Research On Markov Logic Networks And Its Application In Social Networks

Today is the network information age, with more and more social network data collected by people, it becomes very important to find useful informal from these relational data. Markov logic network (MLN) is a model that deals with relational data.An MLN is a first-order knowledge base in essence which consists of formulas attached weight and is a template to construct Markov network. An MLN provides a compact descriptive language and adds uncertain processing. It also serves as a unifying framework for the task of SRL.This paper studies MLNs applying in social network analysis. Research on MLNs includes structure learning and parameters learning. In structure learning we use top-down learning and bottom-up learning algorithm. In parameters learning we use both the generative learning and the discriminative learning. During parameters learning we need to combine optimization algorithm (e.g., Steepest Descent method and Diagonal Newton method) with probabilistic reasoning algorithm (e.g., MaxWalkSat and MC-SAT) to optimize objective function. This paper compares different combinations of these algorithms in order to find the most efficient combination. In addition, this paper proposes a new optimization algorithm called Scaled Conjugate Gradient. It uses Hessian matrix to compute step size in Conjugate Gradient method, instead of line search.At last, we use MLN to model on UW-SCE data set with a variety of learning algorithms, and then compare the experimental results of different combinations to obtain the best one.