| Summary: | 碩士 === 輔仁大學 === 資訊管理學系 === 91 === In time series data mining problems, in order to recognize the characteristics held by data during certain interval, or to identify the state of data which always changes with time, it is necessary to divide time series into segments sometimes. In this thesis, the approach for time series segmentation is the technique which divides time series data into segments with various length, and which describes each segment with specific probability distribution function fitting it.
Even though there have some other time series segmentation algorithms been developed in literatures, they usually need parameters, such as number of segments or upper-bound of residual value in each segment, to be judged by experts. In such cases, not only some professionals have to be involved in time series segmentation tasks, but it is also difficult to avoid the subjective bias held by any person.
Opposite to the methods developed in the literature, we introduce a time series segmentation algorithm based on Ant Colony Optimization to divide a string of time series data into segments and to describe each segment with a probability distribution function that fits it. With Ant Colony Optimization based Segmentation algorithm, parameters used to be judged by experts can be automatically learned after the given number of trials.
After experimenting with data from stock market, it is verified that this algorithm can produce more accurate results than algorithms proposed previously, and causes less information loss than bottom-up algorithm in time series segmentation tasks.
|
|---|
