An Effective Confidence-Based Early Classification of Time Series

Abstract

Early classification of time series aims to predict the class value of a sequence accurately as early as possible, not wait for the full-length data, which is significant in many time-sensitive applications and has attracted great interest in recent years. For instance, early diagnosis can help patients get early treatment and even save their lives. The problem of early classification is how to determine whether the collected data are sufficient to output the class value. Moreover, in practical applications, users also need to know the confidence (reliability) of the prediction results for more appropriate processing. For example, giving a healthy patient the possibility of suffering from some disease can assist physicians in an optimal therapy. However, existing work has not provided an effective measure to indicate how accurate the classification is. Therefore, in this paper, we propose an effective confidence-based early classification of time series. Firstly, based on a set of base time series classifiers trained at different timestamps, we propose a dynamic decision fusion method to measure the confidence of a predicted result by fusing the results of multiple base classifiers. Secondly, by analyzing the distribution of confidence values, we develop an adaptive learning method for the confidence threshold to simultaneously optimize the two conflicting objectives: accuracy and earliness. Finally, the experimental results conducted on 45 equal-length datasets and 8 variable-length datasets clearly show that our proposed approach can achieve the superior in early classification compared to state-of-the-art approaches.INDEX TERMS Time series classification, early classification, confidence measure, decision fusion. I. INTRODUCTION

MultiETSC: automated machine learning for early time series classification

“…This confidence is then used in the triggering mechanism. SR-CF (Mori et al 2018), TEASER (Schäfer and Leser 2020) and ECEC (Lv et al 2019) are all extensions of this idea using slightly different triggering mechanisms and prefix-classifiers. EARLIEST (Hartvigsen et al 2019) is based on a recurrent neural network (RNN) with LSTM cells.…”

Early classification of multivariate data by learning optimal decision rules

“…In the last few decades, various traditional time series classification methods have been studied [2], whereas, in recent times, early classification on TS data has received great research interest [12,23,25,39]. Thus, several methods have been reported in the literature to address an early classification problem such as instance-based learning [5,24,37], shapelet-based methods [10,14,38], model-based approaches [7,21,26], and other methods [12,18].…”

“…A similar approach is adopted in [33], in which the reliability threshold is defined based on uncertainty information in class prediction. Lv et al [23] developed a relatively similar framework, in which the confidence threshold was defined by fusing the classifier's true prediction probabilities at successive time steps. This framework is adaptable for both probabilistic as well as discriminative classifiers.…”

“…(ii) To find the balancing trade-off between the two objectives: accuracy and earliness. In view of these challenges, several approaches have been proposed in the literature [9,23,25,26,37,38] for early class prediction. For example, Xing et al [37], put forth a remarkable work for early classification on time series (ECTS) by introducing the concept of the minimum required length for class prediction based on learning nearest-neighbor relationship in the training dataset.…”

Corrections to “An Effective Confidence-Based Early Classification of Time Series”

“…In the above article [1], reference [2] was missing. One sentence should be appended to the penultimate paragraph of Section II: ''TEASER [2] treats early time series classification as a two-stage classification problem: A slave classifier classifies the time series and obtains the probabilistic results, similar to reference [3], and a master classifier uses the SVM classifier to determine whether the probabilistic results should be trusted or not.…”