Fully Heterogeneous Collective Regression

摘要

Prior work has demonstrated that multiple methods for link-based classification (LBC) can substantially improve accuracy when the nodes of interest are interconnected. To date, however, very little work has considered how methods for LBC could be applied in domains that require continuous, rather than categorical, predictions. In addition, prior work with LBC has learned only one predictive model to use for all nodes of a given type, but some domains exhibit significant node diversity that is not well-suited to this approach. In response, we introduce fully heterogeneous collective regression (FHCR), a new method that learns node-specific models from data and uses these models to jointly predict continuous outputs. We apply FHCR to a voting prediction task, and create novel correlation-based links that outperform alternative methods. In addition, we introduce multiple new methods for inferring continuous outputs that can incorporate link-based information, and show that regression-specific methods based on Bayesian inference outperform the naive approach of inserting regression into existing LBC methods. Overall, we demonstrate the viability of the new FHCR paradigm by producing results that are comparable or better than those of previous link-unaware methods, yet are at least two orders of magnitude faster.