Add ggcm method as example

[xYix]

We add the implementation of GGCM method from the paper From Cluster Assumption to Graph Convolution: Graph-based Semi-Supervised Learning Revisited (same as my last pull request). GGCM is a very practical unsupervised method, with two obvious advantages: 1) it is a “no-learning” method, and 2) it can preserve graph structure.

Table 1: Node classification accuracy.

Acc	CiteSeer	Cora	PubMed
SGC (PyG)	0.716	0.817	0.784
GGCM (PyG)	0.743	0.836	0.807
GGCM (Reported)	0.742	0.836	0.808

For comparison, we use PyG's build-in datasets and codes. As shown above, GGCM outperforms another classical "no-learning" method SGC by a large margin. This observation evidently demonstrates the value of preserving the graph structure information in GCN-type methods.

In addition, the introduced IGC operator can capture the dissimilarity between unlinked nodes. Theoretically, IGC is a numerically stable matrix form of negative sampling for graph scenario. Therefore, it can be used as a “plug-in” in various GCN-type models, to preserve graph structure or fix over-smoothing. If needed, we are very glad to re-modify IGC, to make it as a common "convolution" operator in the torch_geometric.nn part.

Author: Yun Xu (undergraduate, @sjtu) advised by Dr. Zheng Wang@SJTU.