Institute of Information Science, Academia Sinica

The problem of time series classification has been studied for over a decade. In the era of Internet of Things, time series data has become a major data type data, much effort has been devoted to this issue. The approaches to time series classification can be categorized into three types, including distance-based, model-based, and feature-based approaches. In this research, we focus on feature-based methods, which represent time series into a set of characterized values. However, features generated by most of existing representation techniques are not completely interpretable. Due to this fact, a novel time series representation, envelope, is proposed. The envelope is a profiling for a set of time series. This is a supervised feature extraction method that encodes time series into three numbers, -1, 0 and 1. If time series value falls into the envelope then encodes it as 0. We use -1 and 1 to represent the value falls outside below and above respectively. It is always important to find the most discriminating features for data mining tasks. Hence, we need a good heuristic to decide the size of the envelope in order to have a better performance either in data classification or anomaly detection tasks. Moreover, this new representation enjoys the characteristic of sparsity which is an essential property for applying compressed sensing. With this advantage, we can benefit from high transmission efficiency, the reduction of required storage and model complexity. Furthermore, the transformed features are interpretable via visualization. Envelope shows the shape of time series and defines the similarity between which. Disclosed below is the demonstration of the effectiveness of proposed method on numerous benchmark datasets.