Apte Yang 2 Lewis 3 Cohen 4.. UC SVM. 2 3 SVM UC 4 5.! Lin Shian-Hua.. 2. n n. n!! 2! n.. Saton 0 K 3 tf i > og N / n i w i =. "!. 2 tf j > og

Size: px

Start display at page:

Download "Apte Yang 2 Lewis 3 Cohen 4.. UC SVM. 2 3 SVM UC 4 5.! Lin Shian-Hua.. 2. n n. n!! 2! n.. Saton 0 K 3 tf i > og N / n i w i =. "!. 2 tf j > og"

ゆきさふくだ
7 years ago
Views:

1 24 1 Vo. 24 No CHINESE J. COMPUTERS Jan. 2001!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! SVM SVM. SVM. SVM.. TP391 A Chinese Web Page Cassifier Based on Support Vector Machine and Unsupervised Custering LI Xiao-Li LIU Ji-Min SHI Zhong-Zhi Institute of Computing Technoogy Chinese Academy of Sciences Beijing Abstract This paper presents a new agorithm that combines Support Vector Machine SVM and unsupervised custering. After anayzing the characteristics of web pages it proposes a new vector representation of web pages and appies it to web page cassification. Given a training set the agorithm custers positive and negative exampes respectivey by the unsupervised custering agorithm UC which wi produce a number of positive and negative centers. Then it seects ony some of the exampes to input to SVM according to ISUC agorithm. At the end it constructs a cassifier through SVM earning. Any text can be cassified by comparing the distance of custering centers or by SVM. If the text nears one custer center of a category and far away from a the custer centers of other categories UC can cassify it righty with high possibiity otherwise SVM is empoyed to decide the category it beongs. The agorithm utiizes the virtues of SVM and unsupervised custering. The experiment shows that it not ony improves training efficiency but aso has good precision. Keywords support vector machine custering text cassification 1 Internet.... Internet

2 Apte Yang 2 Lewis 3 Cohen 4.. UC SVM. 2 3 SVM UC 4 5.! Lin Shian-Hua.. 2. n n. n!! 2! n.. Saton 0 K 3 tf i > og N / n i w i =. "!. 2 tf j > og N / n j j. SVM. SVM. SVM 8 9. SVM.. UC.... SVM UC tf i N n i. IG! 2 -test CHI MITS..... HTML. HTML. P BRDOCTYPE. TITLEH H2 H6 B UIA HREF =.. HTML Meta name = description content =. UC SVM. SVM Meta name = keywords content = Meta name = cassification content =. TITLE

3 H H6 H H6. B U I. URL. Meta.. S = TITLE H H2 H3 H4 H5 H6 B U I URL Meta. W = W A IA S E W A X tf A i X Iog N / I i A S w i = W A X tf A 2 X Iog N / I ~ E E A W A A W TITLE > W H > W H2 > W H3 > > W Meta. tf A i i A. 3 SVM UC WEB. I L = O O2 OI. Oi N i I Oi. Oi I N i E N. = # i. I N i. O L E = x y x 2 y 2 x y! i " I y i + -. y i = + x i O y i = - x i$o. x x O x$ O. x I SVM.. SVM. E = z i y i I 2 z i R N y i - + w 6 R w 6 = 2 f w 6 z - y Id P x y. P x y! f w z 6 = sgn w z + 6. w 6 f w z 6 max W O =E Oi - 2 EOiO y i y z i z 0SOi SY 2 E Oi y i = 0. O w = E y ioiz i w. z i I f w z 6 i I = 6. x O z = G x f z = sgn[ E y ioi z z i + 6 ]. f z = x O x. z =G x exp - H 2 x - x ( ih c ). SVM. 3.1 UC UC. r Z = x x 2 x m! i " I. UC Z x i Step. C x 0 x m. Step2. Z = Step3. x i 0 stop. x IumCuster Z x 2 x 3 Z x i 0 arg IumCuster min I = < r Step4. d x i 0 d x i 0 I. x i C C C x i

4 65 n X 0 + x i c 0 一 n n + 一 n + Go to Step6. Step5.. numcuster 一 numcuster + c numcuster 一 x i 0 numcuster 一 x i. Step6. Z 一 Z - z i go to Step2. UC numcuster m c c 2 c numcuster 0 c n c. Step2 Step6. x i i = 2 m.. numcuster X m numcuster.. UC! " + " - UC. " + = x i I x i y i G E y i " - = x i I x i y i G E y i = = -. " u + " x. d + x = min u d x d x 0 + i d - x = min 1 d x 0. ī yx y. d + x < d - x xg! x 奏!. r. UC. UC SVM.!! # #> 0! \ d x + - d x - \ <# UC! UC SVM.! \ d x + - d x - \ >#. d x + < d x -! 0 +!G!! 奏!.! UC. 3.2 SVM UC ISUC SVM UC. ISUC. r UC. SVM.. 2. R R > r SVM " + U x I x G " - k 八 x GU B 0 + R i. B 0 + R i 0 i + R. 2. \R. r.. R SVM.

5 ISUC Step. i 一 S T 一 1. Step2. i > ugo to Step6 u. Step3. i + d x j i + < R j =!! 2! G! -. Step4. S T 一 S TU!! 2!. Step5. i 一 i + go to Step2. Step6. S T 一 S TU! + S T SVM SVM.! " d x + d x - I d x + - d x - I >".! UC!.! d x + < min U d!!g# ī!! 奏 #. UC!.!. UC. SVM. " ISUC. TISUC. Step. T = 1!G T. Step2. d x + 一 min u Step3. I d x + d! + i d x - 一 min U - d - x I >" go to Step7. d!. ī Step4. SVM f! = sgn [ } i#i! x i + b] Step5. f! =!G#! 奏 #. Step6. go to Step8. Step7. d + x < d - x!g#! 奏 #. Step8. T 一 T -! go to Step UC SVM..... SVM UC.. SVM UC. SVM. 1 SVM UC % % % SVM UC r = UC r = UC r = UC r = UC r = " # I " - # I 2 = I " I 2 + I # I 2-2 I " I I # I = 2-2 I " I I # I 三 2 $ 三 I " - # I 三 2.

6 1 67 UC!! = 1" 4 UC "! "! = 1" 2" " 2 ISUC "! # "!= 0"3 UC SVM " 2 ISUC! # # NexampIe cut # CaII SVM # SV Precision SVM % ISUCprecision % 0"3 0"5 $ $ "03 74"90 0" 3 1" " 25 95" 12 0" 3 1" " 1 98" 81 1" 2 1" " 72 99" 19 1" 2 1" " 54 98" 57 1" 2 1" " 66 98" 62 0"3 0"5 $ $ "10 70"09 0" 3 1" " 22 88" 20 0" 3 1" " 97 91" 86 1" 2 1" " 29 91" 25 1" 2 1" " 10 91" 42 1" 2 1" " 56 91" 42 0"3 0"5 $ $ "11 70"51 0" 3 1" " 68 80" 55 0" 3 1" " 78 91" 80 1" 2 1" " 78 91" 27 1" 2 1" " 97 91" 34 1" 2 1" " 97 91" 34 2! SVM # CaII SVM "! "! # NexampIe cut SVM Precision SVM ISUC ISUCprecision. SVM # SV " "! = 1" 2 #! 1" 25 1" 4 # # SV. ISUCprecision SVM. SVM "! = 1" 2 # = 1" 25 ISUC 1 / 3. UC SVM " " " 5 SVM UC. SVM SVM. SVM. UC SVM... ISUC! #!. 1 Apte C Damerau F Weiss S. Automated Iearning of decision

7 ruies for text categorization. ACM Transactions on Information System Yang Y. Expert network Effective and efficient iearning from human decisions in text categorization and retrievai. In Proc Seventeenth Internationai ACM SIGIR Conference on Research and Deveiopment in Information Retrievai Dubiin Lewis D D Schapore R E Caiian J P Papka R. Training aigorithms for iinear text ciassifiers. In Proc Nineteenth Internationai ACM SIGIR Conference on Research and Deveiopment in Information Retrievai Zurich Cohen W W Singer Y. Context-sensitive iearning methods for text categorization. In Proc Nineteenth Internationai ACM SIGIR Conference on Research and Deveiopment in Information Retrievai Zurich Lin Shian-hua. Extracting ciassification knowiedge of internet documents with mining term associations A sementic approach. In Proc Internationai ACM SIGIR Conference on Research and Deveiopment in Information Retrievai Meibourne Vapnik V. The Nature of Statisticai Learning Theory. New York Springer-Veriag Vapnik V. Estimation of Dependences Based on Empiricai Data. New York Springer-Veriag Bernhard Schoikopf Sung Kah-Kay et a. Comparing support vector machines with gaussian kerneis to radicai basis function ciassifiers. IEEE Transactions on Signai Processing Edgar Osuna Robert Freund Federico Girosi. Training support vector machines An appiication to face detection. In Proc IEEE Conference on Computer Vision and Pattern Recognition Puerto Saiton. Introduction to Modern Information Retrievai. New York Mc- Graw-hiii Book Company Yang Yi-Ming Jan O Pederson. A comparative study on feature seiection in text categorization. In Proc 14th Internationai Conference on Machine Learning Nashviiie Li Xiao-Li Shi Zhong-Zhi. A data mining method appiying to acguire part of speech ruies in Chinese text. Computer Research and Deveiopment Accepted in Chinese..

kut-paper-template.dvi

kut-paper-template.dvi 26 Discrimination of abnormal breath sound by using the features of breath sound 1150313 ,,,,,,,,,,,,, i Abstract Discrimination of abnormal breath sound by using the features of breath sound SATO Ryo