The invention discloses a method for traffic signal optimization based on a latent semantic model. According to the method, the advantages of the latent semantic model in processing uncertain factors in a recommendation system are utilized, a traffic state is subject to analog to be a user, timing plans are subject to analog to be articles, and traffic indexes (such as time delay) are subject to analog to be scores. The method comprises the steps of preprocessing data; mapping the traffic state and the timing plans to a latent semantic space; training a score prediction model of the traffic state for the timing plans; predicting the scores and selecting an optimal timing plan; and practically applying the timing plan and feeding back traffic information. The method is simple and practical, the uncertain factors which cannot be treated well through a traditional control strategy and have large influence in practice can be treated, the uncertain factors which cannot be accurately quantized easily do not need to undergo modeling accurately, and the method has very good universality and less constraints and can serve as supplementation and optimization of the traditional strategy.