Scribd
Upload
100 views

Similarity Search On Time Series Data

This document discusses techniques for performing similarity search on time series data. It is difficult to index and search time series data using exact matching due to the need to use distance functions to compare sequences. The document explores using the Discrete Fourier Transform (DFT) and Discrete Wavelet Transform (Harr wavelet) to extract features from time series for indexing and search. It evaluates these techniques on stock market and other data, finding they perform similarly in concentrating data energy in the first few coefficients. Later techniques discussed aim to optimize search for subsequence and variable length queries.

Uploaded by

Mehran Rajabi
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
100 views

Similarity Search On Time Series Data

This document discusses techniques for performing similarity search on time series data. It is difficult to index and search time series data using exact matching due to the need to use distance functions to compare sequences. The document explores using the Discrete Fourier Transform (DFT) and Discrete Wavelet Transform (Harr wavelet) to extract features from time series for indexing and search. It evaluates these techniques on stock market and other data, finding they perform similarly in concentrating data energy in the first few coefficients. Later techniques discussed aim to optimize search for subsequence and variable length queries.

Uploaded by

Mehran Rajabi
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

Similarity Search on

Time Series Data


Presented by Zhe Wang

Motivations
z

Fast searching for time-series of real


numbers. (data mining)
z

Scientific database: weather, geological,


astrophysics, etc.
find past days in which solar wind showed similar
pattern to todays

Financial, marketing time series:


Find past sales patterns that resemble last month

Real motivation?

Difficulties for time-series data


z

Cant use exact match like fast string match:


z

Need to use distance function to compare two


time series (next slide)

Cant easily index the time-series data


directly.
z

Need faster algorithm than linear scan (whole


talk)

Distance functions
z

L-p distance function


D(x,y) = ( |xi yi|p )1/p

L-2 distance function (Most popular)


D(x,y) = ( (xi - yi)2 )1/2

Finding similar signals to query signal q


means finding all x such that:
D(q,x) = ( (qi - xi)2 )1/2 <=

Why prefer L2 distance


z

Important feature:
L2 distance is preserved under orthonormal
transforms (For L-p norm, only p=2 satisfy this
property)
Orthonormal transforms: K-L transform, DFT, DWT

Optimal distance measure for estimation


z

If signals are corrupted by Gaussian, additive


noise

Widely used

How to index time-series data


z

Can not direct index the data


z

Very big dimensionality (Even if query is just 512


points)

Need to extract fewer important


representative features to build index upon.
Try to use first few parameters of DFT
(Discrete Fourier Transform) to build index.

DFT: Discrete Fourier transform

Figures taken from: A comparison of DFT and DWT based similarity search in Timeseries Databases (Also figures on slide 9,17,18,24,25)

DFT definition
z

n-point DFT:
(Xf is frequency domain, xt is time domain)
Xf = (1/n1/2) * t=0 to n-1 xt exp(-j2ft/n)2 f = 0,1,, n-1

Inverse DFT:
xt = (1/n1/2) * f=0 to n-1 Xf exp(j2ft/n)2 t = 0,1,, n-1

Energy E(x):
E(x) = ||x||2 = |xt|2

FFT can be done in O(nlogn) time

Parsevals theorem
z

Let X be the DFT of sequence x:


|xt|2 = |Xf|2
Since DFT is a linear transformation:
=> || xt yt ||2 = || Xf Yf ||2
L2 distance of two signal in time domain is same as
their L2 distance in frequency domain

No false dismissal if we just use first few


parameters.
But also do not want too many false hits

Different time series data


Energy distribution
Example
in O(fb)
O(f0)
Totally independent
White noise
time series
Pink noise
O(f -1)
Musical score, work
of art
O(f -2)
Stock movement,
Brown noise
(Brownian walks)
exchange rates
Black noise
O(f -b) b > 2
Water level of river
vs time
Type

Building Index
z

When the signal is not white noise, we can


use first few DFT parameter to capture most
of the energy of the signal
Let Q to be the query time-series data:
first_few_freq (qf - xf )2
<= all_freq (qf - xf |)2
= (qt - xt)2
<= 2

Building index (cont)


z

Use the first few (4-6) DFT parameters, use


R*-tree as index (called F-index)
Given a query Q and , use the index to filter
out all nodes where:
first_few_freq (qf - xf )2 > 2

Can we do better?
z
z

Use DWT (Discrete Wavelet Transform)


Harr wavelet definition:

Where

ij (x) = (2jx i)
i = 0, , 2j-1
1
0 < t < 0.5
(t) =
-1
0.5 < t < 1
0
elsewhere

Harr transform example


z

z
z

Time series data: f(t) = (9 7 3 5)


Resolution

Average

Coefficients

(9 7 3 5)

(8 4)

(1 -1)

(6)

(2)

Harr transform result: (6 2 1 -1)


If we take only first two coefficients (6 2) and
transform back, we get: (8 8 4 4)

Use Harr wavelet with real data

Harr wavelet vs DFT

Preserve L2
distance
Feature
Computation
time
Energy
concentration for
first few params

Harr wavelet

DFT

Yes

Yes

Can capture
localized feature
O(n)

Only global
feature
O(nlogn)

Low resolution

Low frequency

Performance comparison
z

10k feature vectors from HK stock market


using sliding window size =512
Precision = Stime / Stransform
z
z

Stime: # of sequences qualified in time domain


Stransform: # of sequences qualified in transformed
domain

Compare precision using different amount of


coefficients with different method

Performance (HK stock)

DFT fights back


z

Use last few DFT coefficient to improve


quality (Davood Rafiei)
The DFT coefficients of a real valued sequence of
duration n satisfy:
Xn-f = Xf* (f = 1, , n-1)
Note: if X = a+bi, X* = a-bi

Energy distribution: DFT vs DWT

Performance (100 stocks)

What is next: Subsequence


query
z

Up to now, we are focused on whole time


series data match.
What if we need to match subsequence
efficiently?

Query:
Data:

Nave method
z
z

Assume query length fixed at


Using a sliding window with length , slide
through the data.
Insert all possible data points into the index
(using F-index)
Could be twice as slow as sequential scan

ST-index
z

Observation: Successive sliding window tend


to generate similar coefficients
MBR: Minimum Bounding (hyper) Rectangle

How about arbitrary length


time series query?
z

Two basic methods: (Assume dataset is


indexed with window length )
z

Prefix search
Simply use the first of the query to do the search

Multipiece search
If |q| >= k, split q into k pieces, and search DB
with /(k1/2), join the results.

Multi-resolution index
z
z

Base-2 MR index structure


Take as base unit, build index for each
window size of 2i.
Basic algorithm: Longest Prefix Search (LPS)
Eg: if query length = 19
use 16 as the prefix to do search.

Index structure layout

Figure taken from: Optimizing similarity search for arbitrary length time series queries
(also figures on three slides)

Improved algorithm
z
z

Split q into q1q2qt, where |qi|= 2Ci


Eg: Given |q| = 208 = 16 + 32 + 128
q1

q2

q3

16

32

128

Given = 0.6, query example shown in the


next slide

Performance on 556 stocks (again)

Summary
z

Harr DWT and DFT performs similar in feature


extraction for stock data.
Start monitor stock now! (All of these papers
use stock data + synthetic data)
Time series data is hard to optimize for
similarity search.
All these paper are focused on no false
dismissal, approximation might help. (Some
research done.)

Related papers
z
z
z
z
z
z

Tamer Kahveci and Ambuj K. Singh. Optimizing Similarity Search


for Arbitrary Length Time Series Queries
R. Agrawal, C. Faloutsos, and A. Swami, Efficient Similarity
Search in Sequence Databases.
K.-P. Chan and A.W.-C. Fu, Efficient Time Series Matching by
Wavelets.
C Faloutsos, M Ranganathan, Y Manolopoulos, Fast
subsequence matching in time-series databases
D. Rafiei and A. Mendelzon, Efficient Retrieval of Similar Time
Sequences using DFT
YL Wu, D Agrawal, A Abbadi A comparison of DFT and DWT
based similarity search in Time-series Databases

Orthonormal transform
z

Matrix is known as orthonormal if it satisfy


the orthonormality property:

From: http://www.math.iitb.ac.in/~suneel/final_report/node15.html

You might also like