Handling over-fitting and class-imbalance jointly in Pittsburgh learning classifier systems

Abstract

Generalization ability of a classi er is an important issue for any classi cation task. Two prominent problems a ecting the generalization ability are over- tting and class-imbalance. It is thus important to address these problems while developing a classi cation system. There has been an enormous amount of work on classi cation problems in the machine learning literature, but handling over- tting and class imbalance is still an open issue. Most of the existing works su er more or less from these problems. This thesis presents a new evolutionary system, i.e., EDARIC, for rule induction and classi cation. The evolutionary approach used in our new system is based on a destructive method that starts with large-sized rules and gradually decreases the sizes as evolution progresses. The novelty of this thesis lies mainly in the way it addresses and handles the over- tting problem by incorporating an intelligent deletion mechanism for producing smaller-sized, i.e., generalized rules. Another beauty of EDARIC is its simplicity, which is due to using a minimum number of operators and parameters during evolution. Furthermore, EDARIC evolves multiple populations with appropriate operators and uses an ensemble system to classify future unknown instances. These features help in avoiding over- tting and class-imbalance problems, which are bene cial for improving generalization ability of a classi cation system. EDARIC has been tested on 30 standard and 33 imbalanced benchmark data-sets against more than 20 state-of-the-art evolutionary approaches and six state-of-the-art nonevolutionary approaches. EDARIC has also been tested against its own variant (without the intelligent deletion mechanism). The experimental results show that our proposed evolutionary system obtains better generalization performance compared to the existing algorithms. As expected, EDARIC also obtained better generalization performance than its own variant, which did not incorporate the intelligent deletion mechanism.