Habré ã§ã®æ©æ¢°åŠç¿ã³ã³ãã¹ãã®ããŒããç¶ããŠãèªè ã«ããã« XNUMX ã€ã®ãã©ãããã©ãŒã ã玹ä»ããããšæããŸãã 確ãã« kaggle ã»ã©å·šå€§ã§ã¯ãããŸããããééããªã泚ç®ã«å€ããŸãã
å人çã«ã¯ãããã€ãã®çç±ãã kaggle ã¯ããŸã奜ãã§ã¯ãããŸããã
- 第äžã«ãããã§ã®ç«¶æäŒã¯æ°ãæç¶ãããšãå€ããç©æ¥µçã«åå ããã«ã¯å€å€§ãªåªåãå¿ èŠã§ãã
- XNUMX çªç®ã¯ããããªã㯠ã«ãŒãã« (ãããªã㯠ãœãªã¥ãŒã·ã§ã³) ã§ãã Kaggle ã®ä¿¡å¥è ã¯ãããããå§äŸ¶ã®ãããªå·éãã§åœŒãã«æ¥ããããã«ã¢ããã€ã¹ããŠããŸãããå®éã«ã¯ãXNUMX ïœ XNUMX ãæãããŠåãçµãã§ããããšããçªç¶å šå¡ã®éã®å€§ç¿ã«äžŠã¹ãããããšã«ãªãã®ã¯éåžžã«æ®å¿µãªããšã§ãã
幞ããªããšã«ãæ©æ¢°åŠç¿ã³ã³ãã¹ãã¯ä»ã®ãã©ãããã©ãŒã ã§ãéå¬ãããŠããããããã®ã³ã³ãã¹ãã®ããã€ãã«ã€ããŠèª¬æããŸãã
å
¬çšèª: è±èªã äž»å¬è : Yandexããºãã«ãã³ã¯ãHSE |
å
¬çšèªãã·ã¢èªã äž»å¬è : Mail.ru ã°ã«ãŒã |
ãªã³ã©ã€ã³ã©ãŠã³ã: 15幎11æ2019æ¥ïœXNUMXæXNUMXæ¥ã ãªã³ãµã€ã決å: 4 幎 6 æ 2019 ïœ XNUMX æ¥ |
ãªã³ã©ã€ã³ - 7æ15æ¥ããXNUMXæXNUMXæ¥ãŸã§ã ãªãã©ã€ã³ - 30 æ 1 æ¥ãã XNUMX æ XNUMX æ¥ãŸã§ã |
倧åãããã³è¡çªåå éåšå
ã®ç²åã«é¢ããç¹å®ã®ããŒã¿ ã»ãã (è»éãéåéããã®ä»ã®ããªãè€éãªç©çãã©ã¡ãŒã¿ãŒ) ã䜿çšããŠãããããã¥ãŒãªã³ã§ãããã©ãããå€æããŸãã ãã®ã¹ããŒãã¡ã³ãããã2 ã€ã®ã¿ã¹ã¯ãç¹å®ãããŸããã â äºæž¬ãéä¿¡ããã ãã§æžããã®ã§ã¯ã - ãã XNUMX ã€ã¯ãäºæž¬çšã®å®å šãªã³ãŒããšã¢ãã«ã§ãããå®è¡ã«ã¯å®è¡æéãšã¡ã¢ãªäœ¿çšéã«é¢ããŠããªãå³ããå¶éã課ããããŸããã |
SNA ããã«ãœã³ ã³ã³ãã¹ãã§ã¯ã2018 幎 XNUMX æãã XNUMX æã«ãããŠãŠãŒã¶ãŒ ãã¥ãŒã¹ ãã£ãŒãå
ã®ãªãŒãã³ ã°ã«ãŒãããã®ã³ã³ãã³ã衚瀺ã®ãã°ãåéãããŸããã ãã®ãã¹ã ã»ããã«ã¯ãXNUMX æã®æåŸã® XNUMX é±éåãå«ãŸããŠããŸãã ãã°ã®åãšã³ããªã«ã¯ãäœã誰ã«è¡šç€ºããããããŸããã®ã³ã³ãã³ãã«å¯ŸãããŠãŒã¶ãŒã®åå¿ (è©äŸ¡ãã³ã¡ã³ããç¡èŠããã£ãŒãããã®é衚瀺ãªã©) ã«é¢ããæ
å ±ãå«ãŸããŠããŸãã SNA ããã«ãœã³ã®ã¿ã¹ã¯ã®æ¬è³ªã¯ããœãŒã·ã£ã« ãããã¯ãŒã¯ Odnoklassniki ã®åãŠãŒã¶ãŒã®ãã£ãŒããã©ã³ã¯ââä»ããããã¯ã©ã¹ããåãåãæçš¿ãã§ããã ãé«ãããããšã§ãã ãªã³ã©ã€ã³æ®µéã§ã¯ãã¿ã¹ã¯ã¯ 3 ã€ã®éšåã«åå²ãããŸããã 1. ããŸããŸãªå ±åäœæ¥ã®ç¹æ§ã«å¿ããŠæçš¿ãã©ã³ã¯ä»ããã 2. å«ãŸããç»åã«åºã¥ããŠæçš¿ãã©ã³ã¯ä»ããã 3. å«ãŸããããã¹ãã«åŸã£ãŠæçš¿ãã©ã³ã¯ä»ããã |
è€éãªã«ã¹ã¿ã ã¡ããªã¯ã¹ãROC-AUC ãªã© | ãŠãŒã¶ãŒå¥ã®å¹³å ROC-AUC |
第 XNUMX ã¹ããŒãžã®è³å - N äœåã® T ã·ã£ãã第 XNUMX ã¹ããŒãžãžã®ç§»è¡ïŒç«¶æäžã®å®¿æ³è²»ãšé£äºä»£ã¯æ¯æãããŸãïŒ ç¬¬äºæ®µé - ??? ïŒç§ã¯è«žäºæ ã«ããæè³åŒã«åºåžã§ãããæçµçã«è³åã®å 容ãç¥ãããšã¯ã§ããŸããã§ããïŒã 圌ãã¯åªåããŒã ã®ã¡ã³ããŒå šå¡ã«ã©ããããããäžãããšçŽæãã |
第 100 ã¹ããŒãžã®è³å - åªç§ãªåå è
3 åã«ã¯ T ã·ã£ããã¢ã¹ã¯ã¯ãŸã§ã®äº€éè²»ã競æäžã®å®¿æ³è²»ãé£äºä»£ãæ¯æããã第 1 ã¹ããŒãžãžã®åºå Žæš©ãäžããããŸãã ãŸããæåã®ã¹ããŒãžã®çµããã«åããŠãã¹ããŒãž 2080 ã® XNUMX ã€ã®ã¿ã¹ã¯ã§æãåªããè³åãçºè¡šãããŸãããå
šå¡ã RTX XNUMX TI ãã㪠ã«ãŒããç²åŸããŸããã 第 2 ã¹ããŒãžã¯ããŒã ã¹ããŒãžã§ãããŒã 㯠5 ïœ XNUMX 人ã§æ§æãããè³åã¯æ¬¡ã®ãšããã§ãã 1äœ - 300ã«ãŒãã« 2äœ - 200ã«ãŒãã« 3äœ - 100ã«ãŒã㫠審æ»å¡è³ - 100ã«ãŒãã« |
å ¬åŒé»å ±ã°ã«ãŒããåå è ã¯çŽ 190 人ãã³ãã¥ãã±ãŒã·ã§ã³ã¯è±èªã質åã¯åçãŸã§ã«æ°æ¥åŸ ããªããã°ãªããŸããã§ãã | é»å ±ã«ããå ¬åŒã°ã«ãŒããçŽ 1500 人ã®åå è ãåå è ãšäž»å¬è ã®éã§ã¿ã¹ã¯ã«ã€ããŠæŽ»çºãªè°è«ãè¡ããã |
äž»å¬è ã¯ãã·ã³ãã«ãªãœãªã¥ãŒã·ã§ã³ãšé«åºŠãªãœãªã¥ãŒã·ã§ã³ã® 16 ã€ã®åºæ¬çãªãœãªã¥ãŒã·ã§ã³ãæäŸããŸããã ã·ã³ãã«ã§ã¯ 16 GB æªæºã® RAM ãå¿ èŠã§ãã¢ããã³ã¹ã ã¡ã¢ãªã§ã¯ XNUMX GB ã«åãŸããŸããã§ããã åæã«ãå°ãå ãèŠæ®ãããšãåå è ã¯å é²çãªãœãªã¥ãŒã·ã§ã³ãå€§å¹ ã«äžåãããã©ãŒãã³ã¹ãçºæ®ããããšã¯ã§ããŸããã§ããã ãããã®ãœãªã¥ãŒã·ã§ã³ã®ç«ã¡äžãã«ã¯äœã®å°é£ããããŸããã§ããã é«åºŠãªäŸã«ã¯ããœãªã¥ãŒã·ã§ã³ã®æ¹åãã©ãããå§ããã¹ããã«ã€ããŠã®ãã³ããèšèŒãããã³ã¡ã³ããããããšã«æ³šæããŠãã ããã | åã¿ã¹ã¯ã«å¯ŸããŠåºæ¬çãªåå§çãªè§£æ±ºçãæäŸãããŸããããåå è ã¯ããããç°¡åã«è§£æ±ºã§ããŸããã ã³ã³ãã¹ãã®åæ段éã§ã¯ãåå è ã¯ããã€ãã®åé¡ã«çŽé¢ããŸããããŸããããŒã¿ã Apache Parquet 圢åŒã§æäŸãããŠãããPython ãš Parquet ããã±ãŒãžã®ãã¹ãŠã®çµã¿åããããšã©ãŒãªãã§åäœããããã§ã¯ãããŸããã§ããã XNUMX çªç®ã®å°é£ã¯ã¡ãŒã« ã¯ã©ãŠãããåçãããŠã³ããŒãããããšã§ããããçŸæç¹ã§ã¯å€§éã®ããŒã¿ãäžåºŠã«ããŠã³ããŒãããç°¡åãªæ¹æ³ã¯ãããŸããã çµæãšããŠããããã®åé¡ã«ããåå è ã¯æ°æ¥éé ããŸããã |
ã€ããªã 第äžæ®µé
ãã®èª²é¡ã¯ããã¥ãŒç²åãšéãã¥ãŒç²åããã®ç¹æ§ã«å¿ããŠåé¡ããããšã§ããã ãã®ã¿ã¹ã¯ã®éèŠãªç¹åŸŽã¯ããã¬ãŒãã³ã° ããŒã¿ã«éã¿åãååšããããšã§ãããäž»å¬è èªèº«ããã®è¡ã®çãã«å¯Ÿããä¿¡é ŒåºŠãšããŠè§£éããŸããã åé¡ã¯ãããªãã®æ°ã®è¡ã«è² ã®éã¿ãå«ãŸããŠããããšã§ãã
ãã³ãã®ããç·ã«ã€ããŠæ°åéèã (ãã³ãã¯éã¿åã®ãã®æ©èœã«æ³šæãåŒããã ãã§ã)ããã®ã°ã©ããäœæããåŸã3 ã€ã®ãªãã·ã§ã³ããã§ãã¯ããããšã«ããŸããã
1) è² ã®ãŠã§ã€ã(ããã³ããã«å¿ãããŠã§ã€ã)ãæã€ã©ã€ã³ã®ã¿ãŒã²ãããå転ããŸãã
2) éã¿ã 0 ããå§ãŸãããã«æå°å€ã«ã·ããããŸãã
3) æååã®éã¿ã䜿çšããªã
1 çªç®ã®ãªãã·ã§ã³ã¯ææªã§ããããšãå€æããŸããããæåã® XNUMX ã€ã¯çµæãæ¹åããæãè¯ãã£ãã®ã¯ãªãã·ã§ã³ XNUMX ã§ãããã«æåã®ã¿ã¹ã¯ã§çŸåšã® XNUMX äœãXNUMX çªç®ã®ã¿ã¹ã¯ã§ XNUMX äœã«ãªããŸããã
次ã®ã¹ãããã¯ãæ¬ æå€ããªããããŒã¿ã確èªããããšã§ããã äž»å¬è
ã¯ãã§ã«ç²Ÿæ»æžã¿ã®ããŒã¿ãæäŸããŠãããŸããããããªãã®æ°ã®æ¬ æå€ããããããã㯠-9999 ã«çœ®ãæããããŸããã
MatchedHit_{X,Y,Z}[N] åãš MatchedHit_D{X,Y,Z}[N] åã§æ¬ æå€ãèŠã€ãããŸãããããã㯠N=2 ãŸã㯠3 ã®å Žåã®ã¿ã§ãããç解ããŠããããã«ãäžéšã®ç²åã¯æ¬ èœããŠããŸããã§ããã 4 ã€ã®æ€åºåšããã¹ãŠééãã3 çªç®ãŸã㯠4 çªç®ã®ãã¬ãŒãã§åæ¢ããŸããã ããŒã¿ã«ã¯ Lextra_{X,Y}[N] åãå«ãŸããŠãããæããã« MatchedHit_{X,Y,Z}[N] ãšåãããšãèšè¿°ããŠããŸãããããçš®ã®å€æ¿ã䜿çšããŠããŸãã ãããã®ããããªæšæž¬ã¯ãLextra_{X,Y}[N] ã MatchedHit_{X,Y,Z}[N] ã®æ¬ æå€ã眮ãæããããšãã§ããããšã瀺åããŸãã (X 座æšãš Y 座æšã®ã¿)ã MatchedHit_Z[N] ã¯äžå€®å€ã§ååã«åããããŸããã ãããã®æäœã«ãããäž¡æ¹ã®ã¿ã¹ã¯ã§äžéã® 1 äœã«å°éããããšãã§ããŸããã
æåã®ã¹ããŒãžã§åªåããŠãäœãäžããããªãã£ãããšãèãããšãããã§æ¢ããŠãããã£ããããããŸããããç§ãã¡ã¯ç¶ããŠãçŸããçµµãæããæ°ããæ©èœãèãåºããŸããã
ããšãã°ãç²åãš 5 ã€ã®æ€åºãã¬ãŒãã®ãããããšã®äº€ç¹ããããããããšãåãã¬ãŒãäžã®ç¹ãã¢ã¹ãã¯ãæ¯ 4 察 5 ã® 0,0 ã€ã®é·æ¹åœ¢ã«ã°ã«ãŒãåãããäžå¿ã XNUMX 察 XNUMX ã§ããããšãããããŸããç¹ (XNUMX)ãããã³ in æåã®é·æ¹åœ¢ã«ã¯ç¹ããããŸããã
ãã¬ãŒãNo./é·æ¹åœ¢å¯žæ³ | 1 | 2 | 3 | 4 | 5 |
---|---|---|---|---|---|
ãã¬ãŒã1 | 500h625 | 1000h1250 | 2000h2500 | 4000h5000 | 8000h10000 |
ãã¬ãŒã2 | 520h650 | 1040h1300 | 2080h2600 | 4160h5200 | 8320h10400 |
ãã¬ãŒã3 | 560h700 | 1120h1400 | 2240h2800 | 4480h5600 | 8960h11200 |
ãã¬ãŒã4 | 600h750 | 1200h1500 | 2400h3000 | 4800h6000 | 9600h12000 |
ãããã®å¯žæ³ã決å®ããåŸãç²åããšã« 4 ã€ã®æ°ããã«ããŽãªç¹åŸŽ (ç²åãåãã¬ãŒããšäº€å·®ããé·æ¹åœ¢ã®æ°) ãè¿œå ããŸããã
ãŸããç²åãäžå¿ããåŽé¢ã«æ£ä¹±ããŠããããã«èŠããããšã«ãæ°ã¥ãããã®æ£ä¹±ã®ã質ããäœããã®æ¹æ³ã§è©äŸ¡ãããšããã¢ã€ãã¢ãçãŸããŸããã çæ³çã«ã¯ãé¢éžç¹ã«å¿ããŠäœããã®ãçæ³çãªãæŸç©ç·ãèãåºããããããã®åå·®ãæšå®ããããšãå¯èœã§ãããããç§ãã¡ã¯ãçæ³çãªãçŽç·ã«éå®ããŸããã ãã®ãããªçæ³çãªçŽç·ãåäŸµå ¥ç¹ã«å¯ŸããŠæ§ç¯ããã®ã§ããã®çŽç·ããåç²åã®è»éã®æšæºåå·®ãèšç®ããããšãã§ããŸããã target = 1 ã®å¹³åå差㯠152ãtarget = 0 ã®å¹³åå差㯠390 ã§ãã£ãããããã®æ©èœã¯æ«å®çã«è¯å¥œã§ãããšè©äŸ¡ããŸããã ãããŠå®éããã®æ©èœã¯ããã«æã䟿å©ãªæ©èœã®ãããã«ã©ã³ã¯ã€ã³ããŸããã
ç§ãã¡ã¯åãã§ãè¿œå ã® 4 ã€ã®ç¹åŸŽãšããŠãåç²åã® 4 ã€ã®äº€ç¹ãã¹ãŠã®çæ³çãªçŽç·ããã®åå·®ãè¿œå ããŸãã (ãããŠãããããããŸãæ©èœããŸãã)ã
äž»å¬è ããæäŸããããã³ã³ãã¹ãã®ããŒãã«é¢ããç§åŠè«æãžã®ãªã³ã¯ãããç§ãã¡ã¯ãã®åé¡ãæåã«è§£æ±ºããããã§ã¯ãªããããããäœããã®ç¹æ®ãªãœãããŠã§ã¢ãããã®ã§ã¯ãªãããšããèããçãŸããŸããã IsMuonSimpleãIsMuonãIsMuonLoose ã¡ãœãããå®è£ ãããŠãããªããžããªã github ã§çºèŠããã®ã§ãããããå°ãå€æŽããŠãµã€ãã«è»¢éããŸããã æ¹æ³èªäœã¯éåžžã«åçŽã§ãããšãã°ããšãã«ã®ãŒãç¹å®ã®ãããå€æªæºã®å Žåãããã¯ãã¥ãŒãªã³ã§ã¯ãªããããã§ãªãå Žåã¯ãã¥ãŒãªã³ã§ãã ãã®ãããªåçŽãªæ©èœã§ã¯ãæããã«åŸé ããŒã¹ãã£ã³ã°ã䜿çšããå Žåã«å¢å ãããããããšãã§ããªãããããããå€ã«å¥ã®éèŠãªãè·é¢ããè¿œå ããŸããã ãããã®æ©èœãè¥å¹²æ¹åãããŠããŸãã ãããããæ¢åã®ææ³ããã培åºçã«åæããããšã§ããã匷åãªææ³ãèŠã€ããŠæšèã«è¿œå ããããšãã§ããã§ãããã
ã³ã³ãã¹ãã®çµããã«ãXNUMX çªç®ã®åé¡ã«å¯Ÿãããç°¡åãªã解決çãå°ã調æŽããŸããããæçµçã«ã¯æ¬¡ã®ç¹ã§ããŒã¹ã©ã€ã³ãšã¯ç°ãªããŸããã
- è² ã®éã¿ãæã€è¡ã§ã¯ãã¿ãŒã²ãããå転ãããŸãã
- MatchedHit_{X,Y,Z}[N] ã®æ¬ æå€ãåããŸãã
- æ·±ãã 7 ã«åæž
- åŠç¿çã 0.1 ã«äœäž (以å㯠0.19)
ãã®çµæãããã«å€ãã®æ©èœãè©Šã (ããŸãæåããŸããã§ããã)ããã©ã¡ãŒã¿ãŒãéžæã㊠catboostãlightgbmãxgboost ããã¬ãŒãã³ã°ããããŸããŸãªäºæž¬ã®çµã¿åãããè©Šãããã©ã€ããŒããéãåã« XNUMX çªç®ã®ã¿ã¹ã¯ã§èªä¿¡ãæã£ãŠåå©ããæåã®ã¿ã¹ã¯ã§ã¯ãªãŒããŒãã¡ã
ãã©ã€ããŒããéããåŸãæåã®ã¿ã¹ã¯ã§ã¯10äœã1çªç®ã®ã¿ã¹ã¯ã§ã¯3äœã§ããã ãªãŒããŒå šå¡ãå ¥ãä¹±ãããªããŒãããããã©ã€ããŒãã®æ¹ãã¹ããŒããéãã£ãã ããŒã¿ã®éå±€åãäžååã ã£ããã㧠(ãŸãã¯ãããšãã°ããã©ã€ããŒãã«è² ã®éã¿ãæã€è¡ããªãã£ã)ãããã¯å°ãã€ã©ã€ã©ããŸããã
SNA ããã«ãœã³ 2019 - ããã¹ãã 第äžæ®µé
ãã®ã¿ã¹ã¯ã¯ãOdnoklassniki ãœãŒã·ã£ã« ãããã¯ãŒã¯äžã®ãŠãŒã¶ãŒæçš¿ããããã«å«ãŸããããã¹ãã«åºã¥ããŠã©ã³ã¯ä»ãããããšã§ãããããã¹ãã«å ããŠãæçš¿ã«ã¯ããã«ããã€ãã®ç¹åŸŽ (èšèªãææè ãäœææ¥æãé²èŠ§æ¥æ) ããããŸããã ïŒã
ããã¹ããæäœããããã®å€å žçãªã¢ãããŒããšããŠã次㮠XNUMX ã€ã®ãªãã·ã§ã³ãåãäžããŸãã
- é¡äŒŒããåèªãé¡äŒŒãããã¯ãã«ãæã€ããã«ãååèªã n 次å
ã®ãã¯ãã«ç©ºéã«ãããã³ã°ããŸã (詳现ã¯ã
ç§ãã¡ã®èšäº )ã次ã«ããã¹ãã®å¹³ååèªãèŠã€ããããåèªã®çžå¯Ÿäœçœ®ãèæ ®ããã¡ã«ããºã (CNNãLSTM/GRU) ã䜿çšããŸãã - æå šäœãããã«æäœã§ããã¢ãã«ã䜿çšããŸãã ããšãã°ãããŒãã çè«çã«ã¯ããã®ã¢ãããŒãã®æ¹ãããŸãæ©èœããã¯ãã§ãã
ããã¹ãã¯åããŠãªã®ã§ã人ã«æããã®ã¯éãã®ã§ãèªåã§æããŸãã ãããã¯ãã³ã³ãã¹ãã®éå§æã«ç§ãèªåèªèº«ã«äžãããã³ãã§ãã
- äœããæããåã«ãããŒã¿ãèŠãŠãã ããã ããã¹ãèªäœã«å ããŠãããŒã¿ã«ã¯ããã€ãã®åããããããããç§ããã£ããããã¯ããã«å€ãã®ããšãçµãåºãããšãã§ããŸããã æãç°¡åãªæ¹æ³ã¯ãäžéšã®åã«å¯ŸããŠå¹³åã¿ãŒã²ãã ãšã³ã³ãŒãã£ã³ã°ãè¡ãããšã§ãã
- ãã¹ãŠã®ããŒã¿ããåŠã°ãªãã§ãã ããã 倧éã®ããŒã¿ (çŽ 17 äžè¡) ããããŸãããã仮説ãæ€èšŒããããã«ãã¹ãŠã䜿çšããå¿ èŠã¯ãŸã£ãããããŸããã§ããã ãã¬ãŒãã³ã°ãšååŠçã¯éåžžã«é ãã£ãã®ã§ããã£ãšèå³æ·±ã仮説ããã¹ãããæéããã£ãããšã¯æããã§ãã
- <ç©è°ãéžãã¢ããã€ã¹> ãã©ãŒã¢ãã«ãæ¢ãå¿ èŠã¯ãããŸããã ç§ã¯ Elmo ãš Bert ãç解ããã®ã«é·ãæéãè²»ããã圌ããç§ãããã«é«ããšããã«é£ããŠè¡ã£ãŠãããããšãæã¿ãŸããããã®çµæããã·ã¢èªçšã«äºåãã¬ãŒãã³ã°ããã FastText åã蟌ã¿ã䜿çšããŸããã Elmo ã§ã¯ãã以äžã®é床ãå®çŸã§ããŸããã§ããããBert ã§ã¯ãããç解ããæéããŸã ãããŸããã§ããã
- <ç©è°ãéžãã¢ããã€ã¹> ãã©ãŒæ©èœã 1 ã€æ¢ãå¿ èŠã¯ãããŸããã ããŒã¿ãèŠããšãããã¹ãã®çŽ XNUMX% ã«å®éã«ã¯ããã¹ããå«ãŸããŠããªãããšãããããŸããã ããããããã€ãã®ãªãœãŒã¹ãžã®ãªã³ã¯ããã£ãã®ã§ããµã€ããéããŠã¿ã€ãã«ãšèª¬æãåãåºãç°¡åãªããŒãµãŒãäœæããŸããã ããã¯è¯ãã¢ã€ãã¢ã®ããã«æããŸãããããã®åŸã調åã«ä¹ã£ãŠãã¹ãŠã®ããã¹ãã®ãã¹ãŠã®ãªã³ã¯ã解æããããšã«ããåã³å€ãã®æéããã¹ããŠããŸããŸããã ãããã¯ãã¹ãŠãæçµçµæã«å€§ããªæ¹åããããããŸããã§ãã (ããšãã°ãã¹ããã³ã°ãèãåºããŸããã)ã
- å€å žçãªæ©èœãåäœããŸãã ããšãã°ãç§ãã¡ã¯ Google ã§ãããã¹ãæ©èœ kaggleããæ€çŽ¢ãããã¹ãŠãèªãã§è¿œå ããŸãã TF-IDF ã¯ãããã¹ãã®é·ããåèªãå¥èªç¹ã®éãªã©ã®çµ±èšçç¹åŸŽãšåæ§ã«æ¹åããããããŸããã
- DateTime åãããå Žåã¯ãããããããã€ãã®åå¥ã®æ©èœ (æéãææ¥ãªã©) ã«è§£æãã䟡å€ããããŸãã ã©ã®æ©èœã匷調ãã¹ããã¯ãã°ã©ããããã€ãã®ææšã䜿çšããŠåæããå¿ èŠããããŸãã ããã§ã¯ãæ°ãŸããã«ãã¹ãŠãæ£ããå®è¡ããå¿ èŠãªæ©èœã匷調衚瀺ããŸããããéåžžã®åæã§ãåé¡ã¯ãããŸããã§ããïŒããšãã°ã決åã§è¡ã£ãããã«ïŒã
ã³ã³ãã¹ãã®çµæãç§ã¯åèªç³ã¿èŸŒã¿ã§ 7 ã€ã® keras ã¢ãã«ããã¬ãŒãã³ã°ãããã 76 ã€ã¯ LSTM ãš GRU ã«åºã¥ããŠãã¬ãŒãã³ã°ããŸããã ã©ã¡ããããã·ã¢èªçšã«äºåãã¬ãŒãã³ã°ããã FastText åã蟌ã¿ã䜿çšããŸãã (ä»ã®åã蟌ã¿ãããã€ãè©ŠããŸããããããããæãå¹æçã§ãã)ã äºæ³ãå¹³åããçµæãåå è XNUMXåäžæçµXNUMXäœã§ããã
第äžæ®µéçµäºåŸãå
¬éãããŸãã
ã»ã«ã³ãã¹ããŒãžOKïŒIDAO
倧äŒã®ç¬¬XNUMXã¹ããŒãžã¯ã»ãŒé£ç¶ããŠè¡ãããã®ã§ããŸãšããŠèŠãããšã«ããã
ãŸããç§ãšæ°ããè²·åããããŒã ã¯ãMail.ru 瀟ã®å°è±¡çãªãªãã£ã¹ã«å°çããŸãããããã§ç§ãã¡ã®ä»äºã¯ãããã¹ããåçãã³ã©ãã¬ãŒã·ã§ã³ãšãã第 2 段éã® 5 ã€ã®ãã©ãã¯ã®ã¢ãã«ãçµã¿åãããããšã§ããã ãã®ããã« XNUMX æ¥åŒ·ãå²ãåœãŠãããŸããããããã¯éåžžã«çãããšãå€æããŸããã å®éãå䜵ã«ããå©çã¯åŸãããã第 XNUMX 段éã®çµæãåçŸããããšããã§ããŸããã§ããã æçµçã«ã¯XNUMXäœã«ãªããŸããããããã¹ãã¢ãã«ã¯äœ¿ããŸããã§ããã ä»ã®åå è ã®è§£æ±ºçãèŠãåŸãããã¹ããã¯ã©ã¹ã¿ãªã³ã°ããŠã³ã©ãã¬ãŒã·ã§ã³ ã¢ãã«ã«è¿œå ããŠã¿ã䟡å€ããããšæãããŸããã ãã®ã¹ããŒãžã®å¯äœçšãšããŠãæ°ããªå°è±¡ãã¯ãŒã«ãªåå è ãäž»å¬è ãšã®åºäŒããã³ãã¥ãã±ãŒã·ã§ã³ããããŠæ·±å»ãªç¡ç äžè¶³ãããããããIDAOã®æçµã¹ããŒãžã®çµæã«åœ±é¿ãäžããå¯èœæ§ããããŸãã
IDAO 2019 æçµã¹ããŒãžã®èª²é¡ã¯ã空枯ã§ã® Yandex ã¿ã¯ã·ãŒé転æã®æ³šæåŸ ã¡æéãäºæž¬ããããšã§ããã ã¹ããŒãž 2 ã§ã¯ã3 ã€ã®ã¿ã¹ã¯ = 3 ã€ã®ç©ºæž¯ãç¹å®ãããŸããã 空枯ããšã«ã2 ãæéã®ã¿ã¯ã·ãŒæ³šææ°ã«é¢ããååäœã®ããŒã¿ãæäŸãããŸãã ãããŠãã¹ãããŒã¿ãšããŠãéå»1,5é±éã®ç¿æåãšååäœã®æ³šæããŒã¿ãäžããããŸããã æéãã»ãšãã©ãªã (XNUMX æ¥)ãã¿ã¹ã¯ã¯éåžžã«å ·äœçã§ãããŒã ããã³ã³ãã¹ãã«åå ããã®ã¯ XNUMX 人ã ãã§ããããã®çµæãçµããã«åããŠæ²ããçµæãšãªããŸããã èå³æ·±ãã¢ã€ãã¢ã«ã¯ã倩æ°ã亀éæžæ»ãYandex ã¿ã¯ã·ãŒã®æ³šæçµ±èšãªã©ã®å€éšããŒã¿ã䜿çšããè©Šã¿ãå«ãŸããŠããŸããã äž»å¬è ã¯ãããã®ç©ºæž¯ãäœãªã®ããæããã«ããªãã£ãããå€ãã®åå è ã¯ã·ã§ã¬ã¡ãŒããšãŽã©ããã¢ãžã§ããŽã©ããŽããŒã³ãŽã©ã®ããšã ãšæã£ãŠããã ãã®ä»®å®ã¯å€§äŒåŸã«åé§ãããŸããããããšãã°ã¢ã¹ã¯ã¯ã®æ°è±¡ããŒã¿ããã®ç¹åŸŽã«ãããæ€èšŒãšãªãŒããŒããŒãã®äž¡æ¹ã§çµæãåäžããŸããã
ãŸãšã
- MLã®å€§äŒã¯ã¯ãŒã«ã§é¢çœãã§ããïŒ ããã§ã¯ãããŒã¿åæã®ã¹ãã«ããå·§åŠãªã¢ãã«ããã¯ããã¯ã掻çšããããšãã§ããŸããåžžèçãªããšã¯æè¿ãããŸãã
- ML ã¯ãã§ã«èšå€§ãªç¥èäœç³»ã§ãããææ°é¢æ°çã«å¢å ããŠããããã§ãã ç§ã¯ããŸããŸãªåé (ä¿¡å·ãåçãè¡šãããã¹ã) ã«ç²Ÿéãããšããç®æšãèªåã«èšå®ããŸããããå匷ããããšãã©ãã»ã©ãããã¯ãã§ã«èªèããŠããŸããã ããšãã°ããããã®ã³ã³ãã¹ãã®åŸãç§ã¯ã¯ã©ã¹ã¿ãªã³ã° ã¢ã«ãŽãªãºã ãåŸé ããŒã¹ãã£ã³ã° ã©ã€ãã©ãªãæäœããããã®é«åºŠãªãã¯ãã㯠(ç¹ã«ãGPU äžã® CatBoost ãæäœãã)ãã«ãã»ã« ãããã¯ãŒã¯ãã¯ãšãª ããŒå€ã®æ³šæã¡ã«ããºã ãªã©ãç 究ããããšã«ããŸããã
- kaggleã ãã§ã¯ã ãïŒ å°ãªããšã T ã·ã£ããç²åŸããã®ãç°¡åãªã³ã³ãã¹ãã¯ä»ã«ããããããããä»ã®è³åãç²åŸãããã£ã³ã¹ãå€ããããŸãã
- éä¿¡ããïŒ æ©æ¢°åŠç¿ãšããŒã¿åæã®åéã«ã¯ãã§ã«å€§èŠæš¡ãªã³ãã¥ããã£ããããTelegramãSlack ã«ã¯ããŒãå¥ã°ã«ãŒãããããMail.ruãYandexããã®ä»ã®äŒæ¥ã®çå£ãªäººã ã質åã«çããåå¿è ããã®åéã§éãæ©ã¿ç¶ãã人ãã¡ãæ¯æŽããŸããç¥èã®ã
- åã®ãã€ã³ãã«è§Šçºããããã¹ãŠã®äººã«ã蚪åããããšããå§ãããŸã
ããŒã¿ãã§ã¹ã â 10æ11æ¥ãšXNUMXæ¥ã«ã¢ã¹ã¯ã¯ã§éå¬ããã倧èŠæš¡ãªç¡æã«ã³ãã¡ã¬ã³ã¹ã
åºæïŒ habr.com