When noise model is already known, maximum likelihood estimator (MLE) is asymptotically the most optimum one. However, the truth is just the opposite that, noise is unknown in burst communication systems. Aiming that, this paper utilizes the transmitted symbols of burst communications which are eliminated firstly in common methods and proposes a data-aided (DA) frequency estimation algorithm based on least squares support vector classification (LS-SVC). By researching on statistical learning theory (SLT), we construct a structural risk minimization (SRM) function with respect to frequency, and convert the estimation problem into deriving the extremum value of a classification function. Consequently, support vector classification (SVC)’s good learning and generalization capabilities are completely explored and employed. Experimental results show that the proposed algorithm is close to MLE in the case of Gaussian noise, and also exhits good performance in non-Gaussian condition.