Time-Sp

T1School

2Scho

Keywords

Abstract: group behaSo this artpiecewise detection tmatch indiabnormal ganalysis.

Introducti

Networkcharacterisgroup infoworld, movimportant marketexisstrong constrading beh

Related W

The keyextraction Time serie

Time seX=<x (vv is v

fourmodesDecompostime seriescoordinateAbnormal

Time se

pace Dom

TingtingLof Compute

ool of Comp

s: Spatio-tem

When a mavior time-sticle we firslinear repre

to get the abividual abnogroups, we

ion

k group bestics of spoormation, pevement rulepart of fina

st the strongsistency of haviors to an

Work

y technologoftime seriees model eries is an v , t ),xvalue for s:1)Symbolisition(SVD)s into multips together wl extractioneries of abno

main of G

i1, Lingyuer Engineer

puter Scienc

mporal data

major event pace domaist to analyzesentation(Pbnormal timormal time can get the

ehavior is ntaneity, ineople can be on “time”ancial mark

g herd behavtrading behnalyze the s

gies of the es ③ The clu

ordered set(v , t )…xt time r

ic[3];2)Disc);4) Points lple child segwhich is defn of time seormal becom

Group BeI

uXu1, JieYring and Sc

ce and Tech

lttsno

a, Group be

creates a bein. The timeze the time PLR) methome domain.

domain, ane space dom

generally anfectious, embe very intu” and “spaceket.XuChuavior[2], and ithavior is formstock plates

Fig.1 time

behavior ofustering of

t which is m(v , t )>

records, ancreteFourierine represengments alonfined as the eries me a new ho

ehavior RInfluence

Yu1,LeiWcience, Shan

hnology, Un

oopy@sina

ehavior, time

ehavior cluse data of eac

series, appod, and thenAccording nd get the cmain of the

around the motional, a

uitive see the”, as showan Wu (200t also meansmed a certa

s and the cyc

e-space dom

f time and time series.

made up w

nd t is rTransformantation(PLR

ng the time atime series

ot spot in th

Researche

Wang1, Yunghai Unive

niversity of S

a.com

e series, ab

ster, this cluch individuaproximatelyn use activito the time

common abe event with

Internet sand temporahe impact own in the Fi05) in his rs the strong

ain effect oncle on the in

main of the e

space are ①.

with record v

strictly ination(DFT);R). Here weaxis, and gemodel.[4,5

he time serie

h Based

nlanXue1

ersity, Shan

South, Huna

normal dete

uster’s spatial on the spa

express thity measureof the majo

bnormal groh k-Medoid

ocial hot iary. Therefof an event gure.1. Theresearchshoconsistency

n stock pricenfluence of

event

① Time se

values and

ncreasing.Us;3) Se focus on det each secti]

es data minin

on Majo

1,YangLiunghai, 20044

an,421001

ection , k-M

io-temporalace is a time he time serie method foor events ocoups. Finallds clustering

issues, and fore, from th

in the netwe stock marows that Chy trading bees. So we caf major even

eriesmodel②

record tim

suallywe uSingular

discussing Pion of the tw

ing. For tim

r Events

u2 44,China

China

Medoids

l data is theseries data.es with the

or abnormalccurred, wey, for theseg algorithm

it has thehe network

work virtualrket [1] is anhina's stockehavior. Thean use thesents.

② Abnormal

me, refers to

use followValue

PLR, cut thewo endpoint

e series, we

s

e . e l e e

m

e k l n k e e

l

o

w e e t

e

International Conference on Education Technology, Management and Humanities Science (ETMHS 2015)

© 2015. The authors - Published by Atlantis Press 1069

don't care about single sequence points of exceptions, but the abnormal change of time series over a period of time. Ma et al., using support vector regression model for training sequence of historical events, when the new coming time series data deviates from the model, it is defined as the new event of the time series. The TSA - Tree improved algorithm achieve the discovery of Singular Value Decomposition and its singular model was defined as a sudden change in the time series.[6] In this paper, on the basis of these methods, we use activity to extract abnormal pattern. The clustering of time series

Time series similarity measure is the method to weigh the similarity degree of two time series. Euclidean distance and Dynamic Time Warping are two classic methods for Time series similarity measure, but these two methods have defects when application to financial time series. Euclidean distance [7] are vulnerable to the interference of volatility of financial time series, and the complexity of the dynamic time warping measure algorithm limits its application scope. In this paper, first time, weuse k-Medoids algorithm to clustering analysis.

Time-Space Domain of Behavior Research

The Basic Definition Definition 1： 、1 Behavior spatial domain:Ω Ω ,Ω , Ω ,… i 1,2,3, … ，Ω is an individual; 、2 Individual behavior time domain:The time range in Ω of a behavior;T =[t ,t ] 、3 Group behavior time domain:The time range in n individuals;

T=[t ,t ]∪[t ,t ]∪[t ,t ]（ ）1 [t ,t ] is the scope of the behavior’s time inΩ Definition 2： Behavior time-space domain:All individuals involved a behavior is as the spatial domainΩ,the

behavior time domain is T. {Ω, T}={(Ω , Τ , Ω , Τ , Ω , Τ , … Ω , Τ （ ）2 Where Ω , Τ is an individualΩ ‘s time domain Τ on the influence of the event Definition3： Activity:It is the degree of an individual affected by the event, in individual time domain

[t ,t ],K is the variation of an individual attribution,Kis the average variation value Activity: =K K⁄ （ ）3

Algorithm Description (1) Determination of individual time domain ofbehavior

In order to discover individual time domain of group behavior,we will analyze the individual time series. First, we should adopt different length of time in the time series for clustering. Second,weuse the PLR(Piecewise Linear Representation) to express the selected time series.

Algorithm 1: Input: A certain individual period of time series X

X=< ( , ), ( , ) ,…, ( , )> Output: Filter the tiny volatility in time series S, Which contains only rise and drop

interval,called ’. 1)scanX, statistics all the lift range and their corresponding amplitude; 2) take all the absolute amplitude values, then calculate the average as a threshold,

the amplitude value is greater than the threshold value is marked as strong, less than the threshold is marked as weak;

3) if (the current range exist small shock range) we should fuse the range according to all kinds of situations

elsewe can copies the current interval length and fuse the before fused oscillation zone, then calculate the size of each lift range}

4)Finally, return the scope of lift range , of ’and related magnitude Algorithm 2:

1070

Input:Outpu

1） Fima

2） If 3） 4） If return

By Alg

And tBy Al

thetimeseri(2) Determ

AlgorInpOut1)F

r2)K

aap

3) Avalue

Experimen

In this pand post ambehavior ostock tradistocks bar posting beh

From trais similar wevery stocksignificant ehavior in aof event im

: the time seut:Abnormairst,scan Xagnitudef ( >0 && return

f ( <0 &&n /gorithm 1, ti

the k is the vlgorithm 2, ies and its r

mination of srithm 3： put: The all tput: Simila

Find out allrangethat goK-Medoids abnormal pattribute vaperiod(pointAccording twhich are g

nts

paper, the exmounts in thf traders, th

ing volume is an imporhavior is drading volumwith annualk’s trading vevents. The

a certain enmpact.

eries ’ maal time interX’, to

byevery> ) / ,> )

, , ime series X

value of inteeach time se

relative magspace doma

abnormal inar individual the stocksot at Algoritcluster ana

period (poinalues of thet),that is to the resultgot by Step2

xperiment adhe post bar. Bhe people tra

shows the rtant index iven by sign

me trend of al trading vovolume trene consistenc

nvironment,

ade from Sterval

calculatethy month, an

,

X=<x ,x ,x

′

kk t

kerval slope.eries’abnormgnitudechanain

ntervals of sals of time ss of public thm2; alysis was

nt), and dive clustering

, ,t of clusteri2 to have si

dopts stockBecause of ading behavdegree of ireflecting trnificant evea single stoc

olume.As shnd is also simcy of each sat the same

Fig.2 T

ep1

he averagnd we use th

，,… x > ct b ,t b ,

……t b ,. mal interval

nges valueis

stocks and teries. abnormal

carried ouvide these g are the K

; ing, then coimilar judgm

k data of finastock tradin

vior is driveinfluence ofraders’ behaents. ck A, as shohown in figumilar. The astock in abne time, these

Trading volum

ge rise hese values

can be exprt ∈ 1, tt ∈ t , t

…t ∈ t , n

l is individus activity.

theirs relativ

time (or po

ut on all thindividual

K value that

ombine withment again.

ancial markng volume in by signifif the stockavior in the

own in figurure 4, fromabnormal trenormal time e stocks also

me trend

magnitudeas the backg

ressed with

ual time dom

ve magnitud

oints) accor

he individutime seriest before an

h the relative

et, includings an importaicant eventsaffected byvirtual wor

re 3, monthlythe various

end in the fireflects the

o constitute b

andkground.

PLR as foll

main of beha

dechanges v

rding to the

uals which s into two nd after the

e changes o

g stock tradant index re

s, and the chy the event. rld as well,

ly trading vos stocks vo

figure is affee consistencbehavior sp

d decline

lows:

avior ,andof

value.

e abnormal

have sameclusters,the

e Abnormal

of abnormal

ding volumeeflecting thehange of the

Post in theand people

olume trendlume trend,ected by thecy of peoplepace domain

e

f

l

e e l

l

e e e e e

d , e e n

1071

In the e

December abnormal pbright red a

The abnorStarting

time) as thbelow.

The spaceThrough

between eabehavior sp

Tracingabnormal i3,weuse linbetween th

With k class,asshovalueof thenot affecte

Stock

Jiugui -0.

Moutai

-0.

Shanxi

Fenjiu

-0.

experiment, 2012,Firstl

periods at tarea is the a

rmal periodg from each he backgrou

-time modeh the previoach stock.Apatial doma

g to the plainterval of ane segment

he two interv

- Medoids own in figure abnormal d by the eve

Stock Jiugui Moutai WuliangyeYanghe Shanxi FeSwellfun

.150189855

.216869755

.054322847

we analyzely,we analythe same timabnormal ar

Fig.3 Std of stocks.stock (indiv

und, extract

el of groupous analysis

AS shown inain. asticizer evall the stockt to tradingvals of perio

algorithm tre 4. Plasticinterval andent, as show

Abn[11/11/1

e [11/[11/

enjiu [11/[11/

0.494768533

0.432170976

0.477183145

e 26 stocksyze each stome constiturea of each s

ock trading. vidual), witht the unusua

Table1Abn

behaviors of abnormn figure 5, w

ents in Noks in Novemg volume tiods before a

to clusteringcizerevents d the averagwn in figure

normaltime p19,11/29]

19,12/11,12/216,11/19],[19,11/12],1116,11/19], [115,11/19],12

3 -0.4042747

7 -0.2300143

4 -0.5595898

’ trading voock abnorm

ute a behavistock.

g volume of

h historical al period of

normal time perio

mal time intwe also nee

ovember 19mber 19, 20ime series oand after.

Table2 K va

g analysis timpact on

ge K value, e 5.

period

24 2/3,12/5],[12

1/19,12/24,1211/30, 12/5],2/10,[12/24,1

Stoc

479 Wuliae

486 Yang

932 Swell

olume data mal periodsior spatial d

f 26 kinds w

trading dataf stocks by

ods of each stock

terval, we ped to use cu

, 2012 as 012, and theof each sto

alues

to the K valthe stocks cand when t

2/13,12/17]2/27 12/14,[12/26

12/27]

ck

angy -0.181

ghe -0.172

lfun -0.349

of wine ind, and these

domain. As

wine industry

a (that isthe the above

preliminarilyurvesimilarit

the source,en through tock. Table 2

lue, We cancan be meathe ratio is l

6, 12/27]

184553 0.2

267975 0.3

9288851

0.6

dustry in Ne stocks wh

shown in f

y

backgroundalgorithm 2

y establish tyalgorithm

, we can fithe previou2 shows Th

n divide theasured by raless than 1,

266665426

-

300421043

-

613708822

-

ovember tohich appearfigure3,The

d domain of2.As shown

the contactm for related

ind out thes algorithmhe K value

e stock to 2atio ofthe Kthe stock is

-0.327298877

-0.155979363

-0.259637235

o r e

f n

t d

e m e

2 K s

7

6

1072

Finally

domain of analyze postocksaffecamountsch

Conclusio

ThroughNovemberincludingJiThese expbehavior infield.

Acknowle

This woTechnolog(No.40976

Reference

[1] MuGuoH

exchang[2]WuXuCh

we get theplasticizer

ost amountcted by the hanges have

n

h the abov19in2012,thiuguijiu,Laoeriments illn virtual wo

dgement

ork is suppgy (No.2016108)

es

Hua, ZhouWges on an em

huan, HePe

Fig.4 si

Fig.5T

stocks affeevent. In or

ts in relateevent. Fro

e strong con

e experimeheeventin obaigan,Tonlustrate notorld is affec

ported by T1105033)

WeiXing, Chmerge in sto

eng. China

imilar curve clust

The influenc

ected by thrder to verid post bar

om the expesistency dur

Fig.6 post v

ent, we get liquor-

ngpu,Swellft only the tted by the e

he Ocean Pand The

henWei, Keck market.N

open-end

November

ter

ce degree of

he plasticizefy the expe

r in this peerimental re

uring the pla

volume and

.

t a time-spa-making fun,Wuliantrading behevent, theirs

Public WelNational

ertesz JanosNewJoumal

fund herd

r 19, 2012

f plasticizer

er event, theeriment resueriod, we esults, the tasticizer eve

dTrading vo

atial domaifield

ngye,Shanxiavior in thes influence

fare ProjecNatural S

. Order flowlofPhysies,2

behavior a

No

r

ey constitutults, we adoalso obtaintrading voluent, as show

olume

in of the paffects

Fenjiu,Golde real worlalso has the

t of the MiScience Fo

w dynamics 2010,12:075

analysis[J].

ovember 19, 2012

te the behaopt the samen the sameume change

wn in figure

plasticizer es the den seed ald,but also e strong con

inistry of Soundation

s around ext503

Journal o

2

vior spatiale method to

individuales and post6

event, sincestock

and so on,the posting

nsistency of

Science andof China.

treme Price

f Financial

l o l t

e k ,

g f

d .

l

1073

Research, 2005 (5): 60-69.

[3]LIU Yi, BAO De-pei, YANG Ze-hong, ZHAO Yan-nan.Research of New Similarity Measure Method on Time Series Data[J].Application Research of Computers, 2007, 24(5): 112-114. [4] Li Guo’ai, CaoZheng. On-Line Segmentation of Time-Series Data [J].Journal of Software, 2004,

(11). [5]LiuliJiaPengtaohua-can he, etc.Overview of time series data mining [J]. Application Research of

Computers, 2007, (11):15-18. [6] Hung N Q V, Anh D T. Combining SAX and Piecewise Linear Approximation to improve

similarity search on financial time series[C]//Information Technology Convergence, 2007. ISITC 2007.International Symposium on. IEEE, 2007: 58-62.

[7]LUO Xin OUYANG Yuan-Xin XIONG Zhang YUAN Man.The Effect of Similarity Support in K-Nearest-Neighborhood Based Collaborative Filtering[J].Chinese Journal of Computers, 2010.

1074