Embedding Methods or Link-based Similarity Measures, Which is Better for Link Prediction?

The link prediction task has attracted significant attention in the literature. Link-based similarity measures (in short, similarity measures) are the conventional methods for this task, while recently graph embedding methods (in short, embedding methods) are widely employed as well. In this paper, we extensively investigate the effectiveness of embedding methods and similarity measures (i.e., both non-recursive and recursive ones) in link prediction. Our experimental results with three real-world datasets demonstrate that 1) recursive similarity measures are not beneficial in this task than non-recursive one,2) increasing the number of dimensions in vectors may not help improve the accuracy of embedding methods, and 3) in comparison with embedding methods, Adamic/Adar, a non-recursive similarity measure, can be a useful method for link prediction since it shows promising results while being parameter-free.