An Empirical Study of Locally Updated Large-scale Information Network Embedding (LINE)
The problem of embedding very large information networks into low-dimensional vector spaces is useful in many tasks such as visualization, node classification, and link prediction. This paper studies the novel network embedding method called the ''LINE'' , which optimizes a carefully designed objective function that preserves both the local and global network structures. In order to rule out the instability on border vertices' embeddings and its influence on core vertices, we only compute the core graph with LINE after peeling out those border vertices. Then we compute the embedding of peeled nodes with locally updated process. This paper also tried to interpret and visualize the local update process with logistic regression, and optimize the local update process by adding prior and intercept to the objective function. Finally, this paper demonstrate the embeddings on several multi-label network classification tasks for social networks such as BlogCatalog and YouTube. Our results show that the optimized LINE outperforms the initial methods 5% in F1-score with YouTube dataset and speed up the convergence time.