... mining time series data. Literally hundreds of papers have introduced new algorithms to index, classify, cluster and segment time series. In this work we make the following claim. Much of this work has very little utility because the contribution made (speed in the case of indexing, accuracy in the case of classification and clustering, model accuracy in the case of segmentation) offer an amount of "improvement" that would have been completely dwarfed by the variance that would have been observed by testing on many real world datasets, or the variance that would have been observed by changing minor (unstated) implementation details. To illustrate our point

Time series data is perhaps the most frequently encountered type of data examined by the data mining community. Clustering is perhaps the most frequently used data mining algorithm, being useful in it’s own right as an exploratory technique, and also as a subroutine in more complex data mining algorithms such as rule discovery, indexing, summarization, anomaly detection, and classification. Given these two facts, it is hardly surprising that time series clustering has attracted much attention. The data to be clustered can be in one of two formats: many individual time series, or a single time series, from which individual time series are extracted with a sliding window. Given the recent explosion of interest in streaming data and online algorithms, the latter case has received much attention. In this work we make a surprising claim. Clustering of streaming time series is completely meaningless. More concretely, clusters extracted from streaming time series are forced to obey a certain constraint that is pathologically unlikely to be satisfied by any dataset, and because of this, the clusters extracted by any clustering algorithm are essentially random. While this constraint can be intuitively demonstrated with a simple illustration and is simple to prove, it has never appeared in the literature. We can justify calling our claim surprising, since it invalidates the contribution of dozens of previously published papers. We will justify our claim with a theorem, illustrative examples, and a comprehensive set of experiments on reimplementations of previous work. Although the primary contribution of our work is to draw attention to the fact that an apparent solution to an important problem is incorrect and should no longer be used, we also introduce a novel method which, based on the concept of time series motifs, is able to meaningfully cluster some streaming time series datasets.

The problem of efficiently locating previously known patterns in a time series database (i.e., query by content) has received much attention and may now largely be regarded as a solved problem. However, from a knowledge discovery viewpoint, a more interesting problem is the enumeration of previously unknown, frequently occurring patterns. We call such patterns "motifs," because of their close analogy to their discrete counterparts in computation biology. An efficient motif discovery algorithm for time series would be useful as a tool for summarizing and visualizing massive time series databases. In addition, it could be used as a subroutine in various other data mining tasks, including the discovery of association rules, clustering and classification. In this work we carefully motivate, then introduce, a non-trivial definition of time series motifs. We propose an efficient algorithm to discover them, and we demonstrate the utility and efficiency of our approach on several real world datasets.

We introduce a novel wavelet-based tree structure, termed TSA-tree, which improves the efficiency of multilevel trend and surprise queries on time sequence data. With the explosion of scientific observation data (some conceptualized as time-sequences), we are facing the challenge of efficiently storing, retrieving and analyzing this data. Frequent queries on this data set is to find trends (e.g., global warming) or surprises (e.g., undersea volcano eruption) within the original time-series. The challenge, however, is that these trend and surprise queries are needed at different levels of abstractions (e.g., within the last week, last month, last year or last decade). To support these multi-level trend and surprise queries, sometimes huge subset of raw data needs to be retrieved and processed. To

Abstract. A moving cluster is defined by a set of objects that move close to each other for a long time interval. Real-life examples are a group of migrating animals, a convoy of cars moving in a city, etc. We study the discovery of moving clusters in a database of object trajectories. The difference of this problem compared to clustering trajectories and mining movement patterns is that the identity of a moving cluster remains unchanged while its location and content may change over time. For example, while a group of animals are migrating, some animals may leave the group or new animals may enter it. We provide a formal definition for moving clusters and describe three algorithms for their automatic discovery: (i) a straight-forward method based on the definition, (ii) a more efficient method which avoids redundant checks and (iii) an approximate algorithm which trades accuracy for speed by borrowing ideas from the MPEG-2 video encoding. The experimental results demonstrate the efficiency of our techniques and their applicability to large spatio-temporal datasets. 1

The problem of efficiently locating previously known patterns in a time series database (i.e., query by content) has received much attention and may now largely be regarded as a solved problem. However, from a knowledge discovery viewpoint, a more interesting problem is the enumeration of previously unknown, frequently occurring patterns. We call such patterns "motifs", because of their close analogy to their discrete counterparts in computation biology. An efficient motif discovery algorithm for time series would be useful as a tool for summarizing and visualizing massive time series databases. In addition it could be used as a subroutine in various other data mining tasks, including the discovery of association rules, clustering and classification.

The emergence of the field of data mining in the last decade has sparked an increasing interest in clustering of tiate series. Although there has been much research on clustering in general, most classic machine learning and data mining algorithms do not work well for time series due to their unique structure. In particular, the high dimensionaliF, very high feature correlation, and the (typically) large amount of noise that characterize time series data present a difficult challenge. In this work we address these challenges by introducing a novel anytiate version of k-Means clustering algorithm for time series. The algorithm works by leveraging off the multi-resolution property of wavelets. In particular, an initial clustering is perforated with a very coarse resolution representation of the data. The results obtained from this "quick and dirty" clustering are used to initialize a clustering at a slightly finer level of approximation. This process is repeated until the clustering results stabilize or until the "approxiatation" is the raw data. In addition to casting k-Means as an anytime algorithm, our approach has two other very unintuitive properties. The quality of the clustering is often better than the batch algorithm, and even if the algorithm is run to coatpletion, the time taken is typically much less than the time taken by the original algorithm. We explain, and eatpirically demonstrate these surprising and desirable properties with coatprehensive experiatents on several publicly available real data sets.

An overview on the usage of emergent self organizing maps is given. U-Maps visualize the distance structures of high dimensional data sets. P-Maps show their density structures and U*-Maps combine the advantages of the mentioned maps to a visualization suitable to detect nontrivial cluster structures. A concise summary on the usage of Emergent Self-organizing Maps (ESOM) for data mining is given. The tasks of visualization, clustering, and classification as they can be performed with the Databionics ESOM Tools are described. 1

We describe a procedure for identifying major minima and maxima of a time series, and present two applications of this procedure. The first application is fast compression of a series, by selecting major extrema and discarding the other points. The compression algorithm runs in linear time and takes constant memory. The second application is indexing of compressed series by their major extrema, and retrieval of series similar to a given pattern. The retrieval procedure searches for the series whose compressed representation is similar to the compressed pattern. It allows the user to control the trade-off between the speed and accuracy of retrieval. We show the effectiveness of the compression and retrieval for stock charts, meteorological data, and electroencephalograms. Keywords. Time series, compression, fast retrieval, similarity measures. 1

We describe a technique for fast compression of time series and indexing of compressed series. We have tested it on three data sets: stock prices, air and sea temperatures, and wind speeds.