Augmenting statistical relational learning techniques for classification and regression

Abstract

Statistical Relational Learning (SRL) is a subfield of machine learning and it aims at learning from data that exhibit both relational structure and uncertainty. While statistical learning can handle uncertainty and logic learning can handle relations, SRL combines ideas from these two fields, aspiring to learn from relational data having uncertainty. With SRL, we can model worlds with structured objects having complex relationships amongst them, and take advantage of background knowledge. Several SRL approaches evolve probabilistic graphical models to deal with relational data: Markov Logic Network (MLN) and Relational Dependency Network (RDN) extend the graphical models Markov network and dependency network, respectively, to learn from data using conditional dependencies, whereas Probabilistic Logic Programming (PLP) uses Bayesian network for inference. newlineProbabilistic logic learning can handle both relational structure and uncertainty in the data. Incorporating domain knowledge in probabilistic logic approach further augments learning, leading to improved accuracy. Prediction of sports outcome is a fascinating application of predictive analytics. Soccer is a team game where win or loss of a match can be predicted from the history of the matches the teams have played. The historical data about soccer matches are better represented in relational form. A probabilistic logic model is proposed for prediction outcomes of team games, using the history of the matches and the overall composition of the teams. Importantly, the domain knowledge of game experts can be framed as logic rules to enhance prediction. Compared to the traditional machine learning approaches to team game outcome prediction, the probabilistic logic system results in significant improvement in the prediction accuracy. newline newline