Trend similarity and prediction in time-series databases

Many algorithms for discovering similar patterns from time- series databases involve three phases: First, sequential data in time domain is transformed into frequency domain using DFT. Then, the first few data points are considered to depict in an R*-tree. Those points in an R*-tree are compared by their distance. Any pair of data points, if the distance between them is within a certain threshold, are found to be similar. This approach results in performance problem due to emphasis on each data point itself. This paper proposes a novel method of finding similar trend patterns, rather than similar data patterns, from time-series database. As opposed to similar data patterns in the frequency domain, a limited number of points, in the time series, that play a dominant role to make a movement direction are taken into account. Those data points are called a trend sequence. Trend sequences will be defined in various ways. Of many, we focus more on considering trend sequences by a data smoothing technique. We know that a trend sequence contains far fewer data points than an original data sequence, but entails abstract level of sequence movements. To some extent, given a trend sequence, we apply the smoothing algorithm to predict the very next trend data. It is likely that once a trend sequence is found, the very next trend data point is expected. This paper also shows a method for trend prediction. We observed that our approach presented in this paper can be applied to finding similarity among many large time-series data sequences to the prediction of next possible data points to follow.