2018å¹´12æ5æ¥ ãªã¯ã«ã¼ãã¹ã¿ããã£ã³ã°ã®ã¤ãã³ãã§ã®è³æã§ãã ãæ©æ¢°å¦ç¿ã®ã¨ãã»ã³ã¹ãã®è§£èª¬ãã¡ã¤ã³ã«ãªã£ã¦ãã¾ãã
2018å¹´12æ5æ¥ ãªã¯ã«ã¼ãã¹ã¿ããã£ã³ã°ã®ã¤ãã³ãã§ã®è³æã§ãã ãæ©æ¢°å¦ç¿ã®ã¨ãã»ã³ã¹ãã®è§£èª¬ãã¡ã¤ã³ã«ãªã£ã¦ãã¾ãã
SVMãå¦ã³ãã人ã«ã¨ã£ã¦ã¯ããµãã¼ããã¯ã¿ã¼ãã·ã³å ¥éãé称ã赤æ¬ãã¯æé©ãªå ¥éæ¸ã§ããã¨ããããçè«ããå®è·µã¾ã§ãã©ã³ã¹ãã解説ããã¦ãããæ¬æ¸ãèªãã ãã§SVMã®å®è£ ãå¯è½ã«ãªãã ãããæ¬æ¸ã¯SFå°èª¬ã彷彿ã¨ãããç¬ç¹ãªç¿»è¨³ã®æä½ã®ããæ©æ¢°å¦ç¿ã«ä¸æ £ããªèªè ã«ã¨ã£ã¦ã¯èªã¿ããªãã®ã¯è¦ããæ¦ãã¨ãªããæ¬æ¥ãªãåæ¸ããªã¹ã¹ã¡ãããã¨ããã ããããã¯ãã£ã¦ãè±èªã¯ã¡ãã£ã¨ã¨ãã人ãå¤ãã¯ãã ããã§æ¬è¨äºã§ã¯èµ¤æ¬ã®ãªã¹ã¹ã¡ãªèªã¿æ¹ãç´¹ä»ãã¦ã¿ãã 1.ããããã¿ãã§æºåéåãããã æ³³ãã®ãã¾ã人ã§ããããªãæ°´ã«é£ã³è¾¼ãã®ã¯å±éºãã¾ãã¯æºåéåããã¦ä½ã温ãããããã«ã¯ããããããããã¿ã¼ã³èªèãããªã¹ã¹ã¡ãã¨ãããã2ç« ã¾ã§ãä½è£ãããã°3ç« ã¾ã§èªãã§ããã°å åã 2.赤æ¬ãæå ã«ããã¦ä¸æ©å¯ãã ãã¦æºåéåãæ¸ãã ãæ©é赤æ¬ã«ãã©ã¤ï¼ãããã¨ããã ã赤æ¬ã®æ¾ã¤ç´æ°ã§å¿ãèã¾ããªãã
ãã¼ãºãããã®ããã£ã±ãããããªãæ©æ¢°å¦ç¿è¶ å ¥éã ãã©ã²ã£ããç¶ãã¦ããããã ååã¯èå¥é¢æ°ã®åºç¤ã§ãããã¼ã»ãããã³ã®ç°¡åãªèª¬æã¨Perlã«ããå®è£ ã解説ãããå®ã¯ãã®æç¹ã§ãã®æåãªSVM(Support Vector Machineããµãã¼ããã¯ã¿ã¼ãã·ã³)ãã»ã¼å®æãã¦ããã®ã ï¼ã¨ããããã§ä»åã¯SVMãPerlã§ä½ã£ã¦ãã¾ãã話ã åè: ããããã¯ããã人ã®ããã®æ©æ¢°å¦ç¿ã®æç§æ¸ã¾ã¨ã - EchizenBlog-Zwei æ©æ¢°å¦ç¿è¶ å ¥é ããããããã¤ã¼ããã¤ãºã«ã¤ãã¦ã²ã¨ãã¨è¨ã£ã¦ãããã - EchizenBlog-Zwei æ©æ¢°å¦ç¿è¶ å ¥éII ãGmailã®åªå ãã¬ã¤ã§ã使ã£ã¦ããPAæ³ã30åã§ç¿å¾ãããï¼ã - EchizenBlog-Zwei æ©æ¢°å¦ç¿è¶ å ¥éIII ãæ©æ¢°å¦ç¿ã®åºç¤ããã¼ã»ãããã³ã30åã§ä½ã£ã¦å¦ã¶ã - EchizenBlog-Zwei ãã¦
åã«æ¸ããSVMã®è¨äºã§ããL1ã¨ãL2ã¨ããã®ã¯ééããã¨ãã®ããã«ãã£ãã©ãå®ç¾©ããããæå³ãã¦ãããã¨æ¸ãã¦ããããL1ã¨ãL2ã£ã¦æ£ååé ã®è©±ãªããããªãã®ãã¨çåã«æã£ãã1ã¶æã»ã©æéãããã¦ã®ã»ã«ãããã³ãã§ããã確èªãããã¨ãã¦ã«ã¼ãã«å¤å¤é解æãèªãã¨ããã¯ãæ£ååé ã«ã¤ãã¦ã¯L1ã¨L2ã®ä¸¡æ¹ã®èª¬æãæ¸ãã¦ããããæ失ã«é¢ãã¦ã¯æ®éã®Hinge Lossï¼=L1 Lossï¼ããæ¸ãã¦ãªãã ã¨è¨ã訳ã§ããããééãã¡ãã£ããªããã¨ææ¾¹ããæ°æã¡ã§"A dual coordinate descent method for large-scale linear SVM"ãèªã¿ç´ãã¦ã¿ãã¨ããããã£ã±ãL1-SVMã¨ããã®ã¯æ失ãæ®éã®Hinge Lossã§ãL2-SVMã¨ããã®ã¯Hinge Lossã®2ä¹ãæ失ã¨ããã¨æ¸ãã¦ãã£ãã両æ¹ã¨ãæ£ååé ã«ã¤ãã¦ã¯L2æ£ååã使ã£ã¦
Complement Naive BayesãSVMããéããã¼ã¨ä¸»å¼µãã¦ããã®ã§ãSVMããªããæè¿ã¯éããªã£ã¦ããããããã¨ããäºãç´¹ä»ãã¦ã¿ãããè¿å¹´ã¯SVMãªã©ã®å¦ç¿ãé«éã«è¡ãã¨ããææ¡ãè¡ããã¦ãããå®è£ ãå ¬éããã¦ãããã®ãããããã®ä¸ã®ä¸ã¤ã«liblinearã¨ããæ©æ¢°å¦ç¿ã©ã¤ãã©ãªããããã©ã¤ãã©ãªåããæ¨æ¸¬ã§ããéããliblinearã§ã¯ã«ã¼ãã«ã使ããã¨ãåºæ¥ãªãããããããã®åé度ãéãã大è¦æ¨¡ãã¼ã¿ã«é©ç¨ã§ããã¨ããå©ç¹ãããã liblinearãä½ã£ã¦ããã®ã¯libsvmã¨åãç 究ã°ã«ã¼ãã§ãChih-Jen Linãããã¸ã§ã¯ããªã¼ãã¼ã§ããããã ãlibsvmã¯ããªãæåãªã©ã¤ãã©ãªã§ãliblinearã«ã¯ãããã£ãæå³ã§å®å¿æããããï¼liblinearã®æ¹ã¯å ¬éããã¦ãã°ããã¯å²ã¨ãã°ããã£ãããããã©ãï¼ liblinearã«ã¯L1-SVM, L
ãªãªã¼ã¹ãé害æ å ±ãªã©ã®ãµã¼ãã¹ã®ãç¥ãã
ææ°ã®äººæ°ã¨ã³ããªã¼ã®é ä¿¡
å¦çãå®è¡ä¸ã§ã
j次ã®ããã¯ãã¼ã¯
kåã®ããã¯ãã¼ã¯
lãã¨ã§èªã
eã³ã¡ã³ãä¸è¦§ãéã
oãã¼ã¸ãéã
{{#tags}}- {{label}}
{{/tags}}