Shapelet-based techniques are widely utilized in time series classification due to their combination of interpretability and accuracy. However, these methods tend to be less scalable than other state-of-the-art approaches. Because of this, we propose RSAST as a shapeled-based technique that utilizes a stratified technique and certain statistical criteria to randomly select the shapelets. We conducted experiments on 128 datasets from the UCR archive, showcasing the comparative accuracy of RSAST against its baseline methods SAST and STC, as well as other state-of-the-art techniques. Notably, the method we propose, RSAST, preserve the accuracy and interpretability of the baseline methods while reducing the computation time. For example, with Earthquakes, one of the largest datasets from the UCR archive, RSAST took 6 minutes and 36 seconds to complete the training process, whereas SAST took 2 hours and 37 minutes.