Abstract: Improving the accuracy of prediction of travel modes of residents is of a great importance to evaluate the effect of traffic planning and transport strategy.Based on psychology and behavior science, the decision process of travel modes is analyzed.With a structuralized process of decision, a library of travel scenarios is established.A principal component analysis is used to analyze the main factors which have impacts on the decision process of travel modes.The factors are regarded as the inputs of support vector machine (SVM).The differences between SVM and neural network in principles of modeling are analyzed by a statistical learning theory.Then a directed acyclic graph support vector machine (DAG-SVM) model is developed.The results of prediction from different kernel functions are evaluated, and the parameters are optimized by the grid method and genetic algorithm.The results show that among several kernel functions, the radial basis function is the best for prediction.The genetic algorithm is better than the grid method in parameter optimization.The overall accuracy of prediction from the DAG-SVM model is 82.3%, which is nearly 9% higher than that from the neural network model.However the accuracy of prediction for travel by taxi is slightly lower than other ones.This is mainly due to the fact that travel by taxi is an alternative way for residents in particular circumstances, not as regular as other travel modes.