<p>Shapelet identification is crucial for time series classification as it provides discriminative features. Traditional methods often require exhaustive searches through a large pool of candidate subsequences, while learning-based approaches may lead to arbitrary, non-generalizable shapelets. In this paper, we propose a novel dynamic shapelet selection with soft labels(DSSL) for enhanced time series classification. DSSL dynamically selects the most relevant noise-enhanced shapelets for time series classification using an importance transfer strategy where feature importance is generated by the classification model and temporal importance is computed by an RNN-based model. These are integrated as soft labels to guide sample-specific shapelet selection. Experimental results show that our method outperforms eight comparison methods in accuracy and statistical significance, while effectively selecting key time points as shapelet candidates.</p>