TOWARDS OPEN TEMPORAL GRAPH NEURAL NET-WORKS

Abstract

Graph neural networks (GNNs) for temporal graphs have recently attracted increasing attentions, where a common assumption is that the class set for nodes is closed. However, in real-world scenarios, it often faces the open set problem with the dynamically increased class set as the time passes by. This will bring two big challenges to the existing temporal GNN methods: (i) How to dynamically propagate appropriate information in an open temporal graph, where new class nodes are often linked to old class nodes. This case will lead to a sharp contradiction. This is because typical GNNs are prone to make the embeddings of connected nodes become similar, while we expect the embeddings of these two interactive nodes to be distinguishable since they belong to different classes. (ii) How to avoid catastrophic knowledge forgetting over old classes when learning new classes occurred in temporal graphs. In this paper, we propose a general and principled learning approach for open temporal graphs, called OTGNet, with the goal of addressing the above two challenges. We assume the knowledge of a node can be disentangled into class-relevant and class-agnostic one, and thus explore a new message passing mechanism by extending the information bottleneck principle to only propagate class-agnostic knowledge between nodes of different classes, avoiding aggregating conflictive information. Moreover, we devise a strategy to select both important and diverse triad sub-graph structures for effective class-incremental learning. Extensive experiments on three real-world datasets of different domains demonstrate the superiority of our method, compared to the baselines.

1. INTRODUCTION

Temporal graph (Nguyen et al., 2018) represents a sequence of time-stamped events (e.g. addition or deletion for edges or nodes) (Rossi et al., 2020) , which is a popular kind of graph structure in variety of domains such as social networks (Kleinberg, 2007) , citations networks (Feng et al., 2022 ), topic communities (Hamilton et al., 2017) , etc. For instance, in topic communities, all posts can be modelled as a graph, where each node represents one post. New posts can be continually added into the community, thus the graph is dynamically evolving. In order to handle this kind of graph structure, many methods have been proposed in the past decade (Wang et al., 2020b; Xu et al., 2020; Rossi et al., 2020; Nguyen et al., 2018; Li et al., 2022) . The key to success for these methods is to learn an effective node embedding by capturing temporal patterns based on time-stamped events. A basic assumption among the above methods is that the class set of nodes is always closed, i.e., the class set is fixed as time passes by. However, in many real-world applications, the class set is open. We still take topic communities as an example, all the topics can be regarded as the class set of nodes for a post-to-post graph. When a new topic is created in the community, it means a new class is involved into the graph. This will bring two challenges to previous approaches: The first problem is the heterophily propagation issue. In an open temporal graph, a node belonging to a new class is often

