Krishna Kumar Tiwari, V. Vijaya Saradhi, Indian Institute of Technology Guwahati, India

We use composite likelihood for structure learning and parameter estimation in relational dependency networks (RDNs). RDNs currently use pseudolikelihood, to learn parameters, which is a special case of composite likelihood function. Composite likelihood learning is used to give trade-off between computational complexity and performance of the model. Variance of the model is minimum in case of full likelihood and maximum in pseudolikelihood. In particular we focus on modified second order pseudolikelihood function and extend relational Bayesian classifier (RBC) to this setting. Second order RDNs explore pairwise attribute correlation. We evaluate second order learning on synthetic and real world data sets. We observe experimentally second order model has an edge over the pseudolikelihood based model particularly when correlation is high.