Discriminate Supervised Weighted Scheme for the Classification of Time Series Signals

Introduction

Time series is a sequence of data points observed in a continuous interval of time generated largely from various domains of applied science and engineering (Box et al., 2015). Time series analysis is a replica of knowledge discovery from data (KDD) (Zahra et al., 2020) that comprises of significant methods to extract meaningful statistics and other functional information in the time series signals (Fu 2011). KDD process offers several phenomenal services for time series analysis such as query by content (Faloutsos et al., 1994), anomaly detection (Weiss., 2004), classification (Bakshi and Stephanopoulos, 2004), motif discovery (Lin et al., 2004) and segmentation (Keogh et al., 2004) through innovative data mining techniques termed as Time Series Data Mining (TSDM).

In the last decade, several TSDM techniques have been proposed to solve Time Series Classification (TSC) problems due to the emerging of UCR time series repository (Chen et al., 2015). The TSC has applicability in wide range of domains (Ahmed et al, 2020) such as cardiovascular analysis (Kumar et al, 2019), human activity recognition (Trauth et al., 2018), motion sensor analysis, insect sound classification, etc. The proposed state-of-the-art techniques differ in the way they approach the problem such as the utilization of multivariate and univariate signals, extraction of exact, approximate, characteristic, representative patterns, the discovery of parameter-free, ill-known, frequent, significant, interesting, and variable length patterns, etc to efficiently solve the TSC problems as described in Elangovan and Padmavathi. (2019).

Among the techniques, the traditional k- Nearest Neighbor classifier (k-NN) (Willis et al, 2017) with Euclidean distance and Dynamic Time Warping (DTW) performs better due to its effectiveness and simplicity. Several other algorithms have also been proposed in the aspect shapelet based, model-based, and interval-based pattern discovery to extract the global and local features for classification. In which, Bag-of-Words (BoW), a simple and familiar text mining technique has been utilized by Lin et al in 2012 to identify the characteristic patterns in the time series signal for efficient classification. Consequently in later years Schafer. (2015), Senin and Malinchik. (2013) has extended the concept of BoW. With an extension of BoW, this paper proposes a discriminate Supervised Weighted Scheme (SWS) for the process of time series classification. The proposed technique has some similarities with the aforementioned techniques but it differs in searching discriminate representative pattern of a class using a modified weighted scheme rather than the traditional term frequency and inverse document frequency scheme. This discriminate representative pattern of class enables the superior interpretability in classification. This technique searches the representative pattern or word at once rather than the recursive search mechanism used by shapelet and rule based pattern discovery. In addition, the proposed technique uses single BoW for storing the entire corpus of all classes instead of handling individual bags for each class.

The remaining section of the paper has been organized as follows. The following section discusses the related work on various pattern discovery techniques for TSC problems. The next section discusses the proposed discriminative Supervised Weighted Scheme to identify the characteristic pattern for the process of efficient classification and it is followed by results and discussions. Finally, the paper concluded in the conclusion section.