Embedding Networks: Node and Graph Level Representations

Abstract

Graph neural networks gained significant attention for graph representation and classification in the machine learning community. For graph classification, different pooling techniques are introduced, but none of them has considered both the node's neighborhood and the long-range effects on nodes. For this, we propose a novel graph pooling layer R2POOL, which balances both the region-based and the graph-level importance of the nodes and retains the most informative nodes for the next coarser version of the graph. Further, we integrate R2POOL with our multi-level prediction and branch training strategies to learn graph representations and to enhance the model's capability for graph classification. We call the combined model R2MAN. Experiments show that our method has the potential to improve the performance of graph classification on various datasets.

Furthermore, attention mechanism applied on the neighborhood of a node improves the performance of graph neural networks. Typically, it helps to identify a neighbor node which plays a more important role in determining the label of the node under consideration. But in real-world scenarios, a particular subset of nodes together, but not the individual pairs in the subset, may be important to determine the label of the node and the graph. We address this problem and introduce the concept of subgraph attention for graphs. To show the efficiency of this scheme, we use subgraph attention for node classification downstream task. On the other hand, hierarchical graph pooling is promising in the recent literature. But, not all the graphs at different levels play an equal role in graph classification. Towards this end, we propose a novel graph classification algorithm called SubGattPool, which jointly learns the subgraph attention and employs two different attention mechanisms to find the important nodes in a hierarchy and the significance of graphs at different levels. Improved performance over the state-of-the-art on multiple publicly available datasets for both node and graph classification shows the efficiency of the proposed algorithms.