Scholarlink

<p>Time series classification is essential in domains such as the Internet of Things (IoT), healthcare, and finance, supporting tasks including anomaly detection and patient monitoring. This paper presents <i>Autoregressive Time Series Classification</i> (ARTSC), a domain-independent framework that employs a compact and partially interpretable feature set comprising autoregressive coefficients, fitting error, and initial observations. Unlike domain-specific autoregressive methods or complex ensembles such as HIVE-COTE, ARTSC uses only 17 features—more than sixty times fewer than HIVE-COTE’s 1,000+—thereby achieving substantially higher computational efficiency. Lag and differencing parameters are selected through a grid search combined with KPSS stationarity testing, and classification is performed using Random Forests. In addition, a novel <i>Complexity Score</i> paradigm dynamically tunes model parameters according to dataset characteristics, enhancing efficiency without compromising accuracy. Evaluated on 80 UCR/UEA datasets, ARTSC achieves competitive performance with methods such as FCN and TSF, outperforming FCN on datasets including ECG200 (0.900 vs. 0.880) and FordA (0.972 vs. 0.958). Although performance decreases on highly nonlinear datasets (e.g., ShapeletSim, Wine), ARTSC’s lightweight and partially interpretable design suggests potential for resource-constrained IoT and healthcare applications, provided that further real-world validation, robustness analysis, and domain-specific safety assessment are conducted.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A simple, domain-independent autoregressive framework for time series classification

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Aykut Tayyip Altay

摘要

Time series classification is essential in domains such as the Internet of Things (IoT), healthcare, and finance, supporting tasks including anomaly detection and patient monitoring. This paper presents Autoregressive Time Series Classification (ARTSC), a domain-independent framework that employs a compact and partially interpretable feature set comprising autoregressive coefficients, fitting error, and initial observations. Unlike domain-specific autoregressive methods or complex ensembles such as HIVE-COTE, ARTSC uses only 17 features—more than sixty times fewer than HIVE-COTE’s 1,000+—thereby achieving substantially higher computational efficiency. Lag and differencing parameters are selected through a grid search combined with KPSS stationarity testing, and classification is performed using Random Forests. In addition, a novel Complexity Score paradigm dynamically tunes model parameters according to dataset characteristics, enhancing efficiency without compromising accuracy. Evaluated on 80 UCR/UEA datasets, ARTSC achieves competitive performance with methods such as FCN and TSF, outperforming FCN on datasets including ECG200 (0.900 vs. 0.880) and FordA (0.972 vs. 0.958). Although performance decreases on highly nonlinear datasets (e.g., ShapeletSim, Wine), ARTSC’s lightweight and partially interpretable design suggests potential for resource-constrained IoT and healthcare applications, provided that further real-world validation, robustness analysis, and domain-specific safety assessment are conducted.