Paper SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS.
Big Question: semi-supervised classification of graph data
- reason
- computation effective: semi-supervision
- the complex of graphs, the information of nodes and edges are not structural information.
- background
- the improvement of GCNs: spectral GCNs
Key points
The approximation of spectral graph convolution
The lines in paper had confused me at first before I ran the codes.
The difference of graph convolution and valina convolution is the input, as the input is a graph rather than data in same dimension, the key point is how to convert data represented by node and graph to a tensor in fixed dimension.
To solve it, Thomas maps the graph into a spectral space and also, to be computational efficient, approximate the infinite coefficients by second-order Chebyshev polynomial formulas.
After those approximation, it is input into the whole network with features.
Build model
Actually, except the complicated preprocess to represent graph G into a sparse tensor, the other step are not that complex, just the similar as what a convolution layer do. \[ Z = f\left(\mathbf{X},A\right)=softmax\left(\hat{A}ReLU\left({\hat{A}XW^{\left(0\right)}}\right)W^{\left(1\right)}\right) \]