MDL-based time series clustering

摘要

Time series data are pervasive across all human endeavors, and clustering is arguably the most fundamental data mining application. Given this, it is somewhat surprising that the problem of time series clustering from a single stream remains largely unsolved. Most work on time series clustering considers the clustering of individual time series that have been carefully extracted from their original context, for example, gene expression profiles, individual heartbeats, or individual gait cycles. The few attempts at clustering time series streams have been shown to be objectively incorrect in some cases, and in other cases shown to work only on the most contrived synthetic datasets by carefully adjusting a large set of parameters. In this work, we make two fundamental contributions that allow for the first time, the meaningful clustering of subsequences from a time series stream. First, we show that the problem definition for time series clustering from streams currently used is inherently flawed, and a new definition is necessary. Second, we show that the minimum description length framework offers an efficient, effective, and essentially parameter-free method for time series clustering. We show that our method produces objectively correct results on a wide variety of datasets from medicine, speech recognition, zoology, gesture recognition, and industrial process analyses.