Multi model transfer learning with RULES family

Abstract

Due to the rapid growth of computer technologies and the extensive changes in human needs, expertise and digital information were used to induce general conclusions. Such conclusions can be used to deal with future activates and make the life of humans easier. One active filed of machine learning that was developed for this purpose is inductive learning, and several families have emerged from this field. Specifically, RULES family was discovered as covering algorithm that directly induces good and general conclusions in the shape of simple rules. However, it was found that RULES suffer from two major deficiencies. It needs to tradeoff between time and accuracy when inducing the best rule and it did not appropriately handle incomplete data. As a result, this paper will present a new RULES algorithm, which takes advantage of previous versions of RULES family in addition to other advance methods of machine learning, specifically Transfer learning. Moreover, multi-modeling is also merged to transfer the knowledge of a different classification model and further improve the original algorithm. At the end, an empirical test is applied to compare the proposed algorithm with different single-model algorithms to prove that using the past knowledge of other agents in different domains improves specialization accuracy, whether the data is complete or incomplete. © 2013 Springer-Verlag.