Forming an Electoral College for a Graph: a Heuristic Semi-supervised Learning Framework Chen Li Beihang University lichen@act.buaa.edu.cn Xutan Peng The University of Sheffield x.peng@shef.ac.uk Hao Peng Beihang University penghao@buaa.edu.cn Jianxin Li Beihang University lijx@act.buaa.edu.cn Lihong Wang CNCERT wlh@isc.org.cn Philip S. Yu UIC psyu@uic.edu Abstract Recently, graph-based algorithms have drawn much attention because of their impressive success in semi-supervised scenarios. For better model performance, previous studies learn to transform the topology of the input graph. However, these works only focus on optimizing the original nodes and edges, leaving the direction of augmenting existing data unexplored. In this paper, by simulating the generation process of graph signals, we propose a novel heuristic pre-processing technique, namely ELectoral COllege (ELCO), which automatically expands new nodes and edges to refine the label similarity within a dense subgraph. Substantially enlarging the original training set with high-quality generated labeled data, our framework can effectively benefit downstream models. To justify the generality and practicality of ELCO, we couple it with the popular Graph Convolution Network and Graph Attention Network to extensively perform semi-supervised learning evaluations on three standard datasets. In all setups tested, our method boosts the average score of base models by a large margin of 4 points, as well as consistently outperforms the state-of-the-art. Please find our code at https://github.com/ RingBDStack/ELCO. 1 Introduction Numerous real-world data can be represented as graphs, e.g., social networks [18], citation net- works [29], knowledge graphs [22], and protein-interaction networks [6]. In many cases, large-scale annotated data is expensive to obtain. The so-called graph-based Semi-Supervised Learning (SSL), which holds promise to bootstrap applications even with limited supervision, has therefore attracted increasing research interest. Earlier works develop the classical regularization methods, which achieve SSL by smoothing fea- ture representations or model predictions over local neighborhoods using explicit regularization schemes [2, 13, 15, 32]. Although this direction has been well studied, a later thread of algorithms, namely graph convolution network methods, has demonstrated state-of-the-art performance and drawn much attention [3,11,26]. By utilizing various aggregation strategies, these models selectively fuse the local features of the graph into the hidden representations of its target nodes. To further perform downstream tasks, the hidden layers are coupled with specific task layers [11, 26]. One com- mon characteristic of these two strands of models is that, they both adopt the presence of smoothness within the graph structure as a basic assumption. Preprint. Under review. arXiv:2006.06469v1 [cs.SI] 10 Jun 2020