Mining Rules from Real-Valued Time Series: A Relative Information-Gain-Based Approach.
Yuanduo He,Guangju Peng,Xu Chu,Yasha Wang,Zhu Jin,Xiaorong Wang
DOI: https://doi.org/10.1109/compsac.2018.00061
2018-01-01
Abstract:Time series data is collected in almost every industrial field; mining knowledge from it has been attracting extensive attention in the data mining community. In this paper, we focus on temporal association rule mining from real-valued time series. Early work employs symbolization-based methods, but the symbolized representation misses out similarity of the original series, resulting in mining invalid rules. Although state-of-the-art work directly manipulates the original series, it may still find false rules due to the lack of correlation analysis. In our work, we present a hybrid approach combining the idea of direct manipulation and symbolization, which not only preserves the information about the raw data but also realizes the correlation analysis. Specifically, we leverage the similarity-preserving property of motifs, i.e. frequent occurring subsequences in time series, to partially symbolize the raw data. Then, for each rule candidate as a pair of motifs, we propose a rule searching framework to investigate the underlying relationships between them. To evaluate rule candidates, we accommodate the shape similarity by utilizing the relative information gain based on Minimum Description Length principle, and further develop a novel rule interestingness measure R_cos, which generalizes the classical measure cosine for association rules. We perform comprehensive experiments on both artificial and real world datasets, and the results show that the proposed rule searching framework and the rule interestingness measure are effective for mining valid temporal association rules from real-valued time series.