Aiming at the optimization requirements of massive data transmission in data center optical network application scenarios, this paper proposes to use machine learning method to classify the traffic types to be sent before sending data center traffic, and introduces an improved greedy genetic algorithm(IGGA) into data center optical network reconstruction and optimization. The simulation results show that under the same node distribution, the average length of the improved IGGA is significantly reduced compared with the traditional genetic algorithm(GA), and the average length of the improved IGGA is reduced by 3.06% and 6.37% respectively in the case of 20 nodes and 50 nodes topology, and the improvement effect improves with the number of topology nodes increases.