DeepFD: a deep learning approach to fast generate force-directed layout for large graphs

Deep learning techniques have been applied to the graph drawing of node-link diagrams to help figure out user preference of layout in recent research. However, when revisiting existing studies, only stress model and dimensional reduction methods are utilized in the unsupervised learning of graph drawing tasks since their gradient descent conditions can be easily constructed, and few works have explored their scalability on large graphs. In this paper, we propose a framework that can adapt most of the graph layout methods to a form of loss function and develop an implementation DeepFD, which takes the force-directed algorithm as the prototype to design the loss function. Our model is built with the graph-LSTM as encoder and multilayer perceptron as decoder and trained with dataset split from huge graphs with millions of nodes by Louvain. We design a set of qualitative and quantitative experiments to evaluate our method and compare with classical layout techniques such as F-R and K-K algorithms, while deep-learning based models with various architecture or loss function are adopted to perform ablation experiments. The results indicate that our developed approach can generate a high-quality layout of large graph within a low time cost, and the model we propose shows strong robustness and high efficiency.