Highload++ ããã³ DataFest Minsk 2019 ã§ã®ç§ã®ã¹ããŒãã«åºã¥ããŠããŸãã
ä»æ¥ãå€ãã®äººã«ãšã£ãŠãã¡ãŒã«ã¯ãªã³ã©ã€ã³ç掻ã«äžå¯æ¬ ãªéšåã§ãããã®å©ããåããŠãç§ãã¡ã¯ããžãã¹éä¿¡ãè¡ãã財åãããã«ã®äºçŽã泚æãªã©ã«é¢é£ããããããçš®é¡ã®éèŠãªæ å ±ãä¿åããŸãã 2018 幎åã°ã«ãã¡ãŒã«éçºã®è£œåæŠç¥ãçå®ããŸãããçŸä»£ã®ã¡ãŒã«ã¯ã©ãããã¹ãã§ãããã?
ã¡ãŒã«ã¯æ¬¡ã®ãšããã§ã é ããã€ãŸãããŠãŒã¶ãŒãå¢ãç¶ããæ
å ±ãããã²ãŒãããæã䟿å©ãªæ¹æ³ã§æ
å ±ããã£ã«ã¿ãªã³ã°ããæ§é åããæäŸã§ããããã«ããããšã§ãã圌女ã¯ãã£ãšããã§ããã æçšã䜿çšãããšã眰éã®æ¯æããªã©ãããŸããŸãªã¿ã¹ã¯ãã¡ãŒã«ããã¯ã¹å
ã§çŽæ¥è§£æ±ºã§ããŸã (æ®å¿µãªãããç§ããã®æ©èœã䜿çšããŠããŸã)ããããŠãã¡ããåæã«ãã¡ãŒã«ã¯æ
å ±ãä¿è·ããã¹ãã ãé®æãããããã³ã°ããä¿è·ããå¿
èŠããããŸãã å®å
šãª.
ãããã®åéã§ã¯å€ãã®éèŠãªåé¡ãå®çŸ©ãããŠããããã®å€ãã¯æ©æ¢°åŠç¿ã䜿çšããŠå¹æçã«è§£æ±ºã§ããŸããããã§ã¯ãæŠç¥ã®äžéšãšããŠéçºãããæ¢åã®æ©èœã®äŸããæ¹åããšã« 1 ã€ãã€ç€ºããŸãã
- ã¹ããŒããªè¿ä¿¡ãã¡ãŒã«ã«ã¯ã¹ããŒãè¿ä¿¡æ©èœããããŸãããã¥ãŒã©ã« ãããã¯ãŒã¯ã¯æçŽã®æ¬æãåæãããã®æå³ãšç®çãç解ãããã®çµæãè¯å®çãåŠå®çãäžç«ãšãã 3 ã€ã®æãé©åãªå¿çãªãã·ã§ã³ãæäŸããŸããããã«ãããæçŽã«è¿ä¿¡ããæéãå€§å¹ ã«ç¯çŽã§ããæšæºçã§ã¯ãªãé¢çœãæ¹æ³ã§è¿ä¿¡ããããšããããããŸãã
- ã¡ãŒã«ã®ã°ã«ãŒãåãªã³ã©ã€ã³ã¹ãã¢ã§ã®æ³šæã«é¢ãããã®ãç§ãã¡ã¯ãªã³ã©ã€ã³ã§è²·ãç©ãããããšãå€ããååãšããŠåºèã¯æ³šæããšã«è€æ°ã®ã¡ãŒã«ãéä¿¡ã§ããŸããããšãã°ãæ倧ã®ãµãŒãã¹ã§ãã AliExpress ããã¯ã29 åã®æ³šæã«å¯ŸããŠå€§éã®æçŽãå±ããæçµçãªã±ãŒã¹ã§ã¯ãã®æ°ã¯æ倧 XNUMX éã«éããå¯èœæ§ããããšèšç®ãããŸãããããã§ãåºæè¡šçŸèªèã¢ãã«ã䜿çšããŠã泚æçªå·ãæœåºããŸããããã¹ãããã®ãã®ä»ã®æ
å ±ãååŸãããã¹ãŠã®æåã XNUMX ã€ã®ã¹ã¬ããã«ã°ã«ãŒãåããŸãã泚æã«é¢ããåºæ¬æ
å ±ãå¥ã®ããã¯ã¹ã«è¡šç€ºãããããããã®ã¿ã€ãã®é»åã¡ãŒã«ã®åŠçã容æã«ãªããŸãã
- 察è©æ¬ºããã£ãã·ã³ã°ã¯ãç¹ã«å±éºãªã¿ã€ãã®è©æ¬ºã¡ãŒã«ã§ãããæ»æè ã¯ãããå©çšããŠéèæ å ± (ãŠãŒã¶ãŒã®éè¡ã«ãŒããå«ã) ããã°ã€ã³æ å ±ãååŸããããšããŸãããã®ãããªæåã¯ãèŠèŠçã«ãå«ããŠããµãŒãã¹ã«ãã£ãŠéä¿¡ãããå®éã®æåãæš¡å£ããŸãããããã£ãŠãComputer Vision ã®å©ããåããŠã倧äŒæ¥ (Mail.ruãSberãAlfa ãªã©) ããã®ããŽãšã¬ã¿ãŒã®ãã¶ã€ã³ ã¹ã¿ã€ã«ãèªèãããããããã¹ãããã®ä»ã®æ©èœãšãšãã«ã¹ãã ãšãã£ãã·ã³ã°ã®åé¡åã«èæ ®ããŸãã ã
æ©æ¢°åŠç¿
é»åã¡ãŒã«å šè¬ã«ãããæ©æ¢°åŠç¿ã«ã€ããŠå°ã説æããŸããã¡ãŒã«ã¯é«è² è·ã®ã·ã¹ãã ã§ãã1,5 äžäººã® DAU ãŠãŒã¶ãŒã«å¯ŸããŠã30 æ¥ãããå¹³å 30 åéã®æåãåœç€Ÿã®ãµãŒããŒãééããŸããçŽ XNUMX ã®æ©æ¢°åŠç¿ã·ã¹ãã ããå¿ èŠãªæ©èœããã¹ãŠãµããŒãããŠããŸãã
åæåã¯åé¡ãã€ãã©ã€ã³å šäœãééããŸãããŸãã¹ãã ãé®æããé©åãªã¡ãŒã«ãæ®ããŸããã¹ãã ã® 95 ïœ 99% ã¯é©åãªãã©ã«ãã«ããå°éããªãããããŠãŒã¶ãŒã¯ã¹ãã 察çã®åãã«æ°ã¥ããªãããšããããããŸããã¹ãã èªèã¯ç§ãã¡ã®ã·ã¹ãã ã®éåžžã«éèŠãªéšåã§ãããæãé£ããéšåã§ãããªããªããã¹ãã 察çã®åéã§ã¯é²åŸ¡ã·ã¹ãã ãšæ»æã·ã¹ãã ã®éã§çµ¶ããé©å¿ããå¿ èŠãããããããç§ãã¡ã®ããŒã ã«ãšã£ãŠãšã³ãžãã¢ãªã³ã°äžã®ç¶ç¶çãªèª²é¡ãšãªãããã§ãã
次ã«ãæåãšäººéããããããåé¢ããŸãã人ã ããã®ã¡ãŒã«ã¯æãéèŠã§ãããããã¹ããŒã ãªãã©ã€ãªã©ã®æ©èœãæäŸããŠããŸããããããããã®æçŽã¯ 2 ã€ã®éšåã«åãããŠããŸããååŒã«é¢ãããã® - ãããã¯ãµãŒãã¹ããã®éèŠãªæçŽ (ããšãã°ãè³Œå ¥ãããã«ã®äºçŽã®ç¢ºèªã財åã«é¢ãããã®)ãããã³æ å ±ã«é¢ãããã® - ãããã¯ããžãã¹ã®åºåãå²åŒã§ãã
ç§ãã¡ã¯ãååŒã¡ãŒã«ã¯å人çãªéä¿¡ãšåçã®éèŠæ§ãæã£ãŠãããšèããŠããŸããç§ãã¡ã¯æ³šæãèªç©ºåžã®äºçŽã«é¢ããæ å ±ãããã«èŠã€ããå¿ èŠãããããããã®æçŽãæ¢ãã®ã«æéãè²»ããããšãå€ãããããããã®æçŽã¯æå ã«ããã¹ãã§ãããããã£ãŠã䟿å®äžãæ è¡ã泚æã財åããã±ãããç»é²ããããŠæåŸã«çœ°éã® 6 ã€ã®äž»èŠã«ããŽãªã«èªåçã«åé¡ãããŸãã
æ å ±ã¬ã¿ãŒã¯æ倧èŠæš¡ã§ãããããããéèŠåºŠã¯äœãã°ã«ãŒãã§ããããŠãŒã¶ãŒããã®ãããªã¬ã¿ãŒãèªãŸãªããã°ããŠãŒã¶ãŒã®ç掻ã«éèŠãªå€åã¯äœããªããããå³æå¿çããå¿ èŠã¯ãããŸãããæ°ããã€ã³ã¿ãŒãã§ã€ã¹ã§ã¯ããœãŒã·ã£ã« ãããã¯ãŒã¯ãšãã¥ãŒã¹ã¬ã¿ãŒãšãã 2 ã€ã®ã¹ã¬ããã«æããããŸããŠãããããåä¿¡ãã¬ã€ãèŠèŠçã«ã¯ãªã¢ãããéèŠãªã¡ãã»ãŒãžã®ã¿ã衚瀺ãããããã«ãªããŸãã
æŸå
ã·ã¹ãã ã®æ°ãå€ããšãéçšã«å€ãã®å°é£ãçããŸããçµå±ã®ãšãããä»ã®ãœãããŠã§ã¢ãšåæ§ã«ãã¢ãã«ãæéã®çµéãšãšãã«å£åããŸããæ©èœãå£ãããããã·ã³ãæ éããããã³ãŒããæªãã ãããã®ã§ããããã«ãããŒã¿ã¯åžžã«å€åããŠããŸããæ°ããããŒã¿ãè¿œå ããããããŠãŒã¶ãŒã®è¡åãã¿ãŒã³ãå€åããããããããé©åãªãµããŒãããªããã°ã¢ãã«ã®åäœã¯æéã®çµéãšãšãã«ãŸããŸãæªåããŸãã
æ©æ¢°åŠç¿ããŠãŒã¶ãŒã®ç掻ã«æ·±ã浞éããã°ããã»ã©ããšã³ã·ã¹ãã ã«äžãã圱é¿ã¯å€§ãããªãããã®çµæãåžå Žåå è ãåãåãå¯èœæ§ã®ããçµæžçæ倱ãŸãã¯å©çã倧ãããªããšããããšãå¿ããŠã¯ãªããŸããããããã£ãŠããŸããŸãå€ãã®åéã§ããã¬ã€ã€ãŒã¯ ML ã¢ã«ãŽãªãºã ã®åäœã«é©å¿ããŠããŸã (å žåçãªäŸã¯ãåºåãæ€çŽ¢ãããã³ãã§ã«è¿°ã¹ãã¹ãã 察çã§ã)ã
ãŸããæ©æ¢°åŠç¿ã¿ã¹ã¯ã«ã¯ç¹æã®ç¹åŸŽããããŸããã·ã¹ãã ã«å€æŽãå ãããããšãããšãããã軜埮ã§ãã£ãŠããã¢ãã«ã«å€ãã®äœæ¥ãçºçããå¯èœæ§ããããŸããã€ãŸããããŒã¿ã®æäœãåãã¬ãŒãã³ã°ããããã€ã¡ã³ããªã©ãæ°é±éããæ°ãæãããå ŽåããããŸãããããã£ãŠãã¢ãã«ãåäœããç°å¢ã®å€åãéããã°éãã»ã©ããã®ç°å¢ãç¶æããããã«å¿ èŠãªåŽåã¯ããå€ããªããŸããããŒã ã¯å€ãã®ã·ã¹ãã ãäœæããããã«æºè¶³ããŠããŠãããã®ã¡ã³ããã³ã¹ã«ãªãœãŒã¹ã®ã»ãšãã©ãã¹ãŠãè²»ãããæ°ããããšãè¡ãæ©äŒããªãå ŽåããããŸããç§ãã¡ã¯ã¹ãã 察çããŒã ã§ãã®ãããªç¶æ³ã«ééããããšããããŸãããããŠåœŒãã¯ããµããŒããèªååããå¿ èŠããããšããæçœãªçµè«ãäžããŸããã
ãªãŒãã¡ãŒã·ã§ã³
äœãèªååã§ããã®ã§ãããã?å®éãã»ãŒãã¹ãŠã§ããç§ã¯æ©æ¢°åŠç¿ã€ã³ãã©ã¹ãã©ã¯ãã£ãå®çŸ©ãã 4 ã€ã®é åãç¹å®ããŸããã
- ããŒã¿åé;
- è¿œå ã®ãã¬ãŒãã³ã°ã
- å±éãã;
- ãã¹ããšã¢ãã¿ãªã³ã°ã
ç°å¢ãäžå®å®ã§åžžã«å€åããŠããå Žåãã¢ãã«ã®åšå²ã®ã€ã³ãã©ã¹ãã©ã¯ãã£å šäœãã¢ãã«èªäœãããã¯ããã«éèŠã§ããããšãããããŸããããã¯å€ãè¯ãç·åœ¢åé¡åšãããããŸããããé©åãªæ©èœãäžããŠãŠãŒã¶ãŒããè¯ããã£ãŒãããã¯ãåŸãã°ãããããä»å æ©èœãåããæå 端ã®ã¢ãã«ãããã¯ããã«åªããæ©èœãçºæ®ããŸãã
ãã£ãŒãããã¯ã«ãŒã
ãã®ãµã€ã¯ã«ã¯ãããŒã¿åéãè¿œå ã®ãã¬ãŒãã³ã°ãå±éãçµã¿åããããã®ã§ãå®éã«ã¯ã¢ãã«æŽæ°ãµã€ã¯ã«å šäœã§ããã©ãããŠãããéèŠã§ããïŒã¡ãŒã«å ã®ç»é²ã¹ã±ãžã¥ãŒã«ã確èªããŠãã ããã
æ©æ¢°åŠç¿éçºè
ã¯ãããããé»åã¡ãŒã«ã«ç»é²ãããã®ãé²ãã¢ã³ãããã ã¢ãã«ãå®è£
ããŸãããã°ã©ãã¯ãå®éã®ãŠãŒã¶ãŒã®ã¿ãæ®ãå€ãŸã§äœäžããŸãããã¹ãŠãçŽ æŽãããã§ãïŒããã 4 æéãçµéãããšãããããã¹ã¯ãªããã調æŽãããã¹ãŠãéåžžã®ç¶æ
ã«æ»ããŸãããã®å®è£
ã§ã¯ãéçºè
ã¯æ©èœã®è¿œå ãšã¢ãã«ã®åãã¬ãŒãã³ã°ã« 1 ãæãè²»ãããŸããããã¹ãããŒã¯ 4 æéã§é©å¿ã§ããŸããã
èãé£ãã»ã©èŠçã«ãªãããåŸã§ãã¹ãŠãããçŽãå¿ èŠããªãããã«ããã«ã¯ãæåã«ãã£ãŒããã㯠ã«ãŒããã©ã®ããã«ãªãããç°å¢ãå€åããå Žåã«ã©ãããããèããŠããå¿ èŠããããŸããããŒã¿ã®åéããå§ããŸãããããããã¢ã«ãŽãªãºã ã®çæãšãªããŸãã
ããŒã¿åé
ææ°ã®ãã¥ãŒã©ã« ãããã¯ãŒã¯ã«ãšã£ãŠãããŒã¿ã¯å€ããã°å€ãã»ã©è¯ãããšã¯æããã§ãããå®éãããŒã¿ã¯è£œåã®ãŠãŒã¶ãŒã«ãã£ãŠçæãããŸãããŠãŒã¶ãŒã¯ããŒã¿ã«ããŒã¯ãä»ããããšã§ç§ãã¡ãæ¯æŽã§ããŸããããŠãŒã¶ãŒã¯ããæç¹ã§ã¢ãã«ã®å®æã«é£œããŠå¥ã®è£œåã«åãæ¿ããããããããæªçšããããšã¯ã§ããŸããã
æãäžè¬çãªééãã® 1 ã€ã¯ (ããã§ã¯ Andrew Ng ãåç §ããŸã)ããã¹ã ããŒã¿ã»ããã®ã¡ããªã¯ã¹ã«çŠç¹ãåœãŠãããŠããŠãŒã¶ãŒããã®ãã£ãŒãããã¯ã«éç¹ã眮ããªãããšã§ãããŠãŒã¶ãŒããã®ãã£ãŒãããã¯ã¯ãå®éã«ã¯äœæ¥å質ã®äž»ãªå°ºåºŠã§ãããŠãŒã¶ãŒã®ããã®è£œåããŠãŒã¶ãŒãã¢ãã«ã®ä»äºãç解ããªãã£ãããæ°ã«å ¥ããªãã£ãããããšããã¹ãŠãå°ç¡ãã«ãªããŸãã
ãããã£ãŠããŠãŒã¶ãŒã¯åžžã«æ祚ã§ããå¿ èŠãããããã£ãŒãããã¯çšã®ããŒã«ãæäŸãããå¿ èŠããããŸããéèã«é¢é£ããæçŽãã¡ãŒã«ããã¯ã¹ã«å±ãããšèããããå Žåãããã«ãéèããšããŒã¯ãä»ãããŠãŒã¶ãŒãã¯ãªãã¯ããŠããã¯éèã§ã¯ãªãããšã瀺ããã¿ã³ãæç»ããå¿ èŠããããŸãã
ãã£ãŒãããã¯ã®è³ª
ãŠãŒã¶ãŒããã®ãã£ãŒãããã¯ã®è³ªã«ã€ããŠè©±ããŸãããããŸããããªããšãŠãŒã¶ãŒã¯ 1 ã€ã®æŠå¿µã«ç°ãªãæå³ãå
¥ããããšãã§ããŸããããšãã°ãããªããšãããã¯ã ãããŒãžã£ãŒã¯ãã財åããšã¯éè¡ããã®æçŽãæå³ãããšèãããŠãŒã¶ãŒã¯ããã°ãã¡ããããã®å¹Žéã«ã€ããŠã®æçŽã財åãæããšèããŸãã次ã«ãäœã®è«çããªãã«ãã¿ã³ãæŒãã®ã倧奜ããªãŠãŒã¶ãŒãããŸãã第äžã«ããŠãŒã¶ãŒã¯çµè«ã倧ãã誀ã£ãŠããå¯èœæ§ããããŸããç§ãã¡ã®å®è·µã«ãããé¡èãªäŸã¯ãåé¡åã®å®è£
ã§ãã
ãŸãããã¿ã³ã¯äººéã ãã§ãªãããã©ãŠã¶ã®ãµããããããããçš®é¡ã®ãããã«ãã£ãŠãã¯ãªãã¯ãããå¯èœæ§ãããããšãå¿ããªãã§ãã ããããããã£ãŠãçã®ãã£ãŒãããã¯ã¯åŠç¿ã«ã¯è¯ããããŸããããã®æ å ±ã䜿ã£ãŠäœãã§ããã§ãããã?
ç§ãã¡ã¯ 2 ã€ã®ã¢ãããŒãã䜿çšããŸãã
- ãªã³ã¯ããã ML ããã®ãã£ãŒãããã¯ãããšãã°ããªã³ã©ã€ã³ã®ã¢ã³ãããã ã·ã¹ãã ããããåè¿°ããããã«ãéãããæ°ã®å åã«åºã¥ããŠè¿ éãªæ±ºå®ãäžããŸãããããŠãäºåŸçã«åäœãã 2 çªç®ã®é ãã·ã¹ãã ããããŸãããŠãŒã¶ãŒããã®è¡åãªã©ã«é¢ããããå€ãã®ããŒã¿ãå«ãŸããŠããŸãããã®çµæãæãå€ãã®æ å ±ã«åºã¥ããææ決å®ãè¡ãããããæ£ç¢ºæ§ãšå®å šæ§ãé«ãŸããŸãããããã®ã·ã¹ãã ã®åäœã®éãããã¬ãŒãã³ã° ããŒã¿ãšããŠæåã®ã·ã¹ãã ã«éä¿¡ã§ããŸãããããã£ãŠãããåçŽãªã·ã¹ãã ã¯åžžã«ãããè€éãªã·ã¹ãã ã®ããã©ãŒãã³ã¹ã«è¿ã¥ãããšããŸãã
- ã¯ãªãã¯åé¡ãåãŠãŒã¶ãŒã®ã¯ãªãã¯ãç°¡åã«åé¡ãããã®æå¹æ§ãšäœ¿ãããããè©äŸ¡ã§ããŸããããã¯ãã¹ãã 察çã¡ãŒã«ã§ããŠãŒã¶ãŒå±æ§ãå±¥æŽãéä¿¡è å±æ§ãããã¹ãèªäœãåé¡åã®çµæã䜿çšããŠè¡ãããŸãããã®çµæããŠãŒã¶ãŒã®ãã£ãŒãããã¯ãæ€èšŒããèªåã·ã¹ãã ãå®çŸããŸããããŸããåãã¬ãŒãã³ã°ã®å¿ èŠæ§ãã¯ããã«äœããªãããããã®äœæ¥ãä»ã®ãã¹ãŠã®ã·ã¹ãã ã®åºç€ã«ãªãå¯èœæ§ããããŸããäžæ£ç¢ºãªããŒã¿ã«åºã¥ããŠã¢ãã«ããã¬ãŒãã³ã°ãããšçµæã䌎ãããããã®ã¢ãã«ã®äž»ãªåªå äºé ã¯ç²ŸåºŠã§ãã
ããŒã¿ãã¯ãªãŒãã³ã°ããML ã·ã¹ãã ãããã«ãã¬ãŒãã³ã°ããŠããéããŠãŒã¶ãŒã®ããšãå¿ããŠã¯ãªããŸããããªããªããç§ãã¡ã«ãšã£ãŠãã°ã©ãäžã®äœåãäœçŸäžãã®ãšã©ãŒã¯çµ±èšã§ããããŠãŒã¶ãŒã«ãšã£ãŠã¯ãã¹ãŠã®ãã°ãæ²åã ããã§ãããŠãŒã¶ãŒã¯ããã£ãŒãããã¯ãåãåã£ãåŸãäœããã®åœ¢ã§è£œåã®ãšã©ãŒãåãå ¥ããªããã°ãªããªããšããäºå®ã«å ããŠãå°æ¥çã«ã¯åæ§ã®ç¶æ³ãæé€ãããããšãæåŸ ããŸãããããã£ãŠããŠãŒã¶ãŒã«æ祚ã®æ©äŒãäžããã ãã§ãªãããã£ãŒãããã¯ã®ã¯ãªãã¯ããšã«å人çãªãã¥ãŒãªã¹ãã£ãã¯ãäœæãããªã©ãML ã·ã¹ãã ã®åäœãä¿®æ£ããããšãåžžã«äŸ¡å€ããããŸããã¡ãŒã«ã®å Žåãããã¯ãã£ã«ã¿ãªã³ã°æ©èœãªã©ã§ãããã®ãŠãŒã¶ãŒã®éä¿¡è ãšã¿ã€ãã«ã«ããæçŽãªã©ã
ãŸããä»ã®ãŠãŒã¶ãŒãåæ§ã®åé¡ã«æ©ãŸãããªãããã«ãåèªåã¢ãŒããŸãã¯æåã¢ãŒãã§ãµããŒããžã®ã¬ããŒãããªã¯ãšã¹ãã«åºã¥ããŠã¢ãã«ãæ§ç¯ããå¿ èŠããããŸãã
åŠç¿ã®ããã®ãã¥ãŒãªã¹ãã£ãã¯
ãããã®ãã¥ãŒãªã¹ãã£ãã¯ãšæŸèæã«ã¯ 2 ã€ã®åé¡ããããŸãã 1 ã€ç®ã¯ãå¢ãç¶ããæŸèæã®æ°ãé·æã«ããã£ãŠç¶æããã®ãé£ãããšããããšã§ãããŸããŠãããã®å質ãæ§èœãç¶æããã®ãé£ãããšããããšã§ãã 2 çªç®ã®åé¡ã¯ããšã©ãŒãé »ç¹ã«çºçããããã§ã¯ãªããã¢ãã«ãããã«ãã¬ãŒãã³ã°ããããã«æ°åã¯ãªãã¯ããã ãã§ã¯ååã§ã¯ãªãããšã§ãã次ã®ã¢ãããŒããé©çšãããšãããã 2 ã€ã®ç¡é¢ä¿ãªåœ±é¿ãå€§å¹ ã«äžåã§ããããã«èŠããŸãã
- äžæçãªæŸèæãäœæããŸãã
- ããããã¢ãã«ã«ããŒã¿ãéä¿¡ããåãåã£ãããŒã¿ãå«ããŠã¢ãã«èªäœãå®æçã«æŽæ°ããŸãããã¡ãããããã§ã¯ããã¬ãŒãã³ã° ã»ããå ã®ããŒã¿ã®å質ãäœäžãããªãããã«ããã¥ãŒãªã¹ãã£ãã¯ã®ç²ŸåºŠãé«ãããšãéèŠã§ãã
- 次ã«ãæŸèæãäœåããããã«ã¢ãã¿ãªã³ã°ãèšå®ããŸãããã°ããããŠæŸèæãæ©èœããªããªããã¢ãã«ã«å®å šã«èŠãããå Žåã¯ãå®å šã«åãå€ãããšãã§ããŸããçŸåšããã®åé¡ãåã³èµ·ããå¯èœæ§ã¯äœãã§ãã
ãããã£ãŠãæŸèæã®è»éã¯éåžžã«åœ¹ç«ã¡ãŸããéèŠãªããšã¯ã圌ãã®ãµãŒãã¹ã¯ç·æ¥ã§ãããæ°žç¶çãªãã®ã§ã¯ãªããšããããšã§ãã
è¿œå ãã¬ãŒãã³ã°
åãã¬ãŒãã³ã°ã¯ããŠãŒã¶ãŒãŸãã¯ä»ã®ã·ã¹ãã ããã®ãã£ãŒãããã¯ã®çµæãšããŠååŸãããæ°ããããŒã¿ãè¿œå ãããã®ããŒã¿ã«åºã¥ããŠæ¢åã®ã¢ãã«ããã¬ãŒãã³ã°ããããã»ã¹ã§ããè¿œå ã®ãã¬ãŒãã³ã°ã«ã¯ããã€ãã®åé¡ãçºçããå¯èœæ§ããããŸãã
- ã¢ãã«ã¯è¿œå ã®ãã¬ãŒãã³ã°ããµããŒãããŠããªãã ãã§ãæåããã®ã¿åŠç¿ããå¯èœæ§ããããŸãã
- è¿œå ã®ãã¬ãŒãã³ã°ãçç£çŸå Žã§ã®äœæ¥ã®å質ã確å®ã«åäžããããšã¯ãèªç¶çã®æ¬ã®ã©ãã«ãæžãããŠããŸãããå€ãã®å Žåããã®éãã€ãŸãæªåã®ã¿ãèµ·ãããŸãã
- å€åã¯äºæž¬ã§ããªãå ŽåããããŸããããã¯ãç§ãã¡ãç¬èªã«ç¹å®ããããªã埮åŠãªç¹ã§ãã A/B ãã¹ãã§æ°ããã¢ãã«ãçŸåšã®ã¢ãã«ãšåæ§ã®çµæã瀺ãããšããŠãããããåãããã«åäœããããšãæå³ããããã§ã¯ãããŸããã圌ãã®äœæ¥ã¯ããã 1% ç°ãªãå¯èœæ§ããããããã«ãã£ãŠæ°ããªãšã©ãŒãçºçãããããã§ã«ä¿®æ£ãããå€ããšã©ãŒãè¿ããããããå¯èœæ§ããããŸããç§ãã¡ããŠãŒã¶ãŒããçŸåšã®ãšã©ãŒã«å¯ŸåŠããæ¹æ³ããã§ã«ç¥ã£ãŠããŸãããå€æ°ã®æ°ãããšã©ãŒãçºçããå ŽåããŠãŒã¶ãŒã¯äºæž¬å¯èœãªåäœãæåŸ ããŠãããããäœãèµ·ãã£ãŠããã®ãç解ã§ããªãå ŽåããããŸãã
ãããã£ãŠãè¿œå ã®ãã¬ãŒãã³ã°ã§æãéèŠãªããšã¯ãã¢ãã«ãæ¹åãããŠããããå°ãªããšãæªåããŠããªãããšã確èªããããšã§ãã
è¿œå ã®ãã¬ãŒãã³ã°ã«ã€ããŠè©±ããšãã«æåã«æãæµ®ãã¶ã®ã¯ãã¢ã¯ãã£ã ã©ãŒãã³ã°ã®ã¢ãããŒãã§ããããã¯ã©ãããæå³ã§ããïŒããšãã°ãåé¡åšã¯é»åã¡ãŒã«ãéèã«é¢é£ããŠãããã©ãããå€æãããã®å€æå¢çã®åšå²ã«ã©ãã«ä»ãã®äŸã®ãµã³ãã«ãè¿œå ããŸããããã¯ãããšãã°ãå€ãã®ãã£ãŒãããã¯ãããããªã³ã©ã€ã³ã§ã¢ãã«ããã¬ãŒãã³ã°ã§ããåºåã®å Žåã«ããŸãæ©èœããŸãããŸãããã£ãŒãããã¯ãã»ãšãã©ãªãå Žåã¯ãæ¬çªããŒã¿ã®ååžã«å¯ŸããŠéåžžã«åã£ããµã³ãã«ãåŸããããããããã«åºã¥ããŠéçšäžã®ã¢ãã«ã®åäœãè©äŸ¡ããããšã¯äžå¯èœã«ãªããŸãã
å®éãç§ãã¡ã®ç®æšã¯ãå€ããã¿ãŒã³ãæ¢ç¥ã®ã¢ãã«ãä¿åããæ°ãããã¿ãŒã³ãååŸããããšã§ããããã§ã¯ç¶ç¶ãéèŠã§ããããŒã«ã¢ãŠãããããã«å€å€§ãªåŽåãè²»ãããã¢ãã«ã¯ãã§ã«åäœããŠããããããã®ããã©ãŒãã³ã¹ã«éäžã§ããŸãã
ã¡ãŒã«ã§ã¯ãããªãŒãç·åœ¢ããã¥ãŒã©ã« ãããã¯ãŒã¯ãªã©ãããŸããŸãªã¢ãã«ã䜿çšãããŸããããããã«ã€ããŠãç¬èªã®è¿œå ãã¬ãŒãã³ã° ã¢ã«ãŽãªãºã ãäœæããŸããè¿œå ã®ãã¬ãŒãã³ã°ã®ããã»ã¹ã§ã¯ãæ°ããããŒã¿ã ãã§ãªããå€ãã®å Žåã以äžã®ãã¹ãŠã®ã¢ã«ãŽãªãºã ã§èæ ®ãããæ°ããç¹åŸŽãåãåããŸãã
ç·åœ¢ã¢ãã«
ããžã¹ãã£ãã¯ååž°ããããšããŸãã次ã®ã³ã³ããŒãã³ãããæ倱ã¢ãã«ãäœæããŸãã
- æ°ããããŒã¿ã® LogLossã
- æ°ããç¹åŸŽã®éã¿ãæ£èŠåããŸã (å€ãç¹åŸŽã«ã¯è§ŠããŸãã)ã
- ãŸããå€ããã¿ãŒã³ãä¿åããããã«å€ãããŒã¿ããåŠç¿ããŸãã
- ãããŠãããããæãéèŠãªããšã¯ã調åæ£ååãè¿œå ããããšã§ããããã«ãããåºæºã«åŸã£ãŠå€ãã¢ãã«ã«å¯ŸããŠéã¿ãããŸãå€ãããªãããšãä¿èšŒãããŸãã
åæ倱ã³ã³ããŒãã³ãã«ã¯ä¿æ°ããããããçžäºæ€èšŒãéããŠããŸãã¯è£œåèŠä»¶ã«åºã¥ããŠã¿ã¹ã¯ã«æé©ãªå€ãéžæã§ããŸãã
ÐеÑевÑÑ
ãã·ãžã§ã³ããªãŒã«ç§»ããŸããããããªãŒã®è¿œå ãã¬ãŒãã³ã°çšã«æ¬¡ã®ã¢ã«ãŽãªãºã ãã³ã³ãã€ã«ããŸããã
- æ¬çªç°å¢ã§ã¯ãå€ãããŒã¿ã»ããã§ãã¬ãŒãã³ã°ããã 100 ïœ 300 æ¬ã®æšã®ãã©ã¬ã¹ããå®è¡ãããŸãã
- æåŸã«ãM = 5 åãåé€ãã2M = 10 åã®æ°ããããŒã¿ãè¿œå ããŸããããã¯ããŒã¿ã»ããå šäœã§ãã¬ãŒãã³ã°ãããŸãããæ°ããããŒã¿ã®éã¿ãé«ããã¢ãã«ã®æ®µéçãªå€æŽãåœç¶ä¿èšŒãããŸãã
æããã«ãæéã®çµéãšãšãã«æšã®æ°ã¯å€§å¹ ã«å¢å ãããããã¿ã€ãã³ã°ã«åãããŠå®æçã«æšãæžããå¿ èŠããããŸãããããè¡ãããã«ãçŸåšåºãæ®åããŠããç¥èèžç (KD) ã䜿çšããŸãããã®åäœåçã«ã€ããŠç°¡åã«èª¬æããŸãã
- çŸåšã®ãè€éãªãã¢ãã«ããããŸããããããã¬ãŒãã³ã° ããŒã¿ ã»ããã«å¯ŸããŠå®è¡ããåºåã§ã¯ã©ã¹ã®ç¢ºçååžãååŸããŸãã
- 次ã«ãã¯ã©ã¹ååžãã¿ãŒã²ããå€æ°ãšããŠäœ¿çšããŠã¢ãã«ã®çµæãç¹°ãè¿ãããã«ãåŠçã¢ãã« (ãã®å Žåã¯ããªãŒã®æ°ãå°ãªãã¢ãã«) ããã¬ãŒãã³ã°ããŸãã
- ããã§ãããŒã¿ ã»ãã ããŒã¯ã¢ãããäžå䜿çšããŠããªããããä»»æã®ããŒã¿ã䜿çšã§ããããšã«æ³šæããããšãéèŠã§ãããã¡ãããæŠéã¹ããªãŒã ããã®ããŒã¿ ãµã³ãã«ãåŠçã¢ãã«ã®ãã¬ãŒãã³ã° ãµã³ãã«ãšããŠäœ¿çšããŸãããããã£ãŠããã¬ãŒãã³ã° ã»ããã«ãã£ãŠã¢ãã«ã®ç²ŸåºŠãä¿èšŒã§ããã¹ããªãŒã ãµã³ãã«ã«ãã£ãŠæ¬çªååžã§ã®åæ§ã®ããã©ãŒãã³ã¹ãä¿èšŒããããã¬ãŒãã³ã° ã»ããã®åããè£åãããŸãã
ããã 2 ã€ã®ææ³ (ããªãŒãè¿œå ãããã¬ããžèžçã䜿çšããŠããªãŒã®æ°ãå®æçã«æžãã) ãçµã¿åãããããšã§ãæ°ãããã¿ãŒã³ã®å°å
¥ãšå®å
šãªç¶ç¶æ§ãä¿èšŒãããŸãã
KD ã®å©ããåããŠããã£ãŒãã£ã®åé€ãã®ã£ããã®åŠçãªã©ãã¢ãã« ãã£ãŒãã£ã«å¯ŸããŠããŸããŸãªæäœãå®è¡ããŸããç§ãã¡ã®å ŽåãããŒã¿ããŒã¹ã«ä¿åãããŠããéèŠãªçµ±èšçç¹åŸŽ (éä¿¡è ãããã¹ã ããã·ã¥ãURL ãªã©) ãå€æ°ãããŸããããããã¯å€±æããåŸåããããŸãããã¡ããããã¬ãŒãã³ã° ã»ããã§ã¯å€±æç¶æ³ãçºçããªããããã¢ãã«ã¯ãã®ãããªã€ãã³ãã®å±éã«å¯Ÿå¿ããæºåãã§ããŠããŸããããã®ãããªå ŽåãKD ãšæ¡åŒµæè¡ãçµã¿åãããŸããããŒã¿ã®äžéšããã¬ãŒãã³ã°ãããšãã«ãå¿ èŠãªç¹åŸŽãåé€ãŸãã¯ãªã»ããããå ã®ã©ãã« (çŸåšã®ã¢ãã«ã®åºå) ãååŸããŸããããã«ãããã¹ãã¥ãŒãã³ã ã¢ãã«ã¯ãã®ååžãç¹°ãè¿ãããšãåŠç¿ããŸãã ã
ããæ·±å»ãªã¢ãã«æäœãè¡ãããã»ã©ãå¿
èŠãªã¹ã¬ãã ãµã³ãã«ã®å²åã倧ãããªãããšãããããŸããã
æãåçŽãªæäœã§ããç¹åŸŽã®åé€ã«ã¯ããããŒã®ããäžéšã®ã¿ãå¿ èŠã§ããå€æŽãããç¹åŸŽã¯ããã€ãã ãã§ãããçŸåšã®ã¢ãã«ã¯åãã»ããã§ãã¬ãŒãã³ã°ãããŠãããããéãã¯æå°éã§ããã¢ãã«ãç°¡çŽ åããã«ã¯ (ããªãŒã®æ°ãæ°åæžãã)ããã§ã« 50 察 50 ãå¿ èŠã§ãããŸããã¢ãã«ã®ããã©ãŒãã³ã¹ã«é倧ãªåœ±é¿ãäžããéèŠãªçµ±èšçç¹åŸŽãçç¥ãããŠããå Žåããã®äœæ¥ãå¹³æºåããã«ã¯ããã«å€ãã®ãããŒãå¿ èŠã§ããå šæå察å¿ã®è±èœé²æ¢ã¢ãã«ãæ°ç»å Žã
é«éããã¹ã
FastText ã«ç§»ããŸããããåèªã®è¡šçŸ (åã蟌ã¿) ã¯ãåèªèªäœãšãã®ãã¹ãŠã®æå N ã°ã©ã (éåžžã¯ããªã°ã©ã ) ã®åã蟌ã¿ã®åèšã§æ§æãããããšãæãåºããŠãã ãããããªãå€ãã®ããªã°ã©ã ãååšããå¯èœæ§ãããããããã±ãã ããã·ã¥ã䜿çšãããŸããã€ãŸãã空éå šäœãç¹å®ã®åºå®ããã·ã¥ãããã«å€æããŸããçµæãšããŠãã¯ãŒãæ°ïŒãã±ãããããã®å å±€ã®æ¬¡å ã§éã¿è¡åãåŸãããŸãã
ããã«ãã¬ãŒãã³ã°ãéãããšãåèªãããªã°ã©ã ãªã©ã®æ°ããèšå·ã衚瀺ãããŸãã Facebook ã®æšæºçãªãã©ããŒã¢ãã ãã¬ãŒãã³ã°ã§ã¯ãéèŠãªããšã¯äœãèµ·ãããŸãããã¯ãã¹ãšã³ããããŒãæã€å€ãéã¿ã®ã¿ãæ°ããããŒã¿ã§åãã¬ãŒãã³ã°ãããŸãããããã£ãŠãæ°ããæ©èœã¯äœ¿çšãããŸããããã¡ããããã®ã¢ãããŒãã«ã¯ãéçšç°å¢ã«ãããã¢ãã«ã®äºæž¬äžå¯èœæ§ã«é¢é£ããäžèšã®ãã¹ãŠã®æ¬ ç¹ããããŸãããã®ãããFastText ãå°ãå€æŽããŸããããã¹ãŠã®æ°ããéã¿ (ã¯ãŒããšãã©ã€ã°ã©ã ) ãè¿œå ããã¯ãã¹ãšã³ããããŒã§è¡åå šäœãæ¡åŒµããç·åœ¢ã¢ãã«ãšã®é¡æšã«ãã調åæ£ååãè¿œå ããŸããããã«ãããå€ãéã¿ã®ããããªå€åãä¿èšŒãããŸãã
CNN
ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã¯ããå°ãè€éã§ãã CNN ã§æåŸã®å±€ãå®æããŠããå Žåã¯ããã¡ãã調åæ£ååãé©çšããŠé£ç¶æ§ãä¿èšŒã§ããŸãããããããããã¯ãŒã¯å
šäœã®è¿œå ã®ãã¬ãŒãã³ã°ãå¿
èŠãªå Žåããã®ãããªæ£ååããã¹ãŠã®å±€ã«é©çšããããšã¯ã§ããªããªããŸãããã ããããªãã¬ããæ倱ãéããŠçžè£çãªåã蟌ã¿ããã¬ãŒãã³ã°ãããªãã·ã§ã³ããããŸã (
ããªãã«ãã¹
ãã£ãã·ã³ã°å¯Ÿçã¿ã¹ã¯ãäŸãšããŠãããªãã¬ããæ倱ãäžè¬çã«èŠãŠã¿ãŸããããåœç€Ÿã®ããŽã ãã§ãªããä»ç€Ÿã®ããŽã®ããžãã£ããªäŸãšãã¬ãã£ããªäŸãåãäžããŸããæåã®ã¯ã©ã¹éã®è·é¢ãæå°åãã2 çªç®ã®ã¯ã©ã¹éã®è·é¢ãæ倧åããŸããã¯ã©ã¹ã®ã³ã³ãã¯ãæ§ã確ä¿ããããã«ãå°ããªã®ã£ãããèšããŠãããè¡ããŸãã
ãããã¯ãŒã¯ãããã«ãã¬ãŒãã³ã°ãããšãèšé空éã¯å®å
šã«å€åããåã®ç©ºéãšå®å
šã«äºææ§ããªããªããŸããããã¯ãã¯ãã«ã䜿çšããåé¡ã§ã¯æ·±å»ãªåé¡ã§ãããã®åé¡ãåé¿ããã«ã¯ããã¬ãŒãã³ã°äžã«å€ãåã蟌ã¿ãæ··ããŸãã
æ°ããããŒã¿ããã¬ãŒãã³ã° ã»ããã«è¿œå ããã¢ãã«ã® 1 çªç®ã®ããŒãžã§ã³ãæåãããã¬ãŒãã³ã°ããŠããŸãã第 2 段éã§ã¯ããããã¯ãŒã¯ãããã«ãã¬ãŒãã³ã°ããŸã (埮調æŽ)ãæåã«æåŸã®å±€ãå®æãã次ã«ãããã¯ãŒã¯å šäœã®ããªãŒãºã解é€ãããŸããããªãã¬ãããæ§æããããã»ã¹ã§ã¯ããã¬ãŒãã³ã°ãããã¢ãã«ã䜿çšããŠåã蟌ã¿ã®äžéšã®ã¿ãèšç®ããæ®ãã¯å€ãã¢ãã«ã䜿çšããŠèšç®ããŸãããããã£ãŠãè¿œå ã®ãã¬ãŒãã³ã°ã®ããã»ã¹ã§ãèšé空é vXNUMX ãš vXNUMX ã®äºææ§ã確ä¿ããŸãã調åæ£ååã®ãŠããŒã¯ãªããŒãžã§ã³ã
ã¢ãŒããã¯ãã£å šäœ
äŸãšããŠã¢ã³ãã¹ãã ã䜿çšããã·ã¹ãã å šäœãèãããšãã¢ãã«ã¯åé¢ãããŠããããçžäºã«ãã¹ããããŠããŸããåçãããã¹ãããã®ä»ã®æ©èœãæ®åœ±ããCNN ãš Fast Text ã䜿çšããŠåã蟌ã¿ãååŸããŸãã次ã«ãåã蟌ã¿ã®äžã«åé¡åãé©çšãããããŸããŸãªã¯ã©ã¹ (æåã®çš®é¡ãã¹ãã ãããŽã®ååš) ã®ã¹ã³ã¢ãæäŸãããŸããä¿¡å·ãšæšèã¯ãã§ã«æçµæ±ºå®ãäžãããã«æš¹æã«å ¥ãå§ããŠããŸãããã®ã¹ããŒã ã®åã ã®åé¡åã«ããããã¹ãŠã®ããŒã¿ãçã®åœ¢åŒã§ãã·ãžã§ã³ ããªãŒã«å ¥åããã®ã§ã¯ãªããã·ã¹ãã ã®çµæãããé©åã«è§£éããåé¡ãçºçããå Žåã«ã³ã³ããŒãã³ããããå ·äœçã«åãã¬ãŒãã³ã°ããããšãå¯èœã«ãªããŸãã
ãã®çµæãããããã¬ãã«ã§ã®ç¶ç¶æ§ãä¿èšŒãããŸãã CNN ãš Fast Text ã®æäžäœã¬ãã«ã§ã¯èª¿åæ£ååã䜿çšããäžéã®åé¡åšã§ã¯ç¢ºçååžã®äžè²«æ§ãä¿ã€ããã«èª¿åæ£ååãšã¬ãŒã ãã£ãªãã¬ãŒã·ã§ã³ã䜿çšããŸããããã§ãããããªãŒ ããŒã¹ãã£ã³ã°ã¯æ®µéçã«ãã¬ãŒãã³ã°ããããç¥èã®èžçã䜿çšããŠãã¬ãŒãã³ã°ããŸãã
äžè¬ã«ããã®ãããªå ¥ãåã«ãªã£ãæ©æ¢°åŠç¿ã·ã¹ãã ã®ç¶æã¯ãäžäœã¬ãã«ã®ã³ã³ããŒãã³ããäžäœã®ã·ã¹ãã å šäœã®æŽæ°ã«ã€ãªãããããéåžžã¯é¢åã§ããããããç§ãã¡ã®èšå®ã§ã¯åã³ã³ããŒãã³ãããããã«å€æŽããã以åã®ã³ã³ããŒãã³ããšäºææ§ããããããæ§é å šäœãåãã¬ãŒãã³ã°ããããšãªããã·ã¹ãã å šäœãéšåçã«æŽæ°ã§ããé倧ãªãªãŒããŒããããªãã§ãµããŒãã§ããŸãã
å±éãã
ããŒã¿åéãšããŸããŸãªã¿ã€ãã®ã¢ãã«ã®è¿œå ãã¬ãŒãã³ã°ã«ã€ããŠèª¬æããã®ã§ãæ¬çªç°å¢ãžã®å±éã«ç§»ããŸãã
A/B ãã¹ã
åã«è¿°ã¹ãããã«ãããŒã¿åéã®éçšã§éåžžã¯åã£ããµã³ãã«ãåŸãããããããã¢ãã«ã®å®çšŒåããã©ãŒãã³ã¹ãè©äŸ¡ããããšã¯äžå¯èœã§ãããã®ããããããã€æã«ã¯ã¢ãã«ã以åã®ããŒãžã§ã³ãšæ¯èŒããŠå®éã«ã©ããªã£ãŠããã®ããææ¡ãããã€ãŸãA/Bãã¹ããå®æœããå¿ èŠããããŸããå®éããã£ãŒãã®å±éãšåæã®ããã»ã¹ã¯éåžžã«æ¥åžžçãªãã®ã§ãããç°¡åã«èªååã§ããŸããã¢ãã«ã®å¿çãšãŠãŒã¶ãŒã®ãã£ãŒãããã¯ã«é¢ããå©çšå¯èœãªãã¹ãŠã®ææšãåéããªãããã¢ãã«ããŠãŒã¶ãŒã® 5%ã30%ã50%ã100% ã«æ®µéçã«å±éããŸããé倧ãªå€ãå€ã®å Žåã¯ãã¢ãã«ãèªåçã«ããŒã«ããã¯ããŸãããã®ä»ã®å Žåã¯ãååãªæ°ã®ãŠãŒã¶ãŒ ã¯ãªãã¯ãåéããåŸãããŒã»ã³ããŒãžãå¢ããããšã決å®ããŸãããã®çµæãæ°ããã¢ãã«ã¯å®å šã«èªåçã«ãŠãŒã¶ãŒã® 50% ã«æäŸããã察象è å šäœãžã®å±éã¯äººã«ãã£ãŠæ¿èªãããŸããããã®ã¹ãããã¯èªååã§ããŸãã
ãã ããA/B ãã¹ã ããã»ã¹ã«ã¯æé©åã®äœå°ããããŸããå®éã®ãšãããA/B ãã¹ãã¯éåžžã«æéãããã (ç§ãã¡ã®å Žåããã£ãŒãããã¯ã®éã«å¿ã㊠6 ïœ 24 æéããããŸã)ããã®ããéåžžã«è²»çšããããããªãœãŒã¹ãéãããŠããŸããããã«ãæ¬è³ªçã« A/B ãã¹ãå šäœã®æéãççž®ããã«ã¯ããã¹ãã®ãããŒã®ååã«é«ãããŒã»ã³ããŒãžãå¿ èŠã§ã (çµ±èšçã«ææãªãµã³ãã«ãéããŠã¡ããªã¯ã¹ãå°ããªããŒã»ã³ããŒãžã§è©äŸ¡ããã«ã¯éåžžã«é·ãæéããããå¯èœæ§ããããŸã)ã A/B ã¹ãããã®æ°ã¯éåžžã«éãããŠããŸããæããã«ãæãææãªã¢ãã«ã®ã¿ããã¹ãããå¿ èŠããããè¿œå ã®ãã¬ãŒãã³ã° ããã»ã¹äžã«ããªãå€ãã®ã¢ãã«ãåãåããŸãã
ãã®åé¡ã解決ããããã«ãA/B ãã¹ãã®æåãäºæž¬ããå¥ã®åé¡åšããã¬ãŒãã³ã°ããŸããããããè¡ãããã«ããã¬ãŒãã³ã° ã»ãããé 延ã»ãããããã³ã¹ããªãŒã ããã®ãµã³ãã«ã«é¢ããææ決å®çµ±èšã粟床ãåçŸçãããã³ãã®ä»ã®ã¡ããªã¯ã¹ãç¹åŸŽãšããŠååŸããŸãããŸãããã¥ãŒãªã¹ãã£ãã¯ã䜿çšããŠã¢ãã«ãå®çšŒåäžã®çŸåšã®ã¢ãã«ãšæ¯èŒããã¢ãã«ã®è€éããèæ ®ããŸããããããã¹ãŠã®æ©èœã䜿çšããŠããã¹ãå±¥æŽã§ãã¬ãŒãã³ã°ãããåé¡åšãåè£ã¢ãã« (ãã®å Žåã¯æšã®æ£®) ãè©äŸ¡ããA/B ãã¹ãã§ã©ãã䜿çšãããã決å®ããŸãã
å°å
¥æç¹ã§ã¯ããã®ã¢ãããŒãã«ãããA/B ãã¹ãã®æåæ°ãæ°åã«å¢ããããšãã§ããŸããã
ãã¹ããšã¢ãã¿ãªã³ã°
å¥åŠãªããšã«ãæ€æ»ãã¢ãã¿ãªã³ã°ã¯ç§ãã¡ã®å¥åº·ã害ãããã®ã§ã¯ãªãããããéã«å¥åº·ãæ¹åããäžå¿ èŠãªã¹ãã¬ã¹ãã解æŸããŠãããã®ã§ãããã¹ãã«ãã£ãŠé害ãé²ãããšãã§ããç£èŠã«ãã£ãŠé害ãé©æã«æ€åºããŠãŠãŒã¶ãŒãžã®åœ±é¿ã軜æžã§ããŸãã
ããã§ãé ããæ©ããã·ã¹ãã ã¯åžžã«ééããç¯ããã®ã§ããããšãç解ããããšãéèŠã§ããããã¯ããœãããŠã§ã¢ã®éçºãµã€ã¯ã«ã«ãããã®ã§ããã·ã¹ãã éçºã®åæ段éã§ã¯ããã¹ãŠãå®å®ããã€ãããŒã·ã§ã³ã®äž»èŠæ®µéãå®äºãããŸã§ãåžžã«å€ãã®ãã°ãçºçããŸããããããæéã®çµéãšãšãã«ãšã³ããããŒãå¢å€§ããæåã«èª¬æããåšå²ã®ã³ã³ããŒãã³ãã®å£åãããŒã¿ã®å€æŽã«ããããšã©ãŒãåã³çºçããŸãã
ããã§æ³šæãããã®ã¯ãããããæ©æ¢°åŠç¿ã·ã¹ãã ã¯ã©ã€ããµã€ã¯ã«å šäœãéããŠå©çã®èŠ³ç¹ããæ€èšãããã¹ãã§ãããšããããšã§ãã以äžã®ã°ã©ãã¯ããŸããªã¿ã€ãã®ã¹ãã ãææããã·ã¹ãã ã®åäœäŸã瀺ããŠããŸã (ã°ã©ãã®ç·ã¯ãŒãã«è¿ã)ãããæ¥ã誀ã£ãŠãã£ãã·ã¥ãããå±æ§ã«ããã圌女ã¯ãããããªã£ãŠããŸããŸããã幞éãªããšã«ãç°åžžãªããªã¬ãŒãç£èŠããæ©èœããªãã£ããããã·ã¹ãã ã¯ææ決å®ã®å¢çã«ãããã¹ãã ããã©ã«ããŒã«å€§éã®æçŽãä¿åãå§ããŸãããçµæãä¿®æ£ããã«ãããããããã·ã¹ãã ã¯ãã§ã«äœåºŠãééããç¯ããŠããããã5幎çµã£ãŠãå ãåããŸããããããŠãããã¯ã¢ãã«ã®ã©ã€ããµã€ã¯ã«ã®èŠ³ç¹ããèŠããšå®å šã«å€±æã§ãã
ãããã£ãŠãã¢ãã¿ãªã³ã°ã®ãããªåçŽãªããšãã¢ãã«ã®å¯¿åœã«ãããŠéèŠã«ãªãå¯èœæ§ããããŸããæšæºçã§æçœãªææšã«å ããŠãã¢ãã«ã®å¿çãšã¹ã³ã¢ã®ååžãããã³äž»èŠãªç¹åŸŽå€ã®ååžãèæ
®ããŸãã KL ãã€ããŒãžã§ã³ã¹ã䜿çšãããšãçŸåšã®ååžãéå»ã®ååžãšæ¯èŒããããA/B ãã¹ãã®å€ãæ®ãã®ã¹ããªãŒã ãšæ¯èŒãããã§ãããããã¢ãã«ã®ç°åžžã«æ°ã¥ããå€æŽãã¿ã€ã ãªãŒã«ããŒã«ããã¯ã§ããŸãã
ã»ãšãã©ã®å Žåãã·ã¹ãã ã®æåã®ããŒãžã§ã³ã¯ãå°æ¥ã®ç£èŠãšããŠäœ¿çšããåçŽãªãã¥ãŒãªã¹ãã£ãã¯ãŸãã¯ã¢ãã«ã䜿çšããŠèµ·åãããŸããããšãã°ãç¹å®ã®ãªã³ã©ã€ã³ ã¹ãã¢ã®éåžžã®ã¢ãã«ãšæ¯èŒã㊠NER ã¢ãã«ãç£èŠããããããšæ¯èŒããŠåé¡åã®ã«ãã¬ããžãäœäžããŠããå Žåããã®çç±ãç解ããŸãããã¥ãŒãªã¹ãã£ãã¯ã®ãã 1 ã€ã®äŸ¿å©ãªäœ¿ãæ¹!
çµæ
ãã®èšäºã®éèŠãªã¢ã€ãã¢ãããäžåºŠèŠãŠã¿ãŸãããã
- ãã£ãããã¯ãç§ãã¡ã¯åžžã«ãŠãŒã¶ãŒã®ããšãèããŠããŸãããŠãŒã¶ãŒãç§ãã¡ã®ééããã©ã®ããã«åãæ¢ããããã©ã®ããã«å ±åã§ãããããŠãŒã¶ãŒã¯ã¢ãã«ããã¬ãŒãã³ã°ããããã®çŽç²ãªãã£ãŒãããã¯æºã§ã¯ãªããè£å©ç㪠ML ã·ã¹ãã ã®å©ããåããŠã¯ãªã¢ããå¿ èŠãããããšãå¿ããªãã§ãã ããããŠãŒã¶ãŒããã®ä¿¡å·ãåéã§ããªãå Žåã¯ãæ¥ç¶ãããã·ã¹ãã ãªã©ã®ä»£æ¿ãã£ãŒãããã¯æºãæ¢ããŸãã
- è¿œå ãã¬ãŒãã³ã°ãããã§éèŠãªã®ã¯ç¶ç¶æ§ã§ãããããçŸåšã®çç£ã¢ãã«ã«äŸåããŸãã調åæ£ååãåæ§ã®ããªãã¯ã«ãããæ°ããã¢ãã«ã以åã®ã¢ãã«ãšå€§ããå€ãããªãããã«ãã¬ãŒãã³ã°ããŸãã
- å±éãããã¡ããªã¯ã¹ã«åºã¥ãèªåãããã€ã¡ã³ãã«ãããã¢ãã«ã®å®è£ æéãå€§å¹ ã«ççž®ãããŸããçµ±èšãšææ決å®ã®ååžããŠãŒã¶ãŒã®è»¢åæ°ãç£èŠããããšã¯ãå®ãããªç¡ç ãšçç£çãªé±æ«ã«ãšã£ãŠå¿ é ã§ãã
ããããML ã·ã¹ãã ãããè¿
éã«æ¹åããããè¿
éã«åžå Žã«æå
¥ããä¿¡é Œæ§ãé«ããŠã¹ãã¬ã¹ã軜æžããã®ã«åœ¹ç«ã€ããšãé¡ã£ãŠããŸãã
åºæïŒ habr.com