EM algorithm (impute missing value)

mentari_mrh

Minta bantuan....(x10)

Ader sape2 yg tau pasal Em algorithm...best lagik klu ader pengetahuan ttg cara nak program kan method tu...

Thanksss...:love:

[ Last edited by mentari_mrh on 14-6-2005 at 01:26 PM ]

Sarah_Radzi

apa yg u nak tahu pasal EM algo ni ?

cara nak program dia bergantung pada probability distribution u lah ..

programnye iteratif, tak susah sangat nak diprogram, kalau u ade probability expression yg betul

seperti yg u patut tau, EM ade dua step

step E : kira Q(s|s') = E[ (log f(x|s))|y,s']

Step M : cari solution s yg maximize Q(s|s') ==> S=argmax(Q(s|s'), s'=S, repeat E kalau still tak converge

x = complete data, y = incomplete data , s = statistik parameter yg describe x

so, part mana yg u tak paham

[ Last edited by Sarah_Radzi on 16-2-2005 at 01:05 PM ]

Sarah_Radzi

btw, u nak estimate data ape  ? dlm bidang ape ?

mail golok

Originally posted by Sarah_Radzi at 2005-2-16 07:03 PM:
apa yg u nak tahu pasal EM algo ni ?

cara nak program dia bergantung pada probability distribution u lah ..

programnye iteratif, tak susah sangat nak diprogram, kalau u ade probability expres ...

apo kobondo EM algo nih sarah..

citer sikit...sikit pon jadik lerr..:tq:

Atomic_Omnikid

EM algoritma apa benda EM stands for what...

mentari_mrh

tq...SARAH

but actually secara teori nyer mmg dah faham tp bila nak apply tu rasa tak tau lak...

baru2 ni ader baca artikel ttg EM nih...dr Dhowell...dia ader bg step2nyer...so try lah wat dlm excel je....

ini yg aku faham lah yek...(summarized from dr dhowel paper)


Step 1. Find the sufficient statistics of the observed data - guna listwise deletion untuk cari the first estimates, @t (xtau nak wat symbol)

(so kat sini aku regress data yg incomplete to get the parameter estimatation)

Step 2. pastu guna @t to solve for regression coefficients

Step 3. Estimate missing data using regression model developed by using @t

Step 4.pastu prediction data kau ganti kat mana2 yg ader missing data, & the estimates and observed data are now a 搉ew complete data

mentari_mrh

Expectation Maximization (EM) algorithm...

nie sebenarnyer method utk Missing Value Problem (MVP)...method nie menggunakan Maximum Likelihood Estimator (MLE) utk cari dia nyer unbiased estimator bila data kita ilang atau tak lengkap
secara teori dia:

i. Replace missing values by estimated values
ii. Estimate parameter
iii. Repeat
    Step (i) using estimated parameter values as true values, and
    Step (ii) using estimated values as observed values, iterating until
    covergence

EM algorithm leh guna utk solve mcm2 benda cam:
 Filling in missing data in a sample
 Discovering the value of latent variables
 Estimating parameters of finite mixtures
 Unsupervised learning of clusters

aku ni pun bdk baru dlm EM nih...tu yg tgh explore nih...
harap2 leh ler kongsi ngan korang...mklumlah two heads is better than one

Sarah_Radzi

Originally posted by mentari_mrh at Thu 17-Feb-2005 03:35:
tq...SARAH

but actually secara teori nyer mmg dah faham tp bila nak apply tu rasa tak tau lak...

baru2 ni ader baca artikel ttg EM nih...dr Dhowell...dia ader bg step2nyer...so try lah wat dl ...

sekali imbas nampak macam betul ..

biasanya i iterate (radiomobile channel estimation) dlm 3-10 iteration je ..
kalau ikut theory, data convergence rate tu depends kepada Fishers Information ..
lagi informative, lagi slow convergence ..

EM ni gerenti converge ke local stationary point .. cuma persoalannya . adakah stationary point tu estimate yg terbaik ..

EM depends kat initial value tu .. so kena try dapatkan the best estimate lah (walaupun jauh dari sebenar)

kalau u punya program tu converge, utk test, u create data dari model yg u dapat dari EM tu lah .. kalau distribution dia macam betul, ok lah tu . walaupun 3 iteration, hehehe..  

kalau u nak paper original EM , (Dempster, Rubin, Laird, 1976/7) tu, i rasa i still simpan (ataupun i tahu nak dapat kat mana), tapi memang mathematik abis lah kalau nak compare kpd website dhowell tu (yg i baru discover)

flowerjun

Thanks...Sarah

pasal paper rubin tu aku dah baca..tapi susah skit ler nak paham teori nih...bila rasa nak paham...tetiba lain lak...confiusssss   :stp: tu yg aku try cari paper yg apply skit tu...emmm...lagik satu pasai convergence mengikut fishers tu leh terangkan tak :love: tak berapa tau lah....skang ni tgh cuba2 lagi surf artikel2 lain....TQ

Sarah_Radzi

yg pasal fishers tu nanti sat .. kena selongkar kotak cari buku EM i ..  i pun dah tak ingat  camne nak relate.. hik hik hik ..

ps: kalau memory i betul lagi, dlm paper Fessler, Hero pasal Space Alternating Generalized EM (IEEE Trans Comm ataupun Trans Sig Processing thn 80an, 86 kot, ade gak citer sket)

Sarah_Radzi

da, i think u can digest this ..

mentari_mrh

:hmm: dah try wat...iteration stop at 3 to 4...pastu terus converge...

aku try guna SPSS dulu...klu dpt jwpn ok baru try buat dier nyer syntax...

tp nampak cam :gila:...pening kepala dah...anyway thanks a lot sarah...rasanyer cam nak dapat dah...nanti klu dah proceed aku bg tau kejayaan nyer yej...kita kongsi sama2....:pompom:.....kekekkekke

mbhcsf

ni menatang apa Sarah...? nak digest pun raser well...ini apa ker benda ni?

Sarah_Radzi

Originally posted by mbhcsf at Fri 11-Mar-2005 15:11:
ni menatang apa Sarah...? nak digest pun raser well...ini apa ker benda ni?

ini statistik .. dalam kategori matematik (remember Sci, Tech & Math)
utk estimate missing data

lamded

ni ambik degree in mathematic ke?? majoring math?? jom bukan club matematik.. hahaha. me mathematic computer industry skrg tgh pk nak keja kat mane.

mentari_mrh

okey gak tu...aku pun rasa nak gak tau apa benda2 baru yg leh diapplykan dlm statistik...utk future research...nak cari keje yek...tgklah klu ader pleuang terbuka aku bg tau...aku baru je abih master...nak konvo dah bulan april nih...yea yea yea sonok(x100)....abih dah zaman blajar...yuhaaaaaa.....jom kita buka club matematik&statistik....

mentari_mrh

aku pure statictics...tapi masa master tu amik financial engineering..(tapi hehehehe statistics tetap dihati) tu yg keje pun bidang research....harap2...ko dpt ler kongsi ngan aku ilmu maths ko...sbb aku ni bhgn teori nih...malas skit..kekekekke....

mentari_mrh

aku dah try buat regression cam biasa...iteration dia 2 kali terus converge...buka balik artikel dempster nad rubin...baca lagi...buka lak spss nyer artikel pasai em...ader jumpa satu ayat...suruh guna special formula refer to rubin...tgk2 ader...perghhh...pening kepala...skang ni tgh try apply formula tu...agak2 leh dapat kot...hehehehhe

lamded

aku dulu statistik juga..... tapi kantoi so aku swicth ke math computer, less statistik lah major aku ni. ermm finance engineering??? can u give me more detail about that. sound interesting... i want to continue my study.. ambik master. but still wondering nak masuk  math balik or master in science and technology management in um. ermm trying to get into intel doing failure analysis job still waiting for 2nd answer.

mentari_mrh

penat aku type td...ilang lak...kena type balik...emmmppp