SYSTEMS AND METHODS FOR STRUCTURAL CLUSTERING OF TIME SEQUENCES

0Associated
Cases 
0Associated
Defendants 
0Accused
Products 
5Forward
Citations 
0
Petitions 
3
Assignments
First Claim
1. . An apparatus comprising:
 a processor; and
a program storage device readable by machine, tangibly embodying a program of instructions executable by the processor, the program of instructions comprising;
instructions for using a periodogram for extracting period candidates from time series input to the apparatus;
instructions for verifying the period candidates using an autocorrelation function (ACF);
wherein a period candidate is verified responsive to a determination that the period candidate resides on a hill of the ACF; and
instructions for identify a local maximum of a period candidate determined to reside on a hill.
3 Assignments
0 Petitions
Accused Products
Abstract
Arrangements and methods for performing structural clustering between different time series. Time series data relating to a plurality of time series is accepted, structural features relating to the time series data are ascertained, and at least one distance between different time series via employing the structural features is determined. The different time series may be partitioned into clusters based on the at least one distance, and/or the k closest matches to a given time series query based on the at least one distance may be returned.
8 Citations
View as Search Results
Analyzing time series data that exhibits seasonal effects  
Patent #
US 7,865,389 B2
Filed 07/19/2007

Current Assignee
Meizu Technology Company Limited

Sponsoring Entity
HewlettPackard Development Company L.P.

Analyzing time series data that exhibits seasonal effects  
Patent #
US 20090024427A1
Filed 07/19/2007

Current Assignee
Meizu Technology Company Limited

Sponsoring Entity
Meizu Technology Company Limited

Insight and algorithmic clustering for automated synthesis  
Patent #
US 9,336,302 B1
Filed 03/14/2013

Current Assignee
OOL LLC

Sponsoring Entity
OOL LLC

Insight and algorithmic clustering for automated synthesis  
Patent #
US 9,607,023 B1
Filed 05/06/2016

Current Assignee
OOL LLC

Sponsoring Entity
OOL LLC

Insight and algorithmic clustering for automated synthesis  
Patent #
US 10,318,503 B1
Filed 03/27/2017

Current Assignee
OOL LLC

Sponsoring Entity
OOL LLC

Energy forecasting using model parameter estimation  
Patent #
US 20040102937A1
Filed 11/21/2002

Current Assignee
Honeywell International Inc.

Sponsoring Entity
Honeywell International Inc.

Feature classification for time series data  
Patent #
US 6,735,550 B1
Filed 01/15/2002

Current Assignee
University Corporation For Atmospheric Research

Sponsoring Entity
University Corporation For Atmospheric Research

Methods for estimating the seasonality of groups of similar items of commerce data sets based on historical sales data values and associated error information  
Patent #
US 6,834,266 B2
Filed 10/11/2001

Current Assignee
Oracle International Corporation

Sponsoring Entity
ProfitLogic Inc.

8 Claims
 1. . An apparatus comprising:
a processor; and a program storage device readable by machine, tangibly embodying a program of instructions executable by the processor, the program of instructions comprising; instructions for using a periodogram for extracting period candidates from time series input to the apparatus; instructions for verifying the period candidates using an autocorrelation function (ACF); wherein a period candidate is verified responsive to a determination that the period candidate resides on a hill of the ACF; and instructions for identify a local maximum of a period candidate determined to reside on a hill.  View Dependent Claims (2, 3, 4, 5, 6, 7, 8)
1 Specification
This application is a continuation application of copending U.S. patent application Ser. No. 12/115,824, filed May 6, 2008, which is a counterpart to U.S. patent application Ser. No. 12/115,166, filed on May 5, 2008, both of which are continuation applications of and claim priority to U.S. patent application Ser. No. 11/096,485, filed Mar. 31, 2005 and issued as U.S. Pat. No. 7,369,961; the contents of all of these applications are hereby incorporated by reference as if set forth in their entirety.
This invention was made with Government support under Contract No.: H98230043001 awarded by the U.S. Department of Defense. The Government has certain rights in this invention.
The present invention generally relates to the management of data relating to timeseries representations.
Herebelow, numerals set forth in square brackets—[ ]—are keyed to the list of references found towards the end of the present disclosure.
In recent years, in a constant effort to effect ongoing improvements in a crowded field of knowledge, there has been a profusion of timeseries distance measures and representations. The majority of these attempts to characterize the similarity between sequences is based solely on shape. However, it is becoming increasingly apparent that structural similarities can provide more intuitive sequence characterizations that adhere more tightly to human perception of similarity.
While shapebased similarity methods seek to identify homomorphic sequences using original raw data, structurebased methodologies are designed to find latent similarities, possibly by transforming the sequences into a new domain, where the resemblance can be more apparent.
Generally, an evolving need has been recognized in connection with providing an ever more effective and efficient manner of managing timeseries data.
Broadly contemplated herein, in accordance with at least one presently preferred embodiment of the present invention, are methods and arrangements considered for:
(i) efficiently capturing and characterizing (automatically) the periodicity of timeseries;
(ii) characterizing the periodic similarity of time series; and
(iii) combining the above methods to perform periodic clustering of timeseries, where the periodicities of each cluster are also provided.
Techniques such as those outlined above can be applicable in a variety of disciplines, such as manufacturing, natural sciences and medicine, which acquire and record large amounts of periodic data. For the analysis of such data, first there is preferably employed accurate periodicity estimation, which can be utilized either for anomaly detection or for prediction purposes. Then, a structural distance measure can preferably be deployed that can effectively incorporate the periodicity for quantifying the degree of similarity between sequences. It is recognized that a periodic measure can allow for more meaningful and accurate clustering and classification, and can also be used for interactive exploration (and visualization) of massive periodic datasets.
In summary, one aspect of the invention provides a method of performing structural clustering between different time series, said method comprising the steps of: accepting time series data relating to a plurality of time series; ascertaining structural features relating to the time series data; determining at least one distance between different time series via employing the structural features; and partitioning the different time series into clusters based on the at least one distance.
Another aspect of the invention provides an apparatus for performing structural clustering between different time series, said apparatus comprising: an arrangement for accepting time series data relating to a plurality of time series; an arrangement for ascertaining structural features relating to the time series data; an arrangement for determining at least one distance between different time series via employing the structural features; and an arrangement for partitioning the different time series into clusters based on the at least one distance.
A further aspect of the invention provides a program storage device readable by machine, tangibly embodying a program of instructions executed by the machine to perform method steps for performing structural clustering between different time series, said method comprising the steps of: accepting time series data relating to a plurality of time series; ascertaining structural features relating to the time series data; determining at least one distance between different time series via employing the structural features; and partitioning the different time series into clusters based on the at least one distance.
Yet another aspect of the invention provides a method of quantifying the structural similarity between different time series, said method comprising the steps of: accepting time series data relating to a plurality of time series; ascertaining structural features relating to the time series data; determining at least one distance between different time series via employing the structural features; and returning the k closest matches to a given time series query based on the at least one distance.
A yet further aspect of the invention provides an apparatus for quantifying the structural similarity between different time series, said apparatus comprising: an arrangement for accepting time series data relating to a plurality of time series; an arrangement for ascertaining structural features relating to the time series data; an arrangement for determining at least one distance between different time series via employing the structural features; and an arrangement for returning the k closest matches to a given time series query based on the at least one distance.
Furthermore, an additional aspect of the invention provides a program storage device readable by machine, tangibly embodying a program of instructions executed by the machine to perform method steps for quantifying the structural similarity between different time series, said method comprising the steps of: accepting time series data relating to a plurality of time series; ascertaining structural features relating to the time series data; determining at least one distance between different time series via employing the structural features; and returning the k closest matches to a given time series query based on the at least one distance.
For a better understanding of the present invention, together with other and further features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings, and the scope of the invention will be pointed out in the appended claims.
By way of background, provided herebelow is a brief introduction to harmonic analysis using the discrete Fourier Transform, because these tools will be used as the building blocks of algorithms in accordance with at least one embodiment of the present invention.
The normalized Discrete Fourier Transform of a sequence x(n), n=0, 1 . . . N−1 is a sequence of complex numbers X(f):
where the subscript k/N denotes the frequency that each coefficient captures. Herethroughout there will also be utilized the notation F(x) to describe the Fourier Transform. Since we are dealing with real signals, the Fourier coefficients are symmetric around the middle one (or to be more exact, they will be the complex conjugate of their symmetric). The Fourier transform represents the original signal as a linear combination of the complex sinusoids
Therefore, the Fourier coefficients record the amplitude and phase of these sinusoids, after signal x is projected on them.
One can return from the frequency domain back to the time domain, using the inverse Fourier transform F^{−1 }(x)≡x(n):
Note that if during this reverse transformation one discards some of the coefficients (e.g., the last k), then the outcome will be an approximation of the original sequence (see
In order to discover potential periodicities of a timeseries, one needs to examine its power spectral density (PSD or power spectrum). The PSD essentially tells us how much is the expected signal power at each frequency of the signal. Since period is the inverse of frequency, by identifying the frequencies that carry most of the energy, we can also discover the most dominant periods. There are two well known estimators of the PSD; the periodogram and the circular autocorrelation. Both of these methods can be computed using the DFT of a sequence (and can therefore exploit the Fast Fourier Transform for execution in O(N log N) time).
Suppose that X is the DFT of a sequence x. The periodogram P is provided by the squared length of each Fourier coefficient:
Where ∥.∥ denotes the L_{2 }norm of a vector. Notice that one can only detect frequencies that are at most half of the maximum signal frequency, due to Nyquist'"'"'s fundamental theorem. In order to find the k dominant periods, one should preferably pick the k largest values of the periodogram.
Each element of the periodogram provides the power at frequency kIN or, equivalently, at period N/k. Being more precise, each DFT ‘bin’ corresponds to a range of periods (or frequencies). That is, coefficient X(f_{k/N}) corresponds to periods
It is easy to see that the resolution of the periodogram becomes very coarse for longer periods. For example, for a sequence of length N=256, the DFT bin margins will be N/1, N/2, N/3, . . . =256, 128, 64 etc.
Essentially, the accuracy of the discovered periods deteriorates for large periods, due to the increasing width of the DFT bins (N/k). Another related issue is spectral leakage, which causes frequencies that are not integer multiples of the DFT bin width, to disperse over the entire spectrum. This can lead to ‘false alarms’ in the periodogram. However, the periodogram can still provide an accurate indicator of important short (to medium) length periods. Additionally, through the periodogram it is easy to automate the extraction of important periods (peaks) by examining the statistical properties of the Fourier coefficients.
The second way to estimate the dominant periods of a timeseries x, is to calculate the circular AutoCorrelation Function (or ACF), which examines how similar a sequence is to its previous values for different τ lags:
where the sum in n+τ is modulo N.
Therefore, the autocorrelation is formally a convolution, and one can avoid the quadratic calculation in the time domain by computing it efficiently as a dot product in the frequency domain using the normalized Fourier transform:
ACF=F
^{−1}
<X,X*>
The star (*) symbol denotes complex conjugation.
The ACF provides a more finegrained periodicity detector than the periodogram, hence it can pinpoint with greater accuracy even larger periods. However, it is not sufficient by itself for automatic periodicity discovery for the following reasons:
1. Automated discovery of important peaks is more difficult than in the periodogram, because the user must set a significance threshold.
2. Even if the user picks the level of significance, multiples of the same basic period also appear as peaks. Therefore, the method introduces many false alarms that need to be eliminated in a postprocessing phase.
3. Low amplitude events of high frequency may appear less important (i.e., have lower peaks) than high amplitude patterns, which nonetheless appear more scarcely (see
With relation to
The advantages and shortcomings of the periodogram and the ACF are summarized in Table 1.
From the above discussion one can realize that although the periodogram and the autocorrelation cannot provide sufficient spectral information separately, there is a lot of potential when both methods are combined. An approach in accordance with at least one preferred embodiment of the present invention is delineated in the following section.
In accordance with at least one preferred embodiment of the present invention, there is preferably utilized a twotier approach, by considering the information in both the autocorrelation and the periodogram. One may call this method AUTOPERIOD. Since the discovery of important periods is more difficult on the autocorrelation, one can use the periodogram for extracting period candidates. The period candidates may be termed ‘hints’. These ‘hints’ may be false (due to spectral leakage), or provide a coarse estimate of the period (remember that DFT bins increase gradually in size); therefore a verification phase using the autocorrelation is required, since it provides a more finegrained estimation of potential periodicities. The intuition is that if the candidate period from the periodogram lies on a hill of the ACF then one can consider it as a valid period, otherwise one may preferably discard it as false alarm. For the periods that reside on a hill, further refinement may be required if the periodicity hint refers to a large period.
Essentially, there has been leveraged the information of both metrics for providing an accurate periodicity detector. In addition, methods carried out in accordance with at least one embodiment of the present invention are computationally efficient, because both the periodogram and the ACF can be directly computed through the Fast Fourier Transform of the examined sequence in O(N log N) time.
For extracting a set of candidate periodicities from the periodogram, one needs to determine an appropriate power threshold that should distinguish only the dominant frequencies (or inversely the dominant periods). If none of the sequence frequencies exceeds the specific threshold (i.e., the set of periodicity ‘hints’ is empty), then one can regard the sequence as nonperiodic.
In order to specify which periods are important, one should first preferably identify how much of the signal energy is attributed to random mechanisms, that is, everything that could not have been attributed to a random process should be of interest.
Let us assume that one examines a sequence x. The outcome of a permutation on the elements of x is a sequence {tilde over (x)}. The new sequence will retain the first order statistics of the original sequence, but will not exhibit any pattern or periodicities, because of the ‘scrambling’ process (even though such characteristics may have existed in sequence x). Anything that has the structure of {tilde over (x)} is not interesting and should be discarded, therefore at this step one can record the maximum power (p_{max}) that {tilde over (x)} exhibits, at any frequency f.
Only if a frequency of x has more power than p_{max }can be considered interesting. If one would like to provide a 99% confidence interval on what frequencies are important, one should repeat the above experiment 100 times and record for each one the maximum power of the permuted sequence {tilde over (x)}. The 99^{th }largest value of these 100 experiments, will provide a sufficient estimator of the power threshold p_{T }being sought. Periods (in the original sequence periodogram) whose power is more than the derived threshold will be considered:
P_{hint}{N/k:P(f_{k/N})>p_{T}}
Finally, an additional period ‘trimming’ should be performed for discarding periods that are either too large or too small and therefore cannot be considered reliable. In this phase any periodic hint greater than N/2 or smaller than 2 is removed.
In [2] another algorithm for detection of important periods was proposed, which follows a different concept for estimating the periodogram threshold. The assumption there was that the periodogram of nonperiodic timeseries will follow an exponential distribution, which returned very intuitive period estimates for real world datasets. In experimentation, it has been found that the two algorithms return very comparable threshold values. However, because the new method does not make any assumptions about the underlying distribution, it can be applicable for a wider variety of timeseries processes.
By way of concrete examples, there were employed sequences from the MSN query logs (yearly span) to demonstrate the usefulness of the discovered periodic hints. In
After the periodogram peaks have been identified, there has been obtained a candidate set of periodicities for the examined sequence. The validity of these periods will be verified against the autocorrelation. An indication that a period is important, can be the fact that the corresponding period lies on a hill of the autocorrelation. If the period resides on a valley then it can be considered spurious and therefore safely discarded.
After discovering that a periodicity ‘hint’ resides on a hill of the autocorrelation, one can refine it even further by identifying the closest peak (i.e., local maximum). This is a necessary step, because the correct periodicity (i.e., peak of the hill) might not have been discovered by the periodogram, if it was derived from a ‘wide’ DFT bin. This is generally true for larger periods, where the resolution of the DFT bins drops significantly. Below is a discussion of how to address such issues.
The significance of a candidate period ideally can be determined by examining the curvature of the ACF around the candidate period p. The autocorrelation is concave downward, if the second derivative is negative in an open interval (a . . . b):
Nevertheless, small perturbations of the ACF due to the existence of noise, may invalidate the above requirement. There will be sought a more robust estimator of the curvature by approximating the ACF in the proximity of the candidate period with two linear segments. Then it is sufficient to examine if the approximating segments exhibit an upwarddownward trend, for identifying a concave downward pattern (i.e., a hill).
The segmentation of a sequence of length N into k linear segments can be computed optimally using a dynamic programming algorithm in O(N^{2}k) time, while a greedy merge algorithm achieves results very close to optimal in O(N log N) time. For this problem instance, however, one can employ a simpler algorithm, because only a two segment approximation for a specific portion of the ACF is required.
Let Ŝ_{a}^{b }be the linear regression of a sequence x between the positions [a . . . b]and
be the error introduced by the approximating segment. The best split position t_{split }is derived from the configuration that minimizes the total approximation error:
After it has been ascertained that a candidate period belongs on a hill and not on a valley of the ACF, there is a need to discover the closest peak which will return a more accurate estimate of the periodicity hint (particularly for larger periods). One can proceed in two ways; the first one would be to perform any hillclimbing technique, such as gradient ascent, for discovering the local maximum. In this manner the local search will be directed toward the positive direction of the first derivative. Alternatively, one could derive the peak position directly from the linear segmentation of the ACF, which is already computed in the hill detection phase. The peak should be located either at the end of the first segment or at the beginning of the second segment.
Both methods have been implemented for the purpose of experimentation and it has been found that both report accurate results.
Several sequences from the MSN query logs were employed to perform convincing experiments regarding the accuracy of a 2tier methodology in accordance with at least one embodiment of the present invention. The specific dataset is ideal for the present purposes because one can detect a number of different periodicities according to the demand pattern of each query.
The examples in
Query ‘Easter’ (MSN): Examining the demand for a period of 1000 days, one can discover several periodic hints above the power threshold in the periodogram. In this example, the autocorrelation information refines the original periodogram hint (from 333→357). Additional hints are rejected because they reside on ACF valleys (in the figure only the top 3 candidate periods are displayed for reasons of clarity).
 Query ‘Harry Potter’ (MSN): For the specific query although there are no observed periodicities (duration 365 days), the periodogram returns 3 periodic hints, which are mostly attributed to the burst pattern during November when the movie was released. The hints are classified as spurious upon verification with ACF.
Query ‘Fourier’ (MSN): This is an example where the periodogram threshold effectively does not return candidate periods. Notice that if one had utilized only the autocorrelation information, it would have been more troublesome to discover which (if any) periods were important. This represents another validation that the choice to perform the period thresholding in the frequency space was correct.
Economic Index (Stock Market): Finally, this last sequence from a stock market index illustrates a case where both the periodogram and autocorrelation information concur on the single (albeit weak) periodicity.
Through this experimental testbed it has been demonstrated that AUTOPERIOD can provide very accurate periodicity estimates without upsampling the original sequence. In the sections that follow, it will be shown how it can be used in conjunction with periodic similarity measures, for interactive exploration of sequence databases.
Structural measures can preferably be introduced that are based on periodic features extracted from sequences. Periodic distance measures can be used for providing more meaningful structural clustering and visualization of sequences (whether they are periodic or not). After sequences are grouped in ‘periodic’ clusters, using a ‘drilldown’ process the user can selectively apply the AUTOPERIOD method for periodicity estimation on the sequences or clusters of interest. In the discussion of experimentation examples of this methodology using hierarchical clustering trees are provided.
Let us consider first the utility of periodic distance measures with an example. Suppose that one is examining the similarity between the two timeseries of
The new measure of structural similarity presented herein exploits the power content of only the most dominant periods/frequencies. By considering the most powerful frequencies, the present method concentrates on the most important structural characteristics, effectively filtering out the negative influence of noise, and eventually allowing for expedited distance computation. Additionally, the omission of the phase information renders the new similarity measure shift invariant in the time domain. One can therefore discover timeseries with similar patterns, which may occur at different chronological instants.
For comparing the periodic structure of two sequences, one should preferably examine how different is their harmonic content. One may achieve this by utilizing the periodogram and specifically the frequencies with the highest energy.
Suppose that X is the Fourier transform of a sequence x with length n. One can discover the k largest coefficients of X by computing its periodogram P(X) and recording the position of the k frequencies with the highest power content (parameter k depends on the desired compression factor). Let us denote the vector holding the positions of the coefficients with the largest power p^{+} (so p^{+}⊂[1 . . . n]). To compare x with any other sequence q, one needs to examine how similar energies they carry in the dominant periods of x. Therefore, one preferably evaluates P(Q(p^{+})),that describes a sequence holding the equivalent coefficients as the vector P(X(p^{+})). The distance pDist between these two vectors captures the periodic similarity between sequences x and q:
pDist=∥P(Q(p^{+}))−P(X(p^{+}))∥
Example: Let x and q be two sequences and let their respective Fourier Transforms be X={(1+2i),(2+2i),(1+i),(5+1)} and Q={(2+2i),(1+i),(3+i),(1+2i)}. The periodogram vector of X is: P (X)=∥X∥^{2}=(5,8,2,26). The vector holding the positions of X with highest energy is p^{+}=(2,4) and therefore P(X(p^{+}))=(0,8,0,26). Finally, since P(Q)=(8,2,10,5) it follows that: P(Q(p^{+}))=(0,2,0,5)^{1}.
Alternatively, if one doesn'"'"'t want to provide a parameter k, one could extract from a sequence those periodic features that retain e % of the signal energy, or use the algorithm from the periodicity detection to extract the most important periods of a sequence.
In order to meaningfully compare the power content of two sequences one should preferably normalize them, so that they contain the same amount of total energy. One can assign to any sequence x(n) unit power, by performing the following normalization:
The above transformation will lead to zero mean value and sum of squared values equal to 1. Parseval'"'"'s theorem dictates that the energy in the time domain equals the energy in the frequency domain, therefore the total energy in the frequency domain should also be unit:
∥{circumflex over (x)}∥^{2}=∥F({circumflex over (x)})∥^{1}=1
After this normalization, one can more meaningfully compare the periodogram energies.
Presented herein are the results of extensive experimentation that show the usefulness of the new periodic measures, and the measures are compared with widely used shape based measures or newly introduced structural distance measures.
Using 16 sequences which record the yearly demand of several keywords at the MSN search engine, one may preferably perform the hierarchical clustering which is shown in
One can utilize an extended portion of the same dataset for exploring the visualization power of periodic distance measures. Using the pairwise distance matrix between a set of MSN keyword demand sequences (365 values, year 2002), there is evaluated a 2D mapping of the keywords using Multidimensional Scaling (
For a second experiment there is employed a combination of periodic time series that are collected from natural sciences, medicine and manufacturing, augmented by pairs of random noise and random walk data.
All datasets come in pairs, hence, when performing a hierarchical clustering algorithm on this dataset, one expects to find a direct linkage of each sequence pair at the lower level of the dendrogram. If this happens one may consider the clustering to be correct. The dataset is made up of 12 pairs, therefore a measure of the clustering accuracy can be the number of correct pair likings, over twelve, the number of total pairs.
For many datasets that came into 2 pairs (power demand, video surveillance, motor current), all 4 instances demonstrated the same basic period (as suggested by the AUTOPERIOD). However, the periodic measure can effectively separate them into two pairs, because the power content of the respective frequencies was different.
The last experiment is performed on the MITBIH Arrhythmia dataset. There are employed two sets of sequences; one with 2 classes of heartbeats and another one with three (
For both dataset instances, pDist again returns an accurate clustering, while DTW seems to perform badly on the high level dendrogram aggregations, hence not leading to perfect class separation. The Euclidean distance reported worse results. The CDM measure is accurate on the 2 class separation test but does not provide a perfect separation for the 3 class problem (see the original paper [1] for respective results).
By way of recapitulation, there have been presented herein various methods for accurate periodicity estimation and for the characterization of structural periodic similarity between sequences. It is believed that these methods will find many applications for interactive exploration of timeseries databases and for classification or anomaly detection of periodic sequences (e.g., in auto manufacturing, biometrics and medical diagnosis).
It is to be understood that the present invention, in accordance with at least one presently preferred embodiment, includes an arrangement for accepting time series data, an arrangement for ascertaining structural features relating to the time series data, and an arrangement for determining at least one distance between different time series and/or an arrangement for returning the k closest matches to a given time series query. Together, these elements may be implemented on at least one generalpurpose computer running suitable software programs. They may also be implemented on at least one integrated Circuit or part of at least one Integrated Circuit. Thus, it is to be understood that the invention may be implemented in hardware, software, or a combination of both.
If not otherwise stated herein, it is to be assumed that all patents, patent applications, patent publications and other publications (including webbased publications) mentioned and cited herein are hereby fully incorporated by reference herein as if set forth in their entirely herein.
Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be affected therein by one skilled in the art without departing from the scope or spirit of the invention.
 [1] E. Keogh, S. Lonardi, and A. Ratanamahatana. Towards parameterfree data mining. In Proc. of SIGKDD, 2004.
 [2] M. Vlachos, C. Meek, Z. Vagena, and D. Gunopulos. Identification of Similarities, Periodicities & Bursts for Online Search Queries. In Proc. of SIGMOD, 2004.