A Classification with Random SPI: Better Models in Uncertain Environment

This paper addresses the problem of learning optimal classifiers that maximally improve the robustness and accuracy in uncertain environment included a large number of noise and missing values. Recent solutions to the efficiently vertex weight evaluation, such as the Bayes Network, rely on statistics methods, without sufficient robust guarantees. We show how a globally optimal solution can be obtained by formulating predicates and statistical training set evaluation in Markov Logic Network. We then propose a classification algorithm which adopts random selection of the instances and features in Random Statistical Predicate Invention (RSPI) classification model. In a set of experiments on UCI datasets about credit card and CRM information we show that the proposed RSPI can achieve significant gains in robustness of model, compared to decision trees algorithms or other random classification methods.