GEP-Based Classification Algorithms

In 2001, F. Candida proposes a new Evolutionary Computation method named Gene Expression Programming (GEP) and it successfully serves for Classification task. However, F. Candida left some technical issues as open problems. For Example, classification-threshold can only be defined by experience, a long trivial classification way, etc. On the other hand, being a classical decision tree classification algorithm, C4.5 can quickly finish a classification task with good predication accuracy, but it can not deal with the relationship between attributes, which often confines its ability to discrete attributes data. This paper studies and implements the Gene Expression Programming-based Classification model, which is combined with the technologies of Centroid-based clustering technique and evolutionary computation advantages. The main contribution in this paper follows: proposes a dynamic threshold method, introduces a fitness function in terms of sensitivity and specificity, then gives a Gene Expression Programming-based quadrant classification algorithm; Proposes a new concept called Attribute Merge and a new algorithm called GEP-based Attribute Merge on this new concept, and it also includes tow improvement algorithms, the one is called Saving Best Gene for better convergence, the other is Sub-small Attribute Set for finding a sub optimal attribute set according to bisearch. The experiments show GEP-based quadrant classification algorithm simplifies the procedure while keeps the same or even better performance with the traditional GEP-based Classification algorithm and the latter compensates the shortcoming of C4.5.