Mining discriminant temporal patterns is one problematic for the time series classification currently led by the shapelet. We expose this problematic under the angle of a standard feature-space classification task. This approach is enabled by the recent observation that most enumerable subsequences from a time series are redundant and can be discarded. In addition to be simple, the approach turns out to have state of-the-art classification performances with extremely fast computations. It also provides a flexible framework with interesting perspectives.