2019-01-01ãã1å¹´éã®è¨äºä¸è¦§
ã¯ããã« ããæ¥ã¯ãã¹éè¨è¡¨ã¨ãã¤ã¼ããã¤ãºãçºãã¦ãã¦ã ãããããããã¯ã©ãããé¢ä¿æ§ã ã£ãï¼ã¨æã£ãã®ã§ãã®æèã¡ã¢ã§ãã è©¦è¡ ä»¥ä¸ã®ãããªæå±ã¯ã©ã¹Ãæ§å¥Ãå©ãæã®ï¼è»¸ã®ã¯ãã¹éè¨è¡¨ãããã¨ãã ãããç¨ãã¦æ§å¥ã¨å©ãæãããã£ã¦ãã人â¦
ã¯ãã㫠以åã«lightgbmã¯å ¥åã«æ¬ æå¤ããã£ã¦ããã¾ãå¦ç¿ãã¦ãããã¨ããè¨äºãæ¸ããã®ã§ããã ããã¯å¦ç¿æã«æ¬ æãåå¨ãã¦ããå ´åã®è©±ã§ããã ç¾å®ã®åé¡ãèããã¨ãå¦ç¿æã¨ãã®ã¢ãã«ã使ã£ãæ¨è«æã§ã¯æç³»åã®éããç°å¢å¤åã®å½±é¿ ã«ããâ¦
ã¯ããã« ãããããã¹ã«ããã¦ç²ããåæã«ä¸ããå½±é¿ãå®éåãã¦ã¿ãã ã¨ããé¡ç®ã§ä¸è¨ã¤ãã³ãã§LTãã¦ãã¾ããã spoana.connpass.com 使ç¨ããè³æã¯ä»¥ä¸ã«ãªãã¾ãã ããããã¹ã«ããã¦ç²ããåæã«ä¸ããå½±é¿ãå®éåãã¦ã¿ã - Speaker Deck å â¦
ã¯ããã« couseraã§Bayesian Methods for Machine Learningãåãã¦ãããã§ããã ãã®ä¸ã§åºã¦ããVAE(Variational Auto Encoder)ã§å°ã試ãã¦ã¿ãããã¨ããã£ãã®ã§ãã®å®é¨è¨é²ã«ãªãã¾ãã åèæç® VAEã®çè«ãä½ç½®ä»ãã«ã¤ãã¦ã¯ä¸è¨è³æãåèã«ãªâ¦
ã¯ããã« ãã¼ã¿ã®å¤æ°éã®é¢ä¿æ§ãæããã«ãããã¨ããã±ã¼ã¹ã¯å¤ãã¨æãã¾ãã ãã®å ´åç¸é¢ãæ£å¸å³ãã¿ãã®ãä¸è¬çã§ããã交絡ããã¤ãºãå¤ãã±ã¼ã¹ãªã© ããã ãã§ã¯ä¸ååãªå ´åãããããããã®å ´åã«ã対å¿ã§ããããªææ³ã試ãã¦ã¿ã¾ãã 試ãâ¦
â»ãã®è¨äºã§ä½¿ç¨ãã¦ããå¤éä»£å ¥æ³ã®ããã±ã¼ã¸ã¯æ£å¼ãªå¤éä»£å ¥æ³ã®æ çµã¿ã¨ã¯ç°ãªãã¾ãã®ã§ã注æé¡ãã¾ãã ã¯ããã« æè¿å¤éä»£å ¥æ³ã¨ããæ¬ æå¤è£å®ã®ææ³ããããã¨ãç¥ãã¾ããã çµ±è¨å¦ã®çéã§ã¯æ¬ æå¤è£å®ã¯å¤éä»£å ¥æ³ã使ç¨ããã®ããã¿ã¼ã®ãâ¦
ã¯ããã« æ¥ã ã®çæ´»ãå¶ãä¸ã§ãã©ã³ããã©ãã§é£ã¹ããã¨ãããã¨ã¯é常ã«éè¦ãªåé¡ã§ã(2åç®)ã ååã¯ãã®åé¡ãæé©åæ¢åé¡ã¨æãã¦æ¢ç´¢ã¨æ´»ç¨ã®æé©ç¹æ¢ç´¢ãè¡ãã¾ããããæé©åæ¢åé¡ã§ã¯ä¸åº¦è¦éããåºã¯ããé¸æã§ããªãã¨ããå¶ç´ãããã¾ããâ¦
ã¯ããã« æ¥ã ã®çæ´»ãå¶ãä¸ã§ãã©ã³ããã©ãã§é£ã¹ããã¨ãããã¨ã¯é常ã«éè¦ãªåé¡ã§ãã ã©ã³ãã®é¸æè¢ã¨ãã¦ã¯ãæ°ããåºã«å ¥ãï¼æ¢ç´¢ï¼ã¨ãä»ã¾ã§è¡ã£ããã¨ã®ããåºã§è¯ãã£ãåºã«å ¥ãï¼æ´»ç¨ï¼ã®ã©ã¡ãããè¡ãå¿ è¦ãããã¾ããçµé¨çã«è¯ãã£ãåºâ¦
ã¯ããã« ä»åã¯CausalImpactã«ã¤ãã¦æ¸ãã¦ããããã¨æãã¾ãã CausalImpactã¯google製ã®å¹æ測å®ç¨ããã±ã¼ã¸ã§ã主ã«åºåããã£ã³ãã¼ã³ã®å¹æã測å®ããã®ã«ç¨ãããã¾ãã ãªãåºåããã£ã³ãã¼ã³ã®å¹æã測å®ããã®ã«ãããããã®ãå¿ è¦ãªã®ãã¨ããâ¦
ã¯ããã« æåã¯æè¿è©±é¡(ï¼)ã®CausalImpactã«ã¤ãã¦æ¸ããã¨æã£ã¦ããã®ã§ããããã®åºç¤ã¨ãªãå·®åã®å·®åæ³ï¼difference in differenceï¼ã«ã¤ãã¦ã®ç¥èãä¸è¶³ãã¦ããããããã®è¨äºã§ã¯å·®åã®å·®åæ³ã«ã¤ãã¦è©¦ãããã¨ãè¨è¿°ãã¦ããã¾ãã å·®åã®å·®åâ¦
ã¯ããã« ä¸è¨ã®NCAAã³ã³ãã«åå ããåçæã§ãã https://www.kaggle.com/c/womens-machine-learning-competition-2019 https://www.kaggle.com/c/mens-machine-learning-competition-2019 ãã®ã³ã³ãã®ã¿ã¹ã¯ã¯å ¨ç±³å¤§å¦ä½è²åä¼ï¼NCAAï¼ãæ¯å¹´æ¥ã«ä¸»å¬ãâ¦
ã¯ããã« æç³»å解æã«ã¤ãã¦ã¯ä»¥åã«MCMCãç¨ããç¶æ 空éã¢ãã«ã®æ¨å®ãè¡ãªã£ãã®ã§ãããç¶æ 空éã¢ãã«ã®æ¨å®æ¹æ³ã¨ãã¦ã«ã«ãã³ãã£ã«ã¿ãç¥ã£ã¦ãããæ¹ãè¯ãããã ã£ãã®ã§ãä»åã¯ã«ã«ãã³ãã£ã«ã¿ã®å®è£ ãè¡ãªã£ã¦ããã¾ãã ç¶æ 空éã¢ãã«ã®æ¨â¦
ã¯ããã« RCTã使ããªãå ´åã®å ææ¨è«ã®ææ³ã¨ãã¦å¾åã¹ã³ã¢ã使ãæ¹æ³ãããã¾ãã å¾åã¹ã³ã¢ã®ç®åºã¯ãã¸ã¹ãã£ãã¯å帰ãç¨ããã®ãä¸è¬çã§ããããã®é¨åã¯å¥ã«lightgbmã¨ãæ©æ¢°å¦ç¿çãªææ³ã§ãã£ã¦ãããã®ã§ã¯ï¼ã¨æã£ã¦ãã¾ãããæ¢ã«ãã£ã¦ããâ¦
ã¯ããã« æ©æ¢°å¦ç¿ãç¾å®ã®åé¡ã«é©ç¨ããå ´åããã®ã¢ãã«ã«èª¬ææ§ãæ±ãããããã¨ãå°ãªãããåå¨ããã¨æãã¾ãã ãã®å ´åã精度ãç ç²ã«ãã¦ç·å½¢å帰ãå®æ½ããã§ããããï¼æ¨ç³»ã¢ãã«ã®éè¦åº¦ãé å¼µã£ã¦èª¬æããã§ããããï¼ããã¨ãSHAPãLIMEãªã©â¦
ã¯ãã㫠以åã«ãAtCoderã®åé¡é£æ度ã®æ¨å®ãé ç®åå¿çè«ãç¨ãã¦è¡ãã¾ããã ããã¯çµ±è¨ã¢ããªã³ã°ã使ã£ãæ¹æ³ã ã£ãã®ã§ãããåæ§ã®ãã¨ãæ©æ¢°å¦ç¿ã§ã§ããªããã¨ããªããªã ã¨æãä»åã¯æ©æ¢°å¦ç¿ãç¨ãã¦åé¡é£æ度ã®æ¨å®ãè¡ããã¨ã«ãã¾ããã æâ¦
ã¯ãã㫠趣å³ã§ãã¹ãçè«å¨ããããããæ¼ã£ã¦ãã¦ãæ½å¨ã©ã³ã¯çè«ã¨ãããã®ãè¦ã¤ãã¾ããã æ½å¨ã©ã³ã¯çè«ã®ç¹å¾´ã¯ãä¾ãã°ãã¹ãã®å¾ç¹ã§ã¯ï¼ç¹å»ã¿ã®é£ç¶çãªè½åè©ä¾¡ãè¡ãã®ã«å¯¾ãã¦ã 段éçã§é åºçãªè½åè©ä¾¡ãè¡ãç¹ã§ãã ãã¹ãã«ããã1ç¹â¦
ã¯ããã« AtCoderã¯ç«¶æããã°ã©ãã³ã°ã®ãµã¤ãã§ããã»ã¼æ¯é±ã®ããã«ã³ã³ãã¹ããéå¬ãããåå è ãè¤æ°ã®åé¡ã解ãã解ããåé¡æ°ã¨ãã®æ©ãããé ä½ãã¤ãã¾ããã¾ãååå è ã¯ã¬ã¼ãã£ã³ã°ãæã£ã¦ãã¦ããã®ã¬ã¼ãã£ã³ã°ãé ä½ã«ãã£ã¦å¤åããã¨ãâ¦
ã¯ããã« åå ã¯ãã¼ãéå°ãã¢ã½ã³ãç¨ãã¦CMè¦è´ãã²ã¼ã ã®ãã¬ã¤æéã«ä¸ããå¹æã®æ¸¬å®ãè¡ãã¾ããã ã²ã¼ã ã®ãã¬ã¤ã®æç¡ã«ä¸ããå¹æã®é¨åã¯ãã¾ãæ¨å®ã§ãã¾ãããã ã²ã¼ã ããã¬ã¤ããæéã«ä¸ããå¹æã¯ãã¢ã½ã³åå¸ã§ã¯ãã¾ãæ¨å®ãããã¨ãâ¦
ã¯ããã« æè¿ã³ã³ã¤ãã§ã¹ãã¼ã³ãããããã£ã³ãã¼ã³ããã£ã¦ãã¾ããã ãã¤ãºããã£ã¦ããã¨ãããããã¨ãã«ææã¡ã®æ å ±ããå½é¸ç¢ºçãã©ãã¾ã§æ¨å®ã§ãããæ°ã«ãªãã¾ããã ã¨ãããã¨ã§ã5åã³ã³ã¤ãã«è¡ã£ã¦2ã¤ã®ã¹ãã¼ã³ãã²ãããã¦ãã¾ããã â¦
ã¯ããã« æè¿ã¯çµ±è¨ã¢ããªã³ã°ã«ã¤ãã¦å¦ãã§ãã¾ãããå ·ä½çã«ããããã±ã¼ã¹ã§ã¯çµ±è¨ã¢ãã«ã使ãã¹ãã ã¨ããã±ã¼ã¹ãèªåã®ä¸ã§å®ã¾ã£ã¦ãã¾ããããã®ããããã¤ããããã«å®é¨ãè¡ãªã£ã¦ããããã¨æãã¾ãã ãã¼ã¿ã»ãã ä»å㯠ãã¡ãã®å²©æ³¢ãã¼â¦
ã¯ããã« æç³»å解æãè¡ãä¸ã§ãç¶æ 空éã¢ãã«ãç¾ç¶ä½¿ç¨ã§ããææ³ã®ä¸ã§åªãã¦ããã®ã§ã¯ã¨èªåã®ä¸ã§è©±é¡ã ã£ãã®ã§ã 使ã£ã¦ã¿ããã¨ã«ãã¾ããã ç¶æ 空éã¢ãã«ã使ããã¨ã®ç®ç æç³»å解æãè¡ãä¸ã§ãã¢ãã«ã®ç¨®é¡ã¯ç¡æ°ã«ããã®ã§ããã ã©ãããâ¦
ã¯ããã« ã¹ãã¼ãã®ä¸çã«ããã¦ã¯ãå¢ãããããããæµããæ¥ã¦ãããçã®è¡¨ç¾ããããããã¨æãã®ã§ããããããã¯ãªã«ã«ãã§ã¯ãªãå®éã«åå¨ãããã®ãªã®ã§ããããï¼ãã®åéã§ã¯ãã¹ã±ã®ããããã³ãã®è©±ãæåã§ãè²ã ãªæ¹ãç 究ããã¦ããã®ãããªâ¦
ã¯ããã« stanã¨Rã§ãã¤ãºçµ±è¨ã¢ããªã³ã°ãèªã¿çµãã¾ããã æ¬ã®10ç« ã®å°æ£ã®å¼·ãã¨åè² ã ã©ãæ¨å®ããã¨ããå 容ãé¢ç½ãã£ãããã èªèº«ã®èå³ã®ããããã¹ã§åæ§ã®ãã¨ãå®æ½ãã¦ã¿ã¾ããã æ¢ãããåããã¼ãã§ãã£ã¦ããè¨äºããã£ãã®ã§ããã å ä½â¦
ã¯ããã« ç®±æ ¹é§ ä¼ã¯å¥½ãã§å²ã¨æ¯å¹´è¦ã¦ããã®ã§ããã ãã©ã½ã³ç³»ã®ç«¶æã£ã¦ããã©ã¼ãã³ã¹ã®åæ£ã対人競æã¨æ¯è¼ããã¨å°ãããªããããå®ã¯äºåã®æ å ±ããçµæãäºæ¸¬ãããã¨ãæ¯è¼ç容æãªããããªããï¼ã¨ããçåãæµ®ããã ã®ã§ãå大å¦ãåªåãã確â¦