Hybrid beamforming is a promising low-cost solution for large multiple-input multiple-output systems, where the base station is equipped with fewer radio frequency chains. In these systems, the selection of codewords for analog beamforming is essential to optimize the uplink sum rate. In this paper, based on machine learning, we propose a data-driven method of analog beam selection to achieve a near-optimal sum rate with low complexity, which is highly dependent on training data. Specifically, we take the beam selection problem as a multiclass-classification problem, where the training dataset consists of a large number of samples of the millimeter-wave channel. Using this training data, we exploit the support vector machine algorithm to obtain a statistical classification model, which maximizes the sum rate. For real-time transmissions, with the derived classification model, we can select, with low complexity, the optimal analog beam of each user. We also propose a novel method to determine the optimal parameter of Gaussian kernel function via McLaughlin expansion. Analysis and simulation results reveal that, as long as the training data are sufficient, the proposed data-driven method achieves a near-optimal sum-rate performance, while the complexity reduces by several orders of magnitude, compared to the conventional method.