Transcript Time-Series Data Management

Time-Series Data Management
Yonsei University
2nd Semester, 2014
Sanghyun Park
* The slides were extracted from the material presented at ICDM’01
by Eamonn Keogh
Contents




Introduction, motivation
Utility of similarity measurements
Indexing time series
Summary, conclusions
What Are Time Series?

A time series is a collection of observations made
sequentially in time
25.2250
25.2500
25.2500
25.2750
25.3250
25.3500
25.3500
25.4000
25.4000
25.3250
25.2250
25.2000
25.1750
..
..
24.6250
24.6750
24.6750
24.6250
24.6250
24.6250
24.6750
24.7500
29
28
27
26
25
24
23
0
50
100
150
200
250
300
350
400
450
500
Time Series Are Ubiquitous (1/2)

People measure things …






The presidents approval rating
Their blood pressure
The annual rainfall in Riverside
The value of their Yahoo stock
The number of web hits per second
And things change over time and thus time series occur
in virtually every medical, scientific and business domain
Time Series Are Ubiquitous (2/2)

A random sample of 4,000 graphics from 15 of the
world’s newspapers published from 1974 to 1989 found
that more than 75% of all graphics were time series
Time Series Similarity

Defining the similarity between two time series is at the
heart of most time series data mining applications/tasks

Thus time series similarity will be the primary focus of
this lecture
Utility Of Similarity Search (1/2)
Classification
Clustering
Utility Of Similarity Search (2/2)
Rule Discovery

s = 0.5
10
c = 0.3
Query by Content
Query Q
(template)
1
6
2
7
3
8
4
9
5
10
Database C
Challenges Of Research On Time Series
(1/3)

How do we work with very large databases?





1 hour of ECG data: 1 gigabyte
Typical web log: 5 gigabytes per week
Space shuttle database: 158 gigabytes and growing
Macho database: 2 terabytes, updated with 3 gigabytes per day
Since most of the data lives on disk (or tape), we need a
representation of the data we can efficiently manipulate
Challenges Of Research On Time Series
(2/3)

We are dealing with subjective notions of similarity

The definition of similarity depends on the user, the
domain, and the task at hand. We need to handle this
subjectivity
Challenges Of Research On Time Series
(3/3)

Miscellaneous data handling problems



Differing data formats
Differing sampling rates
Noise, missing values, etc
Whole Matching vs.
Subsequence Matching (1/2)

Whole matching
Given a query Q, a reference database C, and
a distance measure, find the Ci that best matches Q
Query Q
(template)
1
6
2
7
3
8
4
9
5
10
Database C
C6 is the best
match
Whole Matching vs.
Subsequence Matching (2/2)

Subsequence matching
Given a query Q, a reference database C, and a
distance measure, find the location that best matches Q
Query Q
(template)
Database C
The best matching
subsection
Motivation Of Similarity Search

You go to the doctor because of chest pains. Your ECG
looks strange …

Your doctor wants to search a database to find similar
ECGs, in the hope that they will offer clues about your
condition …

Two questions


How do we define similar?
How do we search quickly?
Defining Distance Measures

Definition: Let O1 and O2 be two objects from the
universe of possible objects. Their distance is denoted
as D(O1,O2)

What properties should a distance measure have?




D(A,B) = D(B,A)
D(A,A) = 0
D(A,B) = 0 IIf A=B
D(A,B) ≤ D(A,C) + D(B,C)
Symmetry
Constancy of self-similarity
Positivity
Triangluar inequality
The Minkowski Metrics
DQ, C  
 qi  ci 
n
p
p
i 1
p = 1 Manhattan (Rectilinear, City Block)
p = 2 Euclidean
p =  Max (Supremum, “sup”)
D(Q,C)
Euclidean Distance Metric

Given two time series
Q=q1…qn and C=c1…cn,
their Euclidean distance
is defined as:
DQ, C    qi  ci 
n
C
Q
2
i 1
D(Q,C)
Processing The Data Before
Distance Calculation

If we naively try to measure the distance between two
“raw” time series, we may get very unintuitive results

This is because Euclidean distance is very sensitive to
some distortions in the data

For most problems these distortions are not meaningful,
and thus we can and should remove them

Four most common distortions




Offset translation
Amplitude scaling
Linear trend
Noise
Offset Translation
3
3
2.5
2.5
2
2
1.5
1.5
1
1
0.5
0.5
0
0
50
100
150
200
250
300
0
D(Q,C)
0
50
100
150
200
250
300
Q = Q - mean(Q)
C = C - mean(C)
D(Q,C)
0
0
50
100
150
200
250
300
50
100
150
200
250
300
Amplitude Scaling
0
100
200
300
400
500
600
700
800
900 1000
0
100
200
300
400
500
600
700
800
900 1000
Q = (Q - mean(Q)) / std(Q)
C = (C - mean(C)) / std(C)
D(Q,C)
Linear Trend
5
4
Removed offset translation
3
2
Removed amplitude scaling
1
0
12
-1
10
-2
8
-3
0
20
40
60
80
100
120
140
160
180
200
6
4
2
0
5
-2
-4
0
4
20
40
60
80
100
120
140
160
180
200
Removed linear trend
3
The intuition behind removing
linear trend is this:
2
Removed offset translation
1
0
Fit the best fitting straight line to
the time series, then subtract
that line from the time series
Removed amplitude scaling
-1
-2
-3
0
20
40
60
80
100
120
140
160
180
200
Noise
8
8
6
6
4
4
2
2
0
0
-2
-2
-4
0
20
40
60
80
100
120
140
-4
0
20
40
60
80
The intuition behind removing
noise is this:
Q = smooth(Q)
Average each datapoint value
with its neighbors
C = smooth(C)
D(Q,C)
100
120
140
Dynamic Time Warping

We will first see the utility of DTW, then see how it is
calculated
Fixed Time Axis
“Warped” Time Axis
Sequences are aligned “one to one”.
Nonlinear alignments are possible.
Utility of DTW: Example I,
Machine Learning
Cylinder-Bell-Funnel
Cylinder
Funnel
Bell

This dataset has been studied in a machine learning
context by many researchers

Recall that, by definition, the instances of Cylinder-BellFunnel are warped in the time axis
Classification Experiment on
C-B-F Dataset (1/2)

Experimental settings




Training data consists of 10 exemplars from each class
(One) Nearest neighbor algorithm
“Leaving-one-out” evaluation, averaged over 100 runs
Results




Error rate using Euclidean Distance: 26.10%
Error rate using Dynamic Time Warping: 2.87%
Time to classify one instance using Euclidean Distance: 1 sec
Time to classify one instance using Dynamic Time Warping:
4,320 sec
Classification Experiment on
C-B-F Dataset (2/2)

Dynamic time warping can reduce the error rate by an
order of magnitude

Its classification accuracy is competitive with
sophisticated approaches like decision tree, boosting,
neural networks, and Bayesian techniques

But, it is slow …
Utility of DTW: Example II,
Data Mining

Power-demand time series: each sequence
corresponds to a week’s demand for power in a Dutch
research facility in 1997
Wednesday was
a national
holiday
Hierarchical Clustering with
Euclidean Distance
4
5
3
The two 5-day weeks are correctly
grouped.
6
7
Note however, that the three 4-day
weeks are not clustered together.
Also, the two 3-day weeks are also not
clustered together.
2
1
Hierarchical Clustering with
Dynamic Time Warping
6
4
7
5
The two 5-day weeks are correctly
grouped.
3
The three 4-day weeks are clustered
together.
The two 3-day weeks are also
clustered together.
2
1
Time Taken to Create Hierarchical
Clustering of Power-Demand Time Series

Time to create dendrogram using Euclidean Distance:
1.2 seconds

Time to create dendrogram using Dynamic Time
Warping: 3.40 hours
Computing the Dynamic Time Warp
Distance (1/2)
Note that the input sequences can be of different
lengths
Q
|p|
C
Q
w
p
k
j
C

1
w1
1
i
n
|n|
Computing the Dynamic Time Warp
Distance (2/2)
Q
|p|
|n|
C

Every possible mapping from Q to C can be
represented as a warping path in the search matrix

We simply want to find the cheapest one …
DTW (Q, C ) min 



K
k 1
wk K
Although there are exponentially many such paths,
we can find one in only quadratic time using dynamic
programming
(i,j) = d(qi,cj) + min{ (i-1,j-1) , (i-1,j ) , (i,j-1) }
Fast Approximation to Dynamic Time
Warping Distance (1/2)

Simple idea: approximate the time series with some
compressed or downsampled representation, and do
DTW on the new representation
Q
wk
p
C
j
1

w1
1
i
How well does this work …
n
Fast Approximation to Dynamic Time
Warping Distance (2/2)
22.7 sec
1.3 sec

… Strong visual evidence to suggest it works well
Weighted Distance Measures (1/3)

Intuition: for some queries different parts of the
sequence are more important
Weighted Distance Measures (2/3)
DQ, C  
DQ, C ,W  
D(Q,C)
2


q

c
 i i
n
i 1
 wi qi  ci 
n
2
D(Q,C,W)
i 1
The height of this histogram
indicates the relative importance
of that part of the query
W
Weighted Distance Measures (3/3)

How do we set the weights?

One possibility: relevance feedback which is the
reformulation of a query in response to feedback
provided by the user for the results of previous query
Term Vector
Term Weights
[Jordan , Cow, Bull, River]
[
1 , 1 , 1 , 1 ]
Search
Display Results
Gather Feedback
Term Vector
Term Weights
[Jordan , Cow, Bull, River]
[ 1.1 , 1.7 , 0.3 , 0.9 ]
Update Weights
Indexing Time Series (1/6)

We have seen techniques for assessing the similarity of
two time series

However we have not addressed the problem of finding
the best match to a query in a large database …

The obvious solution, to retrieve
and examine every item
(sequential scanning), simply
does not scale to large datasets

We need some way to index
the data
1
6
2
7
3
8
4
9
5
10
Database C
Indexing Time Series (2/6)

We can project time series of length n into n-dimension
space

The first value in C is the X-axis, the second value in C
is the Y-axis, etc.

One advantage of doing this is
that we have abstracted away
the details of “time series”,
now all query processing can
be imagined as finding points
in space …
Indexing Time Series (3/6)

We can project the query time series Q into the same ndimension space and simply look for the nearest points
Q

The problem is that we have to look at every point to
find the nearest neighbor
Indexing Time Series (4/6)

The Minkowski metrics have simple geometric
interpolations
Euclidean
Weighted Euclidean
Manhattan
Max
Indexing Time Series (5/6)

We can group clusters of datapoints with “boxes” called
Minimum Bounding Rectangles (MBR)
R1
R2
R4
R5
R3
R6
R9
R7
R8

We can further recursively group MBRs into larger
MBRs
Indexing Time Series (6/6)

These nested MBRs are organized as a tree (called a
spatial access tree or a multidimensional tree).
Examples include R-tree, Hybrid-tree, etc.
R10
R11
R10 R11 R12
R1 R2 R3
R12
R4 R5 R6
R7 R8 R9
Data nodes containing points
Dimensionality Curse (1/4)

If we project a query into n-dimensional space, how
many additional MBRs must we examine before we are
guaranteed to find the best match?

For the one dimensional space, the answer is clearly 2
Dimensionality Curse (2/4)

If we project a query into n-dimensional space, how
many additional MBRs must we examine before we are
guaranteed to find the best match?

For the two dimensional case, the answer is 8
Dimensionality Curse (3/4)

If we project a query into n-dimensional space, how
many additional MBRs must we examine before we are
guaranteed to find the best match?

For the three dimensional case, the answer is 26
Dimensionality Curse (4/4)

If we project a query into n-dimensional space, how
many additional MBRs must we examine before we are
guaranteed to find the best match?

More generally, in n-dimensional space we must
examine 3n-1 MBRs; n = 21 → 10,460,353,201 MBRs

This is known as the curse of dimensionality
Spatial Access Methods

We can use Spatial Access Methods like the R-tree to
index our data, but …

The performance of R-trees degrades exponentially with
the number of dimensions. Somewhere above 6-20
dimensions the R-tree degrades to linear scanning

Often we want to index time series with hundreds,
perhaps even thousands of features
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (1/8)

Establish a distance metric from a domain expert

Produce a dimensionality reduction technique that
reduces the dimensionality of the data from n to N,
where N can be efficiently handled by your favorite SAM

Produce a distance measure defined on N dimensional
representation of the data, and prove that it obeys
Dindexspace(A,B) ≤ Dtrue(A,B) (lower bounding lemma)

Plug into an off-the-shelve SAM
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (2/8)

We have 6 objects in 3-D space. We issue a query to
find all objects within 1 unit of the point (-3, 0, -2)
A
3
2.5
2
1.5
C
1
0.5
B
F
0
-0.5
-1
3
2
D
1
0
-1
E
-2
-3
-4
-3
-2
-1
0
1
2
3
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (3/8)

The query successfully finds the object E
A
3
2
C
1
0
B
F
-1
3
2
D
1
0
-1
E
-2
-3
-4
-3
-2
-1
0
1
2
3
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (4/8)

Consider what would happen if we issued the same
query after reducing the dimensionality to 2, assuming
the dimensionality technique obeys the lower bounding
lemma

Informally, it’s OK if objects appear “closer” in the
dimensionality reduced space, than in the true space
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (5/8)

Note that because of the dimensionality reduction,
object F appears to less than one unit from the query
(it is a “false alarm”)
3
2.5
A
2
1.5
C
F
1
0.5
0
B
-0.5
-1
D
E
-4
-3
-2
-1
0
1
2
3
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (6/8)

This is OK so long as it does not happen too much,
since we can always retrieve it, then test it in the true,
3-dimensional space

This would leave us with just E, the correct answer
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (7/8)

Now, let’s consider a dimensionality reduction technique
in which the lower bounding lemma is not satisfied

Informally, some objects appear further apart in the
dimensionality reduced space, than in the true space
3
2.5
A
2
E
1.5
1
C
0.5
0
F
-0.5
B
D
-1
-4
-3
-2
-1
0
1
2
3
GEMINI (GEneric Multimedia INdexIng)
{Christos Faloutsos} (8/8)

Note that because of the dimensionality reduction,
object E appears to be more than one unit from the
query (it is a “false dismissal”)

This is unacceptable because we have failed to find the
true answer set to our query

These examples illustrate why the lower bounding
lemma is so important

Now all we have to do is to find a dimensionality
reduction technique that obeys the lower bounding
lemma, and we can index our time series
Notation for Dimensionality Reduction

For the future discussion of dimensionality reduction we
will assume that:

M is the number of time series in our database

n is the original dimensionality of the data

N is the reduced dimensionality of the data

Cratio = N/n is the compression ratio
An Example of a Dimensionality
Reduction Technique (1/5)
Raw
Data
C
0
20
40
60
80
n = 128
100
120
140
0.4995
0.5264
0.5523
0.5761
0.5973
0.6153
0.6301
0.6420
0.6515
0.6596
0.6672
0.6751
0.6843
0.6954
0.7086
0.7240
0.7412
0.7595
0.7780
0.7956
0.8115
0.8247
0.8345
0.8407
0.8431
0.8423
0.8387
…

The graphic shows a time
series with 128 points

The raw data used to produce
the graphic is also reproduced
as a column of numbers (just
the first 30 or so points are
shown)
An Example of a Dimensionality
Reduction Technique (2/5)

We can decompose the data into 64 pure sine waves
using the Discrete Fourier Transform (just the first few
sine waves are shown in the next slide)

The Fourier Coefficients are reproduced as a column of
numbers (just the first 30 or so coefficients are shown)

Note that at this stage we have not done dimensionality
reduction; we have merely changed the representation
An Example of a Dimensionality
Reduction Technique (3/5)
Raw
Data
C
0
20
40
60
80
100
120
140
..............
0.4995
0.5264
0.5523
0.5761
0.5973
0.6153
0.6301
0.6420
0.6515
0.6596
0.6672
0.6751
0.6843
0.6954
0.7086
0.7240
0.7412
0.7595
0.7780
0.7956
0.8115
0.8247
0.8345
0.8407
0.8431
0.8423
0.8387
…
Fourier
Coefficients
1.5698
1.0485
0.7160
0.8406
0.3709
0.4670
0.2667
0.1928
0.1635
0.1602
0.0992
0.1282
0.1438
0.1416
0.1400
0.1412
0.1530
0.0795
0.1013
0.1150
0.1801
0.1082
0.0812
0.0347
0.0052
0.0017
0.0002
...
An Example of a Dimensionality
Reduction Technique (4/5)

Note that the first few sine waves tend to be the largest
(equivalently, the magnitude of the Fourier coefficients
tends to decrease as you move down the column)

We can therefore truncate most of the small coefficients
with little effect

Instead of taking the first few coefficients, we could take
the “best” coefficients

This can help greatly in terms of approximation quality,
but make indexing hard
An Example of a Dimensionality
Reduction Technique (5/5)
C
C’
0
20
40
60
80
100
We have
discarded 15/16
of the data.
120
140
Raw
Data
0.4995
0.5264
0.5523
0.5761
0.5973
0.6153
0.6301
0.6420
0.6515
0.6596
0.6672
0.6751
0.6843
0.6954
0.7086
0.7240
0.7412
0.7595
0.7780
0.7956
0.8115
0.8247
0.8345
0.8407
0.8431
0.8423
0.8387
…
Fourier
Coefficients
1.5698
1.0485
0.7160
0.8406
0.3709
0.4670
0.2667
0.1928
0.1635
0.1602
0.0992
0.1282
0.1438
0.1416
0.1400
0.1412
0.1530
0.0795
0.1013
0.1150
0.1801
0.1082
0.0812
0.0347
0.0052
0.0017
0.0002
...
Truncated
Fourier
Coefficients
1.5698
1.0485
0.7160
0.8406
0.3709
0.4670
0.2667
0.1928
n = 128
N=8
Cratio = 1/16
0
20
40
60
DFT
80 100 120
0
20
40
60
DWT
80 100 120
0
20
40
60 80 100 120
SVD
0
20
40
60 80 100 120
APCA
0
20 40 60 80 100 120
PAA
0
20 40 60 80 100 120
PLA
Directions for Future Research





Time series in 2, 3, K dimensions
Transforming other problems into time series problems
Weighted distance measures
Relevance feedback
Approximation to SVD