and the extent to which the latter influences the success of such learning process. Our motivation to do this stems from empirical results presented in several recent papers showing that static methods are sometimes on par or better than methods designed specifically for learning on dynamic graphs. To assess the importance of temporal information, we first propose a similarity measure between nodes based on the time distance of their edges with an explicit control over the decay of forgetting over time. We then devise a novel approach that combines the proposed time distance with static properties of the graph when learning temporal node embeddings. Our results on 3 different tasks (link prediction, node and edge classification) and 6 real-world datasets show that finding the right trade-off between static and dynamic information is crucial for learning good node representations and allows to significantly improve the results compared to state-of-the-art methods.