# DIFF GRAPH U-NET: A DIFFERENTIABLE U-SHAPE NETWORK FOR GRAPH CLASSIFICATION
# DIFF GRAPH U-NET: A DIFFERENTIABLE U-SHAPE NETWORK FOR GRAPH CLASSIFICATION
## Introduction
## Introduction
Regarding GCNs, there are mainly three tasks: node classification, link prediction, and graph classification. This project focuses on the classification of graphics through Differentiable Pooling (DIFFPOOL). DIFFPOOL is a state-of-the-art pool module that can learn the hierarchical expression of graphics and combine various end-to-end GCN structures. Hence, in order to research the changes of a graph after several DIFFPOOL modules respectively, the visualizations for pooling has been fulfilled via rebuilding the graph. Apart from DIFFPOOL, this project also explored the possibility to build an encoder-decoderarchitecture via building such a U-net-like method for graph data. A novel GCN architecture DIFF Graph U-Net is proposed.
Recently, graph convolutional networks (GCNs) have been demonstrated efficient in learning graph representations. Regarding the down-sampling and up-sampling of non-Euclidean data, most existing methods are flat and lack robustness. We visualize the process of a state-of-the-art work DiffPool, and develop a novel differentiable module for upsampling called DiffUnpool. DiffPool and DiffUnpool learn soft cluster assignment for nodes via GCNs and multi-layer perceptrons respectively. To address the graph classification problem, based on DiffPool and DiffUnpool, we further propose an end-to-end encoder-decoder architecture, diff graph U-Net. Different from other U-shape models before, diff graph U-Net learns node embeddings hierarchically, and collect global features in residual fashion. Our experimental results show that our model yields an overall improvement of accuracy on 4 different data sets, compared with previous methods.
