TIME SERIES REPRESENTATION

Abstract

Efficient and accurate measure for time series data is very important and difficult task. A lot of solutions and been proposed for reducing the dimensionality of time series and have been found very useful as well, such as “Singular valued decomposition”, “discrete Fourier transform”, “Adaptive piecewise aggregate approximation (APAA)”, “segmented sum of variation”, “perpetually important points”, “symbolic aggregate approximation (SAX)”, “bit level representation”, “Major extrema”, “matrix representation”, etc.Among these techniques SAX is one of the most used and accurate techniques. Although it is one the mostly used techniques but is has its own disadvantages. It allows a very good dimensionality reduction and distance measure, but it does not solve the problem of data loss due to dimensionality reduction. This is because SAX is based on PAA which compresses the time series by taking equal sized segmented mean of the time series. This method provides good dimensionality reduction but leads to data loss. Due to this problem of data loss, SAX have been improved in many ways like Extended sax(E-SAX), Extreme SAX, genetic algorithm SAX(GASAX), Transitional SAX(TM-SAX), Trend SAX. This thesis focuses on the time series representation, where the time series must be compressed without losing information. Many techniques have been proposed (discussed in Chapter 2: Literature survey and research gaps). There are two challenges in time series representation, the first one being the compression and the second one being the distance measure. The compression should be done in a way such that the trend information loss in minimal. And the representation of time series should be able to perform comparison efficiently i.e., distance measure should be efficient. In Chapter 2, various techniques are discussed, and most the techniques are focusing on distance measure rather than the compression technique. In Chapter 3: Fast-SAX, one of the proposed works (Fast-SAX) is discussed where the compression technique is improved by using “major extrema” [8] algorithm. In Chapter 4: FT-SAX, a modification of Transitional SAX [14] is proposed where the execution time is reduced as compared to original technique by modifying algorithm procedure and using hashing with binary search. Chapter 5 has the conclusion for both Chapter 3 and Chapter 4. Chapter 6 has the Future work.