Research On Relational Markov Networks And Its Application In Social Networks

There are many relational data in real-world social networks. These datasets consist of different types of entities with extensive and complex links. The mining of these links can provide us rich and accurate information of the social networks. To study how to use of these link attributes sufficiently is of great significance in social network analysis.Relational Markov network is one of discriminative probabilistic graph models. Through the combination of probabilistic inference model and relational schema, it can deal with complicated relational data effectively. Applying the model to social network datasets classification can fully capture the dependencies among them, thus effectively improve the accuracy of classification.This paper researches the learning process of Relational Markov network. We deeply analyse the likelihood approach to construct the objective function. However, the experiments found that the time complexity becomes higher and higher along with the increasing of the size of datasets. To solve this problem, this paper introduces pseudo-likelihood approach instead of likelihood approach to construct the objective function. While in the parameter optimization, we research some effective nonlinear optimization algorithms, such as conjugate gradient algorithm, gradient descent algorithm and quasi-Newton algorithm. And we also research some one-dimensional search algorithms, such as golden section algorithm, Newton algorithm and Armijo-Goldstein algorithm. In addition, this paper compares both the advantages and disadvantages of each algorithm according to the classification accuracies and the time complexities, and then seeks to provide an optimal combination of these algorithms.We use Relational Markov network framework on a classification task with Cora dataset and WebKB dataset. And the experimental results show that the time complexity of pseudo-likelihood is much lower than likelihood approach. While in the parameter optimization, the combination of quasi-Newton algorithm and the golden section algorithm can achieve higher classification accuracy and lower time complexity.