FARS: A Multi-relational Feature and Relation Selection Approach for Efficient Classification

Abstract

Feature selection is an essential data processing step to remove the irrelevant and redundant attributes for shorter learning time, better accuracy and better comprehensibility. A number of algorithms have been proposed in both data mining and machine learning area. These algorithms are usually used in single table environment, where data are stored in one relational table or one flat file. They are not suitable for multi-relational environment, where data are stored in multiple tables joined each other by semantic relationships. To solve this problem, in this paper we propose a novel approach called FARSto do both feature and relation selection for efficient multi-relational classification. By this approach, we not only extend traditional feature selection method to selects relevant features from multi-relations, but also develop a new method to reconstruct the multi-relational database schema and get rid of irrelevant tables to further improve classification performance. Results of experiments conducted on several real databases show that FARScan effectively choose a small set of relevant features, enhancing the classification efficiency significantly and improving prediction accuracy.