Selection of Relevant and Non-Redundant Multivariate Ordinal Patterns for Time Series Classification

Transformation of multivariate time series into feature spaces are common for data mining tasks like classification. Ordinality is one important property in time series that provides a qualitative representation of the underlying dynamic regime. In a multivariate time series, ordinalities from multiple dimensions combine together to be discriminative for the classification problem. However, existing works on ordinality do not address the multivariate nature of the time series. For multivariate ordinal patterns, there is a computational challenge with an explosion of pattern combinations, while not all patterns are relevant and provide novel information for the classification. In this work, we propose a technique for the extraction and selection of relevant and non-redundant multivariate ordinal patterns from the high-dimensional combinatorial search space. Our proposed approach Ordinal feature extraction (ordex), simultaneously extracts and scores the relevance and redundancy of ordinal patterns without training a classifier. As a filter-based approach, ordex aims to select a set of relevant patterns with complementary information. Hence, using our scoring function based on the principles of Chebyshev’s inequality, we maximize the relevance of the patterns and minimize the correlation between them. Our experiments on real world datasets show that ordinality in time series contains valuable information for classification in several applications.