Advanced Database Systems

Advanced Database Systems
1-D Time Series Data Indexing
Donghyun Jeong
Contents
n
n
n
n
n
n
n
1-D Time Series Data

Generate Data
Overall Processing Method
Preprocessing (I),(II),(III)
Searching (I), (II)
Implementation (I), (II)
Demo & Reference
Ill briefly introduce about 1-D time series data indexing. In fact, I
cannot get currently used data such as stock data. Therefore I
generate artificial data to process and show how to do indexing.
Ill have a look all processes of designed and implemented
procedures. (More brief information can find in the book;
ADVANCED DATABASE SYSTEMS)
1-D Time Series Data

n
Stock Data(example)
ABC NEWS
70
60
Stock
50
40
30
20
10
M9
M1
0
M1
1
M1
2
M8
M7
M6
M5
M4
M3
M2
M1
Month
An example of 1-D time series data is stock data. I have

designed indexing procedures with computer-generated stock
data instead of real data. As you see the histogram on the above,
it denotes time series data.
Sample Data (a part)
M1
M2

M3
NAME
M4
M5
57.709375 59.903125 52.584375

News
45.6375
38.69375
69.34375
62.35625
58.6125
67.8875
75.690625
60.2
61.340625
. .. MBS TV
69.859375 77.396875 73.921875

DB
ABC
Hallym
Generate Data
n
1-D Time Series Data Generating
As I said on the previous page, time series data(Stock data)

generated on special program we made. The generating program
made with MFC. Stock data has fluctuations about 10. I
generate 100 samples include modeled company names to show
it looks like real data. (See on the previous page)
Overall Processing Method

n
Preprocessing
n
1-D Time Series Data -> Spatial Data

n
Using DFT, Clustering(Iterative Method)
Indexing (R-tree like method)
Searching
n
n
1-D Time Series Data -> Spatial Data

Search nearest data
n
Using Euclidean Distance
Refining Process
1-D Time Series Data Indexing has two parts of processing

procedures.(Preprocessing/ Searching)
1, Preprocessing
First change 1-D Time Series Data to spatial data on the
frequency domain using DFT(Discrete Fourier Transform).
Second grouping with similar data using clustering method. Last
indexing grouped ID(Index) to get a possibility of searching.
2. Searching
After preprocessing procedure, we can search data which we
want. To find data we need to change our data to spatial data
using DFT as we used in preprocessing. After get new spatial
data which we want to search, calculate Euclidean distance with
original spatial data. And then choose a nearest one. In result,
we do refining process to get final data.
? see more on next page
Preprocessing (I)
n
DFT(Discrete Fourier Transform)
No fault dismissals(Parsevals theorem )

Dfeature(F(x),F(y)) <= D(x,y)
Using DFT, we can change 1-D time series data to spatial data
on frequency domain. After change to spatial data, choose three
data (f = 1 ~ 3). If the wanted data is included(false alarm), we
can find data with the refining process. But if the wanted data is
not included(false dismissals), we cannot find data. The reason
of changing data to spatial data is that the Euclidean distance of
original data is always greater than the Euclidean distance of
spatial data which is changed to spatial data using DFT. It
denotes no fault dismissals. [2][3]
DFT Pseudo Code
for k=0 to sizeofdata -1
real_tmp = 0
// real data
imag_tmp = 0
// image data
for n=0 to sizeofdata -1

real_tmp = real_tmp + data(n)*cos(2*PI/12*k*n)
imag_tmp = imag_ tmp + -1*data(n)*sin(2*PI/12*k*n)
next
power(k) = sqr (real_tmp*real_ tmp + imag_tmp*imag_ tmp)
// calculate power of data on frequency domain
next
Preprocessing (II)
n
Clustering
n
Calculate Cluster Center using K-means

n
K-means is a method to search nearest

point using Squared Euclidean Distance.
Grouping some elements which have a

same center value.
I use the iterative method to cluster data. Especially I used Kmeans algorithm to cluster. K-means algorithm is to make K
numbers of cluster center using squared Euclidean distance. [4]
----- GROUP LIST(Clustering) ----Cluster Number + {Elements}
0{0}
1 { 28 44 87 }
2 { 19 30 67 84 }
3 { 3 32 70 }
4 { 4 39 81 91 95 }
5 { 5 37 69 83 86 }
6 { 6 46 48 62 80 93 }
7 { 7 41 71 74 }
8 { 8 36 40 63 }
9 { 9 60 61 78 88 89 90 97 }
10 { 10 31 54 }
11 { 11 72 73 92 }
12 { 12 96 }
13 { 13 42 49 59 85 }
14 { 14 35 47 77 }
15 { 15 26 34 43 55 57 58 65 82 }
16 { 16 50 64 75 }
17 { 17 68 94 98 }
18 { 18 38 51 56 }
19 { 25 53 66 }
Preprocessing (II)
n
Clustering
3-D Space(without Z value)
80
70
DFT2(Y)
60
50
40
30
20
10
0
0
20
40
60
DFT1(X)
80
100
120
After clustering data as you can see on the above, we can

search with cluster center values. The total data(100) changed to
25 cluster center.
----- Cluster Center ---DFT1
DFT2
DFT3
23.340117 66.733917 10.157041

50.151770 25.601297 4.004073
61.989542 15.195794 22.199620
7.714025 28.789136 24.238079
40.608878 13.905752 7.586689
67.583971 34.087372 15.156011
45.037389 37.401815 18.088304
12.462210 29.154218 6.057019
43.047011 20.739349 31.244777
16.180685 15.162112 11.154085
8.316712 9.745249 3.354249
37.904022 31.389206 11.713221
26.594263 6.602002 19.067617
26.666832 18.493810 13.540259
16.295715 23.074414 19.648849
93.036619 44.201457 28.903012
30.942753 10.509283 10.143203
38.378123 18.404003 17.045466
25.086263 30.154396 19.368506
83.322350 16.709450 20.495816
Searching (I)
n
Search Data
As people input data on the web, the preprocessing procedure

change data to spatial data using DFT. With this spatial data, we
choose the nearest cluster. It denotes false alarm, we can find a
result data after using refining process.
Searching (II)
n
Search Data
Similary
Search
Result
Refining
Result
As you see on the above, we can find similar search results and
refining result. On the refining process, we determined e value
must be smaller than 1.0 (e<1.0). Otherwise the exact data which
we want to search cannot be found.
Serch
Found Similary Search Result
Company Name : GVE
Company Name : Medal co.
Company Name : YAYAWA
Refining Result
Company Name : Medal co.
10
Implementation (I)
n
Random Generate & insert data

n
User Interface
n
Visual C++ 6.0

ASP, Msoffice2000 component(Web
Application)
Database
n
SQL Server 7.0
We generate 1-D time series data using random generator in

Visual C++. Also we designed user interface using ASP(Active
Server Page) to search data on the web. 1-D time series data
shows graphical view using graph components in MSOffice200.
In fact we should make database system to test data indexing.
Instead of making database system, we use well-known
database system(SQLServer 7.0).
11
Implementation (II)
n
Datagram(table descript)
Cluster
values
Data &
DFT
values
Table consists of cluster data and spatial data. Cluster data and
ID are connected with foreign key with GRP(cluster group
number).
CREATE TABLE [dbo].[timeseries ] (
[ID] [float] NULL ,
[M1] [float] NULL , [M2] [float] NULL , [M3]
[float] NULL , [M4] [float] NULL , [M5] [float] NULL , [M6] [float] NULL , [M7]
[float] NULL , [M8] [float] NULL , [M9] [float] NULL , [M10] [float] NULL , [M11]
[float] NULL , [M12] [float] NULL , [DFT1] [float] NULL , [DFT2] [float] NULL ,
[DFT3] [float] NULL , [GRP] [int] NULL , [NAME] [nvarchar] (255) NULL
) ON [PRIMARY]
GO
CREATE TABLE [dbo].[cluster] (
[ID] [int] NOT NULL , [C1] [float] NULL , [C2] [float] NULL , [C3] [float]
NULL ) ON [PRIMARY]
GO
ALTER TABLE [dbo].[cluster] WITH NOCHECK ADD
CONSTRAINT [PK_cluster] PRIMARY KEY NONCLUSTERED
( [ID] ) ON [PRIMARY]
GO
ALTER TABLE [dbo].[timeseries ] ADD
CONSTRAINT [FK_timeseries_cluster] FOREIGN KEY
( [GRP] ) REFERENCES [dbo].[cluster] ( [ID] )
GO
12
Demo
The image shows web designed 1-D time series data.
MAIN PAGE (overall composition)

|
+ Search by name
|
+ Search by data
| |
| + DFT & Similar Search & Refining
|
+ Graph Form Display (graphical display)
13
Result
n
A possibility of 1-D Time Series Data

Searching on the Internet.
Dont know how fast it is compare
with another applications.
Should use real data instead of
generated data
We designed and implemented 1-D time series data searching

on the web. Actually we has a look a possibility of searching
stock data on the web in relevantly short time. In fact there is no
1-D times series data searching product on the web.
The problem is that searching 1-D time series data searching on
the web is possible but there is no analysis method to measure
the performance of 1-D time series data searching.
14
Reference
n
[1]ADVANCED DATABASE SYSTEMS,Carlo Zaniolo et al.

Morgan Kaufmann Publishers, pp.295-305, 1997.
[2]Fourier Transform of an image,
http://www.postech.ac.kr/~yirin/fft/fft.html
[3]C++ ALGORITHMS for DIGITAL SIGNAL PROCESSING
Second Edition, Paul M. Embree, Damon Danieli, pp.331339, 1998.
[4]Pattern Recognition with Neural Networks in C++, Abhjit
S. Pandya, Robert B. Macy, pp.213-230, 1995.
15

Advanced Database Systems

Uploaded by

Copyright:

Available Formats

Advanced Database Systems

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Database Systems

Uploaded by

Copyright:

Available Formats

Advanced Database Systems

1-D Time Series Data Indexing

1-D Time Series Data

1-D Time Series Data

An example of 1-D time series data is stock data. I have

57.709375 59.903125 52.584375

69.859375 77.396875 73.921875

1-D Time Series Data Generating

As I said on the previous page, time series data(Stock data)

Overall Processing Method

1-D Time Series Data -> Spatial Data

Using DFT, Clustering(Iterative Method)

Indexing (R-tree like method)

1-D Time Series Data -> Spatial Data

Using Euclidean Distance

1-D Time Series Data Indexing has two parts of processing

DFT(Discrete Fourier Transform)

No fault dismissals(Parsevals theorem )

for n=0 to sizeofdata -1

Calculate Cluster Center using K-means

K-means is a method to search nearest

Grouping some elements which have a

After clustering data as you can see on the above, we can

23.340117 66.733917 10.157041

As people input data on the web, the preprocessing procedure

Random Generate & insert data

Visual C++ 6.0

SQL Server 7.0

We generate 1-D time series data using random generator in

The image shows web designed 1-D time series data.

MAIN PAGE (overall composition)

A possibility of 1-D Time Series Data

We designed and implemented 1-D time series data searching

[1]ADVANCED DATABASE SYSTEMS,Carlo Zaniolo et al.

You might also like