Non-Boolean Features Set Covering Machine Mixed With Rough Set Theory

In recent years, the study of classifiers is a hot spot in machine learning. Various studies have shown that classifiers generated by current methods are accurate but difficult to interpret.Mario Marchand and John Shawe-Taylor proposed the theory of set covering machine in 2002. It was proposed as an alternative to the support vector machine when the objective is to obtain a sparse classifier with good generalization. The set covering machine allows a trade-off between accuracy and complexity, and it can return compact conjunctions and disjunctions of Boolean attributes for any feature sets that are constructed from the data. It performs as efficiently as the other learning machines, and outperforms other learning methods in some aspects. But most of the feature sets in information systems are non-Boolean, so the algorithm cannot be used directly for general practical learning tasks.To promote the application of the set covering machine, it introduces a method to chang non-Boolean attributes to Boolean ones and also does some changes to the original set covering machine algorithm in the paper. It extends the set covering machine for Boolean attributes to the cases whose conditional attributes are non-Boolean ones. The paper also extends the SCM algorithm beyond binary classification to the multi-class setting. By combining with knowledge of the rough set theory, it gets a more intuitional decision rules. By analyzing and dealing with the practical data, it returns good results in expectation and proves that the improving algorithm is a useful tool for general learning tasks.