This study applies optimization-based algorithm to develop combination classification methods. We propose a genetic algorithm-based combination classification method of multiple decision trees to improve predictive accuracy, optimize classification rules, and interpret classification results. The basic algorithm for decision tree has been constructed in a top-down recursive divide-and-conquer manner. Based on different split measures (attribute selection measures), different decision tree algorithms can be produced, and then multiple decision trees can be formed. We proposed the parallel combination model of multiple decision trees. On top of the combination model, multiple decision trees are parallel combined. Each decision tree produces its own classification rules according to training samples from which one can present the classification result using probability distribution of target class label. At the bottom, the classification result of each decision tree serves as the input for combination algorithm in producing classification result and rules for the combination model. Combination algorithm adopts weighted summation of the outputs of probability measurement levels from individual decision trees, while genetic algorithm optimizes connection weight matrix. Finally the target class label with the largest probability output value is selected as the decision result for combination classification methods. The proposed method is applied to the issues of customer credit rating assessment and customer response behaviour pattern recognition in CRM. From the simulation results it is concluded that the proposed method has higher predictive accuracy than single decision tree. Moreover, it retains good interpretability and optimizes classification rules.