Time Series Classification Methods Based On Deep Shapelet Learning

Time series classification is a research hotspot in the field of time series data mining.The essence of this task is to assign a predefined label to an input time series.Shapelets are discriminative subsequences of time series,thus they are useful for time series classification tasks.The traditional shapelet-based classification methods search shapelets from candidate subsequences.With the development of deep learning,there are some efforts to learn shapelets directly using the back-propagation algorithm,and have achieved good results.However,the current learning-based shapelet methods have the following issues.First,the learned shapelets are fixed after learning,which is not conducive to capturing important local patterns in the test data.Second,the shapelets learned through back-propagation may not be similar to any real subsequences,which reduces the interpretability of the shapelet method.Finally,the current shapelet learning methods tend to learn shapelets that can separate most samples of different categories rather than shapelets that can perfectly separate one class from other classes,and thus will ignore the minority class in imbalanced dataset.To address the above issues,this paper proposes two novel deep shapelet learning models.1)A novel deep shapelet learning model called Adversarial Dynamic Shapelet Networks(ADSNs)is proposed in this work.ADSNs use a shapelet generator to dynamically produce sample-specific shapelets conditioned on input time series,and then extract the feature representation of time series through shapelet transformation algorithm.In addition,ADSNs employ an adversarial training strategy to make the generated shapelets similar to actual subsequences.Thus,the proposed model has high modeling flexibility while retaining the interpretability of shapelet methods.Experiments conducted on extensive time series datasets prove the good performance of the model,and the visualization analysis shows the effectiveness of dynamic shapelet generation and adversarial training strategy.2)A novel deep multi-granularity shapelet learning model called Triple-shapelet Networks(TSNs)is proposed in this work.TSNs simultaneously learn three types of shapelets with different granularities.Specifically,in addition to using the back-propagation algorithm to learn general shapelets directly,TSNs also use auxiliary binary classifiers to learn the category-specific shapelets and use a shapelet generator to generate sample-specific shapelets conditioned on input data.These three different types of shapelets are used in the shapelet transformation algorithm simultaneously to extract the features representation of the input time series.The category-specific shapelets can improve the model performance on imbalanced datasets,while the sample-specific shapelets can improve the modeling flexibility.Experiments on the UCR datasets prove the performance of the model and the visualization analysis demonstrates the effectiveness of category-specific shapelets and sample-specific shapelets.