Unsupervised Segmentation of Categorical Time Series into Episodes

This paper describes an unsupervised olgorirhm f o r segmenting categorical time series inro episodes. The VOTING-EXPERTS algorithm first collects starisrics about the frequency and boundav entmpy of ngrams. then passes a window over rhe series and has two “expert methods ” decide where in rhe window boundaries should be drawn. The algorirhm successfully segments t a r into words in four languages. The algorithm also segments time series of mbot sensor data inro subsequences rhar represenr episodes in rhe life of rhe robot. We claim that VOTING-EXPERTS finds meaningful episodes in categorical time series because it exploits two statistical characteristics of meaningful episodes.