Early classification of multivariate data by learning optimal decision rules

Abstract

Early classification on time series has emerged as an active research area in the field of machine learning. It covers a wide range of applications in agriculture, medical and multimedia systems, including drought prediction, health monitoring, event detection, and many more. The early classification aims to predict the class label of a time series as soon as possible without waiting for the complete series. A critical issue in early classification is the learning of decision policy that determines the adequacy of the collected data required for reliable class prediction. It is more challenging for Multivariate Time Series (MTS) data, where the decision depends on multiple variables to achieve a trade-off between earliness and accuracy. Therefore, this work proposes an optimization-based early classification model for MTS data based on optimal decision rule learning. The proposed model adopts a two-layered approach. The first layer employs the Gaussian process probabilistic classifiers for each variable in MTS that provides the class probabilities at the successive time steps in the series. The second layer defines Early Stopping Rule (ESR) that performs the class prediction task. The ESR learns its parameters through the particle swarm optimization by simultaneously minimizing the misclassification cost and delaying the decision cost. This work has utilized publicly available MTS datasets to validate the proposed early classification model. The experimental results show that the proposed model achieves promising results in terms of accuracy and earliness compared to existing methods.

Decoupled Early Time Series Classification Using Varied-Length Feature Augmentation and Gradient Projection Technique

“…Among them, the traditional methods leveraged hand-crafted features to train multiple classifiers, and set exiting strategies to quit classification. According to different exiting strategies [ 15 ], these methods can be divided into Prefix based [ 23 , 24 , 25 , 26 ], Shapelet based [ 27 , 28 , 29 , 30 ], and Model based [ 31 , 32 , 33 , 34 , 35 , 36 ] categories. Although these methods have achieved impressive performance and obtained extensive research, it is difficult to obtain the expert knowledge for constructing hand-crafted features as well as multiple classifiers.…”