Learning to Evolve on Dynamic Graphs (Student Abstract)

Abstract

Representation learning in dynamic graphs is a challenging problem because the topology of graph and node features vary at different time. This requires the model to be able to effectively capture both graph topology information and temporal information. Most existing works are built on recurrent neural networks (RNNs), which are used to exact temporal information of dynamic graphs, and thus they inherit the same drawbacks of RNNs. In this paper, we propose Learning to Evolve on Dynamic Graphs (LEDG) - a novel algorithm that jointly learns graph information and time information. Specifically, our approach utilizes gradient-based meta-learning to learn updating strategies that have better generalization ability than RNN on snapshots. It is model-agnostic and thus can train any message passing based graph neural network (GNN) on dynamic graphs. To enhance the representation power, we disentangle the embeddings into time embeddings and graph intrinsic embeddings. We conduct experiments on various datasets and down-stream tasks, and the experimental results validate the effectiveness of our method.

Learnable Spectral Wavelets on Dynamic Graphs to Capture Global Interactions

“…Now, we present comprehensive experiments to evaluate our proposed framework. We borrow datasets and its preprocessing/splitting settings used in previous best baselines (Xiang, Huang, and Wang 2022;Pareja et al 2020). Datasets: Table ??…”

“…The best hyperparameters search has the range as: Number of layers ∈ {1, 2}, Hidden dimension ∈ {32, 64, 128}, Number of heads ∈ {4, 8, 16}, Filter order ∈ {4, 8, 16}, Wavelet scales ∈ [0.1, 10]. The rest of the parameters and settings are borrowed from previous works (Pareja et al 2020;Xiang, Huang, and Wang 2022). Variants of our Framework: In the main results we show three variants of our method with different aggregators.…”

“…Furthermore, autoencoder-based methods have been introduced, such as in (Goyal et al 2017;Goyal, Chhetri, and Canedo 2020;Xu, Singh, and Karniadakis 2022) that focus on reconstructing the graphs in future timesteps as an objective. Xiang, Huang, and Wang (2022) proposed a metalearning framework in which the objective is to predict the graph at future timestep. However, these works focus only on learning spatial properties to capture local dependencies and ignores the global dependencies that may emerge due to dynamic nature of the graph.…”