- Aishwarya Pothula

# Graphgan: Graph Representation Learning With Generative Adversarial Nets

Graph feature learning or network feature learning’s main purpose is to map each node in the graph to a low-dimensional vector space while maintaining the structure information or distance information of the original graph. It can also be called Network Embedding, Graph Embedding or GRL. GRL has a wide range of applications including link prediction, node classification, recommendation systems, vision, knowledge graph representation, clustering, text embedding, and social network analysis.

GRL when classified based on input exists in four types namely isomorphic, heterogenous, graph with side information and graph transformed from non-relational data. Similarly, when classified based on output, it outputs four types namely node embedding, output edge embedding, subgraph embedding and embedding of the full image.

GraphGAN's algorithm is as above. The input is some hyperparameters. It outputs the generative model G and discriminative model D. Lines 3-12 are each iteration of the algorithm. In each of the iterations, the generator is used to generate s points, and these points are used to update the parameters of θG.

We use GraphGAN in the following three test scenarios. The first is link prediction, which predicts the probability of whether there is an edge between two points. Figure 4 shows the learning curve of GraphGAN. For the generator, it stabilizes quickly after about ten rounds of training and maintains performance afterwards. For D, its performance increases first and then decreases slowly. This is also consistent with the GAN framework we described earlier. Table 1 shows that GraphGAN obtained the best results on both datasets. The second test scenario is Node Classification. In this scenario, we want to classify nodes. The datasets we use are BlogCatalog and Wikipedia. In such data, our method achieved the best results. The third test scenario is recommended. The data set used was MovieLens, and our method achieved the best results.

GraphGAN proposed in this paper is a framework that combines a generative model and discriminative model. G and D are actually playing a minimax game, where G tries to generate some fake points. These fake points cannot be discriminated by D, and D tries to separate the true value from the fake points to avoid being deceived by G.In addition, we also propose a Graph Softmax implementation as G, which overcomes the shortcomings of softmax and hierarchical softmax. We conduct experiments on five real-world datasets in three scenarios, and the results demonstrate that GraphGAN significantly outperforms strong baselines in all experiments due to its adversarial framework and proximity-aware graph softmax.