Summarizing time-evolving data

Abstract

We live in a highly dynamic environment where we have learned to appreciate time as a very important aspect in our lives. The world, as we know it, is not the same as it was yesterday and it will not be the same tomorrow. Things evolve over time. As we try to better understand our environment, we need to model this continuous change. To satisfy this need, the research community has developed temporal database systems. Summarizing time-evolving data is a challenging problem. Not only is this problem challenging because of the large size of the data sets usually involved but also because the temporal ordering and validity of every entry in the database must be considered. In this dissertation we present effective and efficient solutions to the problem of summarizing time-evolving data from two different perspectives. The first perspective considers capturing the collective behavior of temporal data in a process known as temporal aggregation. For this, we propose a model that reduces the evaluation of temporal aggregation queries to the problem of selecting qualifying tuples and grouping these tuples into collections to which an aggregate function is to be applied. In addition, we propose IO efficient algorithms for the evaluation of temporal aggregation queries. The second perspective we study is aimed toward the detection of individual entities in the database whose temporal behavior (evolution) matches a given pattern. The process of matching the temporal evolution of an entry in the database to a query pattern is known as similarity search. To address this problem, we propose a new indexing paradigm called Skyline Index and a new and compact representation of time series called Self COntained Bit Encoding (SCoBE). The techniques proposed in this dissertation provide significant performance improvements over the current state-of-the-art. Our empirical evidence shows that the proposed algorithms for the evaluation of temporal aggregation queries significantly outperform previous approaches. We experimentally show that the Skyline Index can be coupled with the state of the art dimensionality reduction techniques and significantly improve the performance on the evaluation of similarity search queries. Similarly, we show that SCoBE consistently outperforms previously proposed transformations of time series data for the evaluation of similarity search queries under a variety of scenarios.