åæ師ããå¦ç¿ã®ã¢ãã«ä»®å®
Machine Learning Advent Calendar 2014ã®12æ¥ç®ã
æè¿åæ師ããå¦ç¿ã«èå³ããã£ã¦ã¡ãã£ã¨åå¼·ãã¦ã¿ãã®ã§ããã«ã¤ãã¦æ¸ãã¦ã¿ããèªåãåå¼·ããæã«èªãã æç®ãä¸ã®æ¹ã«æ¸ããã®ã§ããèå³ãããã°ã
åæ師ããå¦ç¿ã¯ã©ãã«ä»ããã¼ã¿ã«å ãã¦ã©ãã«ç¡ããã¼ã¿ã使ã£ã¦å¦ç¿ã§ããã¨ãããã¨ã§ãããé åçãªãã ãã©ãä½ãèããã«ä½¿ãã¨æ師ããå¦ç¿ãã精度ãè½ã¡ããã¨ãããããã
ã©ãã«ç¡ããã¼ã¿ã¯ãã®åã®éãã©ã®ã¯ã©ã¹ã«å±ãããåãã£ã¦ããªããã¼ã¿ãªã®ã§ãä½ãããã® ã¢ãã« ã«åºãã¦ãã®ãã¼ã¿ãã©ã®ã¯ã©ã¹ã«å±ããããä»®å®ãã¦ãããªãã¨ãããªããã¤ã¾ããã¼ã¿ã®åå¸ï¼ã¢ãã«ï¼ã«ä»®å®ãç½®ããªãã¨ãããªããåæ師ããå¦ç¿ãããä¸ã§ã¯ããã ä¸çªéè¦ã
Introduction to Semi-Supervised Learningã®p.15ã«ãããæ¸ãã¦ããã
the model assumption plays an important role in semi-supervised learning.
ã¨ããããã§ããã§ã¯ãããããªåæ師ããå¦ç¿ææ³ã«ã¤ãã¦ãã®ææ³ãä»®å®ãã¦ããã¢ãã«ã«ã¤ãã¦ã¾ã¨ããã
Self-training
æ¦è¦
ããåé¡å¨fã使ã£ã¦ã©ãã«ä»ããã¼ã¿ã®ã¿ã§å¦ç¿ããã©ãã«ãªããã¼ã¿ã®åé¡ãããããã®å¾åé¡ããã©ãã«ãªããã¼ã¿ã®ãã¡é«ã 確信度 ã§åé¡ã§ãããã®ã®ããã¤ããã©ãã«ãããã¼ã¿ã¨ã¿ãªãã¦å度å¦ç¿ãããããã®ç¹°ãè¿ããSelf-trainingã¯wrapper methodãªã®ã§ã©ããªåé¡å¨ã§ã使ããã
Unsupervised word sense disambiguation rivaling supervised methods, ACL 1995
ã¢ãã«ä»®å®
èªåãï¼é«ã確信度ã§ï¼åé¡ãããã¼ã¿ã®åé¡çµæã¯æ£ãããã¨ãããã¯ç°ãªãã¯ã©ã¹ã«å±ããã¼ã¿ãwell-separatedã§ãããã¨ãæå³ããã
Mixture Models (semi-supervised GMM)
æ¦è¦
EMã§æ¨å®ããæã®ãã¼ã¿ã®å¯ä¸çï¼ã©ã®ã¯ã©ã¹ã«å±ããï¼ããã©ãã«ä»ããã¼ã¿ã«ã¤ãã¦ã¯ãã®ãã¼ã¿ãå±ãã¯ã©ã¹ã§ã®ã¿å¤1ãåãããã«åºå®ããã©ãã«ç¡ããã¼ã¿ã«ã¤ãã¦ã¯é常ã®GMMã®æã¨åæ§ã«è¨ç®ãããGMMãä¾ã«åºãããã©æ··åå¤é åå¸ã使ã£ãããã¹ãåé¡ã®è«æãæåã£ã½ãã
Text Classification from Labeled and Unlabeled Documents using EM, Machine Learning 2000
ã¢ãã«ä»®å®
ãã¼ããã¸ã¼ã®ããã«èããããããããªããã©ã ä»®å®ããã¢ãã«ãæ£ãããã¨ããä¾ãã°å®éã®ãã¼ã¿ã¯æ··åæ£è¦åå¸ã«å¾ã£ã¦ãªãã®ã«semi-supervised GMMã§è§£ããã¨ãã¦ããã¾ããããªãã
Co-training
æ¦è¦
ãã¼ã¿ãäºã¤ã®ç´ æ§ã«å解ã§ããã¨ããï¼ä¾ãã°webãã¼ã¸ã®åé¡ãããã¨ãã¯ç»åã¨æç« ã¨ãï¼ãããããã®ç´ æ§ã«ã¤ãã¦åé¡å¨ãåå¥ã«ç¨æããã©ãã«ä»ããã¼ã¿ã®ã¿ã§ããããå¦ç¿ãããåé¡å¨ï¼ã§é«ã確信度ã§åé¡ã§ããã©ãã«ãªããã¼ã¿ãåé¡å¨ï¼ã§ã®æ¬¡åã®å¦ç¿ã«ç¨ããéã«åé¡å¨ï¼ã§é«ã確信度ã§åé¡ã§ããã©ãã«ãªããã¼ã¿ãåé¡å¨ï¼ã§ã®æ¬¡åã®å¦ç¿ã«ç¨ããããããç¹°ãè¿ããSelf-trainingã«ä¼¼ã¦ããã©åé¡å¨ãäºã¤ä½¿ãã¨ããããã¤ã³ãããããWrapper methodã
Combining labeled and unlabeled data with co-training, COLT 1998
ã¢ãã«ä»®å®
- åã ã®åé¡å¨ã ãã§ã è¯ã åé¡ãåºæ¥ããã¨
- å解ãããäºã¤ã®ç´ æ§ã¯ã©ãã«ãä¸ããããä¸ã§ æ¡ä»¶ä»ãç¬ç« ã§ãããã¨
2ãæãç«ããªãã¨ãã©ã¡ãã®åé¡å¨ãåããããªçµæãåºãããæå³ãç¡ãã
Graph-based SSL (Label propagation)
æ¦è¦
Label propagationã¯Graph-based SSLã®ä¸ã¤ãã©ãã«ããã©ãã«ãªãã«é¢ä¿ãªãããã¼ã¿éã®é¡ä¼¼åº¦ã«åºãã¦ã°ã©ããæ§ç¯ããããã¼ãã¯ãã¼ã¿ã§ï¼éã¿ä»ãï¼ã¨ãã¸ã¯ãã¼ã¿éã®é¡ä¼¼åº¦ãé¡ä¼¼åº¦ã®è¨ç®ã«ã¯ã¬ã¦ã·ã¢ã³ã«ã¼ãã«ãç¨ãããã¨ãå¤ããæ§ç¯ããã°ã©ãä¸ã§ã©ãã«ãããã¼ã¿ããã©ãã«ãªããã¼ã¿ã¸ ã©ãã«ãä¼æ¬ ãããã©ãã«ãªããã¼ã¿ã®ã©ãã«ãæ¨å®ãããã©ãã«ãªããã¼ã¿ã®è¿ãã«ã©ãã«ï¼ã®ã©ãã«ãããã¼ã¿ãå¤ãæãã°ãã©ãã«ï¼ã«åé¡ãããã¤ã¡ã¼ã¸ãã¡ãªã¿ã«ããã¯å ã ã°ã©ãã§è¡¨ããããããªãã¼ã¿ï¼ã½ã¼ã·ã£ã«ãããã¯ã¼ã¯ã¨ãï¼ã«å¯¾ãã¦ãããç¨ããããã
Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions, ICML 2003
ã¢ãã«ä»®å®
ï¼å¤§ããéã¿ã§ï¼ æ¥ç¶ãã¦ããï¼ãã¼ãã¯åãã©ãã«ãæã¤ãã¨ã ã©ãã«ãéããã¼ã¿ã§ãé¡ä¼¼åº¦ã大ãããªããã¨ããããããããªå ´åã¯ãã¾ããããªãã
S3VM
æ¦è¦
Semi-supervised support vector machineã®ãã¨ãé常ã®SVMã¯ã©ãã«ãããã¼ã¿ã®ã¿ãè¦ã¦æ±ºå®å¢çã決ããããS3VMã§ã¯ã©ãã«ãªããã¼ã¿ãèæ ®ãã¦ãåºæ¥ãã ã決å®å¢çãã©ãã«ãªããã¼ã¿ã®ä¸ãéããªãããã«ãããå®éã«ä¾ãè¦ã¦ã¿ãã¨ããããããããªãã¡ãªã¿ã«é常ã®SVMã¨éã£ã¦ã³ã¹ãé¢æ°ãå¸ã«ãªããªãããæ±ãã¥ããã
Statistical Learning Theory (æ¬)
ã¢ãã«ä»®å®
ç°ãªãã¯ã©ã¹ã«å±ããã¼ã¿ãwell-separatedã§ãããã¨ãè¨ãæããã¨ãç¹å¾´ç©ºéã«ããã¦æ±ºå®å¢çä»è¿ã«ã¯ï¼ã©ãã«ãªãï¼ãã¼ã¿ãå°ãªããã¨ã
è¯ãæç®
å¤åèªåã¯ãã®é ã§èªãã ã¯ãããã¨ãã¨ããç¨åº¦åæ師ããå¦ç¿ã«ã¤ãã¦ã®ç¥èã¯ãã£ããã©ã
åæ師ããå¦ç¿ã¨ã¯çµå±ã®ã¨ããä½ãªã®ã? - Seeking for my unique color. åæ師ããå¦ç¿ - SlideShare
åæ師ããå¦ç¿ã«é¢ããç°¡åãªä¿¯ç°ãè¯ããå¤åä¸ã®æ¹ã«æããæ¬ãèªãã°ãã®è¨äºã«æ¸ãã¦ãããã¨ã¯å ¨ã¦æ¸ãã¦ãããã ãã©ãã¾ãã¯æ¥æ¬èªã§ä¿¯ç°ããã人ã«ããããã
satomacoto: Pythonã§ã©ãã«ä¼æ¬æ³ã試ãã¦ã¿ã
Label spreadingã£ã¦ããGraph-based SSLã«ã¤ãã¦ã®ç°¡åãªèª¬æã¨ãã®å®è£ ãæ¸ãã¦ãããå®è£ ããããç°¡åã§ãããã
ã°ã©ããç¨ãããã¼ã¢ããã¯é¢æ°ã«ããåæ師ããå¦ç¿
Label propagationã«ã¤ãã¦è©³ç´°ã«èª¬æãã¦ãè¨äºãç°¡åãªä¾ã使ã£ã¦èª¬æãã¦ãã®ã§ãããããããæé©ååé¡ã¨ãã¦ã®å®å¼åã¨ãã«ã¤ãã¦ãæ¸ãã¦ããã
ããã°ãã¼ã¿æ代ã®ãµã³ã¿ç©ã - ML Advent Calendar 2013 æçµæ¥
@sleepy_yoshiããã«ããå»å¹´ã®ML Advent Calendarã®ããã¨ããSemi-supervised GMMï¼ã¨è¿å¹´åé¡ã«ãªã£ã¦ãããµã³ã¿ç©ãï¼ã«ã¤ãã¦æ¸ãã¦ãããå ¨ã¦ã©ãã«ãªãã®å ´åãå ¨ã¦ã©ãã«ããã®å ´åãã©ã£ã¡ãããå ´åï¼åæ師ããï¼ã«ã¤ãã¦åãã¦æ¯è¼ãã¦ãããã¡ããã¡ã詳細ã«èª¬æãã¦ãããããããããããããè¶ ããããã
Node Classification in Social Networks
ã°ã©ãä¸ã®ãã¼ãåé¡ã¨ããåé¡ãããããããªGraph-based SSLã«ã¤ãã¦è§¦ãã¦ããããããã®ç°¡åãªéãã¨ãã©ã³ãã ã¦ã©ã¼ã¯ã¨ãã¦ã®è§£éã¨ãæé©ååé¡ã¨ãã¦ã®è§£éã¨ãæ¸ããã¦ã¦è¯ãã
- ä½è : Xiaojin Zhu,Andrew B. Goldberg
- åºç社/ã¡ã¼ã«ã¼: Morgan and Claypool Publishers
- çºå£²æ¥: 2009/09/15
- ã¡ãã£ã¢: ãã¼ãã¼ããã¯
- è³¼å ¥: 1人 ã¯ãªãã¯: 52å
- ãã®ååãå«ãããã° (9件) ãè¦ã
SSLã®å ¥éæ¸ããã®è¨äºã«ã¾ã¨ãããã¨ã¯å ¨ã¦ãã®æ¬ã«æ¸ãã¦ãããLabel propagationãææ¡ãã人ãæ¸ããæ¬ãåèæç®ãªã¹ãã¨Appendixãé¤ãã°100ãã¼ã¸ããªããèªã¿ãããã®ã§ããèªãããå ¨ä½ã俯ç°ããã人ã«ã¯è¶ ããããã
Semi-Supervised Learning (Adaptive Computation and Machine Learning series)
- ä½è : Olivier Chapelle,Bernhard Schoelkopf,Alexander Zien
- åºç社/ã¡ã¼ã«ã¼: The MIT Press
- çºå£²æ¥: 2010/01/22
- ã¡ãã£ã¢: ãã¼ãã¼ããã¯
- è³¼å ¥: 1人 ã¯ãªãã¯: 4å
- ãã®ååãå«ãããã°ãè¦ã
åãæ¬ã ãã©ãããSSLã®å ¥éæ¸ãIntroduction to SSLãããã£ã¨è©³ããè²ã æ¸ãã¦ãããããããªæç®ããå¼ç¨ããã¦ããããã£ã¡ã®ã»ããå®çªãªã®ããªãèªãã§ãéä¸ãªã®ã§å ¨ä½ã¨ãã¦ããããã§ãããã¯ããããªãã