Cylindrical algebraic decomposition (CAD) is one of the basic tools in computational real algebraic geometry. It has important applications in many fields. It's shown in both theory and practice that different variable orders have a great influence on the efficiency of CAD. Most of the existing algorithms for selecting variable orders for CAD are based on experience. Recently, some scholars have studied applying machine learning methods to select best empirical methods for variable order selection. Different from the existing methods, we apply machine learning to select variable order directly. Based on a graph structure of the polynomial system, we propose a group of new features. Experiments show that the multi-classifier trained with these features outperforms not only the random order selection method, but also the traditional heuristic method implemented by Maple command of SuggestVariableOrder.