Abstract
Many applications in different domains generate time series data at an increasing rate. The continuous flow of emitted data may concern personal activities (e.g., through smart-meters or smart-plugs for electricity or water consumption) or professional activities (e.g., for monitoring heart activity or through the sensors installed on plants by farmers). This results in the production of large and complex data, usually in the form of time series.In recent years, there has been an explosion of interest in time series data mining. As a general rule, large time series come along with super-high dimensionality. As a consequence, it is difficult and inefficient to directly mine the raw time series without relying on dimensionality reduction. Therefore, the representation of the data is the key to efficient and effective solutions. Given this high data volumes in time series applications, or simply the need for fast response times, it is usually necessary to rely on alternative, shorter representations of these series, usually with loss. This incurs approximate comparisons of time series where precision is a major issue.In this thesis, we focus on the problem of segmenting time series before their transformation into symbolic representations. For this, we propose solutions to adaptively segment time series databases by adopting a variable segment size that depends on the time series distribution. These methods reduce significantly the information loss incurred by possible splittings at different steps of the representation calculation, particularly for datasets with unbalanced (non-uniform) distributions.We provide theoretical guarantees on the lower bound of similarity measures, and our results illustrate that our techniques can significantly improve the time series representation quality.