Link Completion using Prediction by Partial Matching

Abstract

Prediction by partial matching (PPM) is typically used as a powerful method for data compression. Recently, PPM was applied to solve link prediction problem, e. g., predictive prefetching on the Web. Link completion is a link analysis problem and is almost identical to link prediction but harder and more general. This research applies PPM to impute the missing links in single (directed) graph-structured data model with node and link labels. The experiments use the co-authorship dataset for case-study. Our proposed algorithm not only uses original PPM forward method but also PPM backward and hybrid methods. The algorithm can predict any missing position at any position of a given query link. The experimental results show the prediction accuracy in several dimensions depending on the testing data.