Meta-Prediction for Collective Classification

When data instances are inter-related, as are nodes in a social network or hyperlink graph, algorithms for collective classification (CC) can significantly improve accuracy. Recently, an algorithm for CC named Cautious ICA (ICAC) was shown to improve accuracy compared to the popular ICA algorithm. ICAC improves performance by initially favoring its more confident predictions during collective inference. In this paper, we introduce ICAMC, a new algorithm that outperforms ICAC when the attributes that describe each node are not highly predictive. ICAMC learns a meta-classifier that identifies which node label predictions are most likely to be correct. We show that this approach significantly increases accuracy on a range of real and synthetic data sets. We also describe new features for the meta-classifier and demonstrate that a simple search can identify an effective feature set that increases accuracy