Abstract: Classification is an important sector of Data Mining,and decision tree is one of the efficient classification methods used constantly.Nowadays,thereare several classification algorithms which use decision tree,for instance ID3,C4.5 and CART.But there are some disadvantages to use them,for example,classset must be input manually,attribute must be separated and the best class setis needed.This paper makes the following contributions to avoid these disadvantages:on the one hand,proposing a new concept of using information gain to lineGEP chromosome's head;on the other hand,proposing the algorithms of IG-GEPDTAR (constructing decision tree attribution reduction algorithms with gene expression programming based on information gain) and validate it by using experiment data.The result shows that decision tree constructed by IG-GEPDTAR is absolutely correct and better,it has less redundancies than Candida Ferreira's,has 82.9%less nodes than that of ID3 algorithms,and has 31.2% less nodes than that of Wang Chunnian's.

Key words: GEP (gene expression programming), information gain, decision tree induction, entropy