An Evaluation of Large-scale Information Network Embedding based on Latent Space Model Generating Links

Graph representation learning encodes vertices as low-dimensional vectors that summarize their graph position and the structure of their local graph neighborhood. These methods give us beneficial representation in continuous space from big relational data. However, the algorithms are usually evaluated indirectly from the accuracy of applying the learning results to classification tasks because of not giving the correct answer when graph representation learning is applied. Therefore, this study proposes a method to evaluate graph representation learning algorithms by preparing correct learning results for the data by distributing objects in the latent space in advance and probabilistically generating relational graph data from the distributions in the latent space. Using this method, we evaluated LINE: Large-scale information network embedding, one of the most popular algorithms for learning graph representations. LINE consists of two algorithms optimizing two objective functions defined by first-order proximity and second-order proximity. We prepared two link-generating models suitable for these two objective functions and clarified that the corresponding LINE algorithm performed well for the link data generated by each model.