Learning time-series shapelets

Abstract

Shapelets are discriminative sub-sequences of time series that best predict the target variable. For this reason, shapelet discovery has recently attracted considerable interest within the time-series research community. Currently shapelets are found by evaluating the prediction qualities of numerous candidates extracted from the series segments. In contrast to the state-of-the-art, this paper proposes a novel perspective in terms of learning shapelets. A new mathematical formalization of the task via a classification objective function is proposed and a tailored stochastic gradient learning algorithm is applied. The proposed method enables learning nearto-optimal shapelets directly without the need to try out lots of candidates. Furthermore, our method can learn true top-K shapelets by capturing their interaction. Extensive experimentation demonstrates statistically significant improvement in terms of wins and ranks against 13 baselines over 28 time-series datasets.

Time Series Classification by Sequence Learning in All-Subsequence Space

“…The shapelets are typically first extracted from the data, then used as a new representation in which to learn standard classifiers such as decision trees, random forest and SVM [16]. Approaches for finding the best shapelets vary from brute force search, to bounding quality metrics such as the information gain [15], searching in a lower dimensional SAX space [17], or learning discriminative fixedlength shapelets [7]. These methods deliver high accuracy but have a high computational complexity for training or testing and many are sensitive to noise [4].…”

“…This approach was originally designed for classification of sequences of discrete items, such as text or DNA. An important aspect of this approach is that it can efficiently select the best variable-length subsequences as driven by the training data and loss function (by combining learning and feature selection), and does not require a user to provide candidate subsequence lengths, such as in [7]. SEQL was shown to perform well in dense feature spaces and with very long sequences and large vocabularies.…”

“…This approach achieves accuracy comparable to 1NN-DTW with less computational cost, as only one centroid per class is needed during the testing stage, but has an expensive training stage. Other work aims to learn discriminative subsequences from raw numeric data (aka shapelets) [7] using an optimization formulation that involves searching for fixed-length time series subsequences that best predict the target variable. This approach has promising accuracy, but has a significant training time complexity of O(N 2 n 3 ), where N is the number of time series and n is the length of time series [4].…”

Discovering Malware with Time Series Shapelets

“…Despite the existence of many other approaches for time series classification [38], we use shapelets in our work, as (i) they find local and discriminative features from the data, (ii) they impose no assumptions on the nature of the data unlike autoregressive or ARIMA time series models [38,39] and they work even on non-stationary time series, (iii) they work on data instances of different lengths (unlike popular classifiers such as support vector machines, feed-forward neural networks, and random forests in their standard forms), (iv) they are easy to interpret and visualize for domain experts, and (v) they have been shown to be more accurate than other methods for some datasets [11,12,15,18,[20][21][22][23][24]27,39].…”

“…We consider a binary (two-class) classification scenario. Time series shapelets were first proposed by Ye and Keogh [39] and there have been optimizations on the initial method to make it faster or more advanced [12,15,18,20,24,27].…”