Temporal Pattern Mining for Multivariate Time Series Classification
Temporal Pattern Mining for Multivariate Time Series Classification
复制标题
DOI：
10.1166/jmihi.2011.1019
复制
发表时间：
2011-06-01
影响因子：
--
通讯作者：
Singh, Harpreet
中科院分区：
文献类型：
--
作者：
Dua, Sumeet;Saini, Sheetal;Singh, Harpreet
时间序列数据存在于许多科学学科中，如金融、气象学、生物医学和航空航天。因此，时间序列数据的时态模式挖掘和基于时态模式的分类是必要的。在本文中，我们专注于时间模式挖掘的时间间隔为基础的事件数据和时间模式为基础的多变量时间序列数据集的分类。时间模式挖掘尚未广泛应用于癫痫发作检测。因此，我们测试我们的算法使用颅内脑电图数据在癫痫发作记录。首先，我们展示了现有的时间模式表示方案，如时间序列知识表示，TPrefixSpan，IEMiner的歧义。其次，我们解决这些时间模式表示方案的不足之处，提出了一个新的时间模式表示方案的时间间隔为基础的事件数据和开发一个类似Apriori算法的时间模式生成。此外，我们定义了时间规则的建议时间规则为基础的分类。对于我们的实验，我们应用时间模式挖掘癫痫发作颅内脑电图数据，以确定癫痫发作检测的发作和非发作条件。该分类器利用从多变量时间序列颅内脑电图（IEEG）数据中提取的多变量时间模式将事件分类为癫痫发作或非癫痫发作。
Time series data is found in many disciplines of science, such as finance, meteorology, biomedicine, and aerospace. Thus, efficient techniques for temporal pattern mining and temporal pattern-based classification of time series data are imperative. In this paper, we focus on temporal pattern mining for interval-based event data and temporal pattern-based classification for the multivariate time series datasets. Temporal pattern mining has not been applied extensively for epileptic seizure detection. Therefore, we test our algorithm using intracranial electroencephalography data generated during epileptic seizure recording. First, we demonstrate the ambiguities in existing temporal pattern representation schemes such as Time Series Knowledge Representation, TPrefixSpan, and IEMiner. Second, we address the deficiencies of these temporal pattern representation schemes by proposing a new temporal pattern representation scheme for interval-based event data and developing an apriori-like algorithm for temporal pattern generation. Furthermore, we define temporal rules for the proposed temporal rule-based classifier. For our experiments, we apply temporal pattern mining on epileptic seizure intracranial electroencephalography data to determine the seizure and non-seizure conditions for epileptic seizure detection. The proposed classifier exploits multivariate temporal patterns extracted from multivariate time series intracranial electroencephalography (IEEG) data to classify events as seizure or non-seizure occurrences.