Reformulating Downstream Tasks
Method I ·
To further narrow the gap between various graph tasks and state-of-the-art pre-training strategies, we further study the task space of various graph applications and reformulate downstream problems to the graph-level task.
Specifically, we reformulate node-level and edge-level tasks to graph-level tasks by building induced graphs for nodes and edges, respectively. As shown in Figure a, an induced graph for a target node means its local area in the network within π distance, which is also known as its π-ego network. This subgraph preserves the nodeβs local structure by neighboring node connections and its semantic context by neighboring node features, which is the main scope of most graph neural encoders. When we treat the target nodeβs label as this induced graph label, we can easily translate the node classification problem into graph classification; Similarly, we present an induced graph for a pair of nodes in Figure b. Here, the pair of nodes can be treated as a positive edge if there is an edge connecting them, or a negative edge if not. This subgraph can be easily built by extending this node pair to their π distance neighbors. We can reformulate the edge-level task by assigning the graph label with the edge label of the target node pair. Note that for unweighted graphs, the π distance is equal to π-hop length; for weighted graphs, the π distance refers to the shortest path distance, where the induced graph can be easily found by many efficient algorithms.