Comparative study of complexity measures for time series classification and clustering

• Grant number: 13/03115-3
• Support type: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): May 01, 2013
• Effective date (End): October 31, 2013
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal researcher: Gustavo Enrique de Almeida Prado Alves Batista
• Grantee: Gabriela Pinto Cesar Duque
• Home Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

Recently, there is an increasing interest in time series processing due to the large number of application domains that generate data with such property. Such interest can be measured by the vast amount of methods recently proposed in the literature to tasks such as classification, clustering, summarization, abnormality detection, and motif discovery. Recent studies have shown for several problems that methods based on similarity present an efficacy that is hardly surpassed, even when compared to more sophisticated methods. This is mainly due to the fact that the community has studied and proposed several invariances to distance measures for time series. The invariances make the distance measures ignore certain undesired data properties. The most well-known example is the invariance to local differences in time scale, obtained with the warping technique. Other invariances include the invariance to differences in amplitude and offset phase and occlusion. Recently, we demonstrated to the scientific community that time series similarity classification methods can be largely benefited by a new invariance: complexity invariance. However, time series complexity can be measured by different means, such as auto-similarity (fractal dimension), entropy, Kolmogorov complexity, and compression, among others. The main objective of this research project is to investigate how different complexity measures can aid in the classification and clustering of time series. (AU)