PROJECT TITLE :

Efficient Shapelet Discovery for Time Series Classification

ABSTRACT:

Recently, it was discovered that time-series shapelets, which are discriminative subsequences, are effective for the classification of time series ( tsc ). It should come as no surprise that the accuracy of tsc is directly correlated to the quality of the shapelets. However, a significant amount of research has been directed toward developing accurate models from a variety of shapelet candidates. Existing studies, which are used to determine such candidates, are surprisingly straightforward. For example, they may involve enumerating subsequences of some fixed lengths or randomly selecting some subsequences as candidates for shapelets. After that, the construction of the model based on the candidates takes up the vast majority of the computational time. In this paper, we propose a brand new efficient shapelet discovery method that we call bspcover. The goal of this method is to find a set of shapelet candidates that are of a high quality and can be used for model building. To be more specific, bspcover generates a large number of candidates by using Symbolic Aggregate approXimation with a sliding window. It then uses Bloom filters and similarity matching, respectively, to remove candidates who are identical or highly similar to one another. Next, we present a pp-Cover algorithm with the goal of quickly determining discriminative shapelet candidates that most adequately represent each time-series class. In conclusion, a classification model can be constructed using any shapelet learning method that is already in existence. Extensive tests with well-known time-series datasets and representative examples of state-of-the-art methods have been carried out by our team. According to the findings, bspcover is more than 70 times faster than the methods that are currently considered to be state-of-the-art, and its accuracy is frequently on par with or even higher than that of existing works.