Unsupervised Algorithm for Post-Processing of Roughly Segmented Categorical Time Series

Many types of existing collections often contain repeating sequences which could be called as patterns. If these patterns are recognized they can be for instance used in data compression or for prediction. Extraction of these patterns from data collections with components generated in equidistant time and in finite number of levels is now a trivial task. The problem arises for data collections that are subject to different types of distortions in all axes. This paper discusses possibilities of using the Voting Experts algorithm enhanced by the Dynamic Time Warping (DTW) method. This algorithm is used for searching characteristic patterns in collections that are subject to the previously mentioned distortions. By using the Voting Experts high precision cuts (but with low level of recall) are first created in the collection. These cuts are then processed using the DTW method to increase resulting recall. This algorithm has better quality indicators than the original Voting Experts algorithm.