Time series classification has attracted much attention in the last two decades. However, in many real-world applications, the acquisition of sufficient amounts of labeled training data is costly, while unlabeled data is usually easily to be obtained. In this paper, we study the problem of learning discriminative features (segments) from both labeled and unlabeled time series data. The discriminative segments are often referred to as shapelets. We present a new Semi-Supervised Shapelets Learning (SSSL for short) model to efficiently learn shapelets by using both labeled and unlabeled time series data. Briefly, SSSL engages both labeled and unlabeled time series data in an integrated model that considers the least squares regression, the power of the pseudo-labels, shapelets regularization, and spectral analysis. The experimental results on real-world data demonstrate the superiority of our approach over existing methods.