In order to maximize the total spectrum efficiency and improve the Quality of Experience (QoE) on Elastic Optical Network (EON), an adaptive modulation method based on Deep Neural Network (DNN) is proposed. Firstly, the network transmission quality in optical fiber is estimated. Then, the route length, hop number and video quality are taken as the three parameters of the request characteristics, and the deep learning method is adopted to selected the modulation system with the maximum spectrum efficiency according to the desired QoE of each requirement. The simulation results show that the average spectral efficiency of the proposed method is 51% and 32% higher in National Science Foundation (NSF) network than that of distance adaptive method and integer programming method, and 43% and 29% higher in China Network (CN)backbone network. In addition, the blocking probability of the proposed method is 0.01 lower than that of the distance adaptive method and integer programming method.