æ©æ¢°å¦ç¿ã¢ãã«ã®äºæ¸¬çµæã説æããããã®åã欲ããã...ï¼
ã¯ããã«
æè¿ã¯AIãæ©æ¢°å¦ç¿ãªã©ã®åèªããã¸ãã¹ã§æµè¡ã£ã¦ãã¦ãä¸ã¯AIæ代ãè¿ãã¦ãããQiitaãTwitterãçºãã¦ããã¨ãã®å½±é¿ãåãã¦ãä¸ã®å¤ãã®ã¨ã³ã¸ãã¢ãAIã®åå¼·ãå§ãåºãã¦ããããã«è¦åããããã
ããã«ãè¿å¹´ã§ã¯æ©æ¢°å¦ç¿ã®ã©ã¤ãã©ãªãå å®ãã¦ããã誰ã§ãæ©æ¢°å¦ç¿ãå®è£ ãããã¨ãã§ããè¯ãæ代ã«ãªã£ã¦ããã
ãã®ä¸æ¹ã§ãç¹å¾´é¸æãè¡ã精度ãåä¸ãããããæ©æ¢°å¦ç¿ã®åºããçããã©ã®ç¹å¾´ã«åºã¥ãã¦å¤æãããã®ããç解ãããããã«ã¯ãã¢ãã«ã«å¯¾ããç解ããã¯ããã¯ãå¿ è¦ã¨ãªãå ´åãå¤ã ãããè¤éãªã¢ãã«ã«ãªãã¨äººéã«ã¯è§£éãå°é£ã§èª¬æãé£ãããè¿é æµè¡ãã®Deep Learningç³»ã®ã¢ãã«ã ã¨é »ç¹ã«ããªããããããããªããã©ãã¾ãããã¾ãããã¨ãªã£ã¦ããã¨æãã
ä¸è¬çãªã¨ã³ã¸ãã¢ã¨ãã¦ã¯ããã®ç¹ãå²ã¨èª²é¡ãªããããªããã¨åæã«æã£ã¦ãããã¨ããããç§ã課é¡ã«æãã¦ãããï¼ç¹ã«å®æ¥åã§æ©æ¢°å¦ç¿ãã¦ããªãä¸ã«ãã¨ã³ã¸ãã¢ã§ããªããï¼
ãããªããã§ãä»åã¯ãã®èª²é¡ã解決ããããã®ãã¼ã«ã§ããLIME(Local Interpretable Model-agnostic Explainations)ãèå³æ·±ãã£ãã®ã§ãç´¹ä»ãã¦ããããã¨æãã
â»æ¬è¨äºã¯LIMEã®ã¢ã«ã´ãªãºã ã®èª¬æã¨ãªããããLIMEãå®éã«å©ç¨ãããæ¹ã¯GitHub - marcotcr/lime: Lime: Explaining the predictions of any machine learning classifierã§pythonã®ã©ã¤ãã©ãªã¤ã³ã¹ãã¼ã«æ¹æ³ã¨ãã¥ã¼ããªã¢ã«ãè¼ã£ã¦ããã®ã§ããã¡ãããåç
§ãã ããã
ã¢ãã«ã®èª¬æã¨ã¯ä½ã
LIMEã®ç´¹ä»ã«ç§»ãåã«æ©æ¢°å¦ç¿ã¢ãã«ã説æããã¨ã¯ã©ããããã¨ãªã®ãæ´çãã¦ããããã
æ©æ¢°å¦ç¿ã¢ãã«ã®èª¬æã«ã¯ä¸è¨ã®èª¬æã®2種é¡ãèããããã
- explaining prediction(äºæ¸¬ã®èª¬æ)
ãã¼ã¿ä¸ã¤ã«å¯¾ããæ©æ¢°å¦ç¿ã¢ãã«ã®åé¡å¨ã«ããäºæ¸¬çµæã«å¯¾ãã¦ãã©ããã¦åé¡ãè¡ãããã®ãã説æãããã¨ã(ä¸è¨ã®å³ã¯ã¤ã¡ã¼ã¸)
(åºå
¸: https://arxiv.org/pdf/1602.04938.pdf)
- explaining models(ã¢ãã«ã®èª¬æ)
åé¡å¨ãã©ãããæ§è³ªãæã£ã¦ããã®ãã説æãããã¨ã
(åºå
¸: Introduction to Local Interpretable Model-Agnostic Explanations (LIME) - O'Reilly Media)
LIMEã¯ãã®ãã¡ãexplaining predictionãè¡ãããã®ã¢ã«ã´ãªãºã ã§ããã
explaining modelsã«ã¤ãã¦ã¯ãSP-LIMEã¨å¼ã°ããã¢ã«ã´ãªãºã ãè«æã«è¨è¼ããã¦ããã®ã§ããã¡ããåç
§ããããã(æ°ãåãã°ãSP-LIMEã«ã¤ãã¦ãè¨äºãæ¸ã)
LIME(Local Interpretable Model-agnostic Explainations)ã®ç´¹ä»
LIMEã¨ã¯ï¼
KDD2016ã§æ¡æããããâWhy Should I Trust You?â Explaining the Predictions of Any Classifierãã¨ããã¿ã¤ãã«ã®è«æã§çºè¡¨ãããã¢ã«ã´ãªãºã ãåé¡å¨ãã©ã®ããã«å¤æãã¦ã©ããªã³ã°ãè¡ãªã£ãã®ãã人éã§ã解éã§ãããããªå½¢ã§æ示ãã¦ãããã
ãã®ã¢ã«ã´ãªãºã ã¯ãããã¼ã¿ãåé¡ããçµæãããããã®ç¹å¾´ãã©ã®ç¨åº¦åé¡ã«è²¢ç®ãã¦ãããã調ã¹ããã¨ã§åé¡å¨ã®äºæ¸¬çµæã説æãã¦ãããã¾ããåé¡å¨ã®äºæ¸¬çµæãç¨ãããããä»»æã®åé¡å¨ã«é©ç¨ã§ããç¹å¾´ãããã
LIMEã®ã¢ã¤ãã¢
ãã¼ã¿ã®å¨è¾ºãããµã³ããªã³ã°ãããã¼ã¿ãç¨ãã¦ã説æãããåé¡å¨ã®åºåã¨è¿ä¼¼ããããã«è§£éå¯è½ãªï¼ãã¤åç´ãªï¼ã¢ãã«ãå¦ç¿ãããããã®å¾ãå¾ãããåé¡å¨ãç¨ãã¦åé¡çµæã®è§£éãè¡ããä¸è¨ãã¤ã¡ã¼ã¸å³ï¼è«æããæç²ããå³ãç·¨éï¼ã
(åºå
¸: https://arxiv.org/pdf/1602.04938.pdf)
説æç¨åé¡å¨ã®å¦ç¿æ¹æ³
説æç¨åé¡å¨ã¯ãã¼ã¿ã®å¨è¾ºã§ã®çµæã¨è¿ä¼¼ããããã«ãããã
ããããããã«ãä¸è¨ã®ç®çé¢æ°ãå©ç¨ãã¦å¦ç¿ããã
- : 解éå¯è½ãªã¢ãã«ã®éå
- : Gã®ãã¡ã®ä¸ã¤ã®ã¢ãã«ãä¾ãã°ãç·å½¢ã¢ãã«ãªã©
- : 説æãããåé¡å¨
- : ãã¼ã¿ã¨ã®è·é¢
- : ãã¼ã¿ã®å¨è¾ºã§ã¨ã®çµæãã©ãã ãéã£ã¦ãããï¼ã¯æ失é¢æ°ã¨ãããï¼
- : 説æç¨åé¡å¨ã®è¤éã
ä¸è¨ã®å
容ãããã¯ãã¼ã¿ã®å¨è¾ºã§ã¨ã®çµæã«ã¤ãã¦ã®é£ãéãã¨èª¬æç¨åé¡å¨ã®è¤éãã®åãæå°ã«ãã ã®éåãæ±ãããã®ã§ããã¨è¨ããã
ããã§ãã¯ããã¹ãåé¡ã®å ´åã解éå¯è½ãªã¢ãã«ã®ç¹å¾´è¡¨ç¾ãåèªã®æç¡ã®Bag-of-Wordsæ³(åèªè¢è©°ã)ã¨ããåèªã®æ°(次å
æ°)ã«é度Kãè¨å®ãããã¨ã§ã説æã解éå¯è½ã§ãããã¨ãä¿è¨¼ããããã®ãã®ãããã
ç»åãã¼ã¿ã®å ´åã¯super-pixelsã¨å¼ã°ããä»»æã®ã¢ã«ã´ãªãºã ã使ç¨ãã¦è¨ç®ããããã®ãç¨ãã¦è§£éå¯è½ãªã¢ãã«ã®ç¹å¾´è¡¨ç¾ã¨ããã
ããã§ããã®ç¹å¾´è¡¨ç¾ã¯ã®2å¤ã§è¡¨ããã1ã¯å
ã®super-pixelsã0ã¯ã°ã¬ã¼ã¢ã¦ããããsuper-pixelsã示ãã
ããã¾ã§ã§ã«ã¤ãã¦ãä½ã¨ãªãã¨ããã¬ãã«ã§ã¯ç解ãã§ããã¨æãããã
ããã§ã次㯠ã®æ°å¼ã«ã¤ãã¦ãè¦ã¦ãããã
- : ã®å¨è¾ºã®ãã¼ã¿ã®éå
- : éã¼ãè¦ç´ ãä¸é¨ã ãå«ããµã³ããªã³ã°ã«ããçæããã2å¤ã®ã¹ãã¼ã¹ãªç¹ã
ãã§å®ç¾©ããã
- : ãç¨ãã¦å¾©å ãããå ã®ãµã³ãã«ã®ç¹å¾´è¡¨ç¾ãã§å®ç¾©ããã
ãã®å¼ãè¦ãéããã®å¨è¾ºã®ãã¼ã¿ã«ãããã§éã¿ä»ãããæ®å·®å¹³æ¹åãåºãã¦ããã
æ®å·®å¹³æ¹åèªä½ã¯æ£è§£ãã¼ã¿(ä»åã®å ´åã説æãããåé¡ã¢ãã«ã®äºæ¸¬çµæ)ã¨æ¨å®ã¢ãã«ã®äºæ¸¬çµæã¨ã®éã®ä¸ä¸è´ãè©ä¾¡ãã尺度ãªã®ã§ããããããããã¨æãã
ã¾ããã§éã¿ä»ããã¦ããçç±ã«ã¤ãã¦ç解ãããããã®å¼ãè¦ã¦ãããã
- : ã¨ã¨ã®è·é¢é¢æ°(ä¾ãã°ãããã¹ããªãã³ãµã¤ã³é¡ä¼¼åº¦ãç»åãªãL2ãã«ã ãªã©ãå©ç¨ãã)
- : ææ°ã«ã¼ãã«ã®ã«ã¼ãã«å¹
ã®å¼ã¯ã«ã¼ãã«é¢æ°ã§ãããxã¨zã®2å¤æ°éã®é¡ä¼¼åº¦ãç®åºãã¦ãããã¯ãããã¼ã«0ãã1ã¾ã§ã®å¤ãå
¥ãã¦è¦ã¦è¨ç®ããã°ãããã¨æããããµã³ãã«ãè¿ããã°è¿ãã»ã©å¤ãå°ãããªããããã§éã¿ä»ããããã¨ã§ãã¨ã¨ã®è·é¢ãè¿ããµã³ãã«ã®å ´åã¯æ失ãå°ãããªãããããªããéã«è·é¢ãé ããµã³ãã«ã®å ´åã¯æ失ãé«ããªãããã®éã¿ä»ãã®ãããã§ãããã¹ããªã¢ãã«ã¨ãªã£ã¦ããã
æå¾ã¯ã«ã¤ãã¦æãä¸ãã¦ãããã®å¼ãè¦ã¦ãããã
ã¯å©ç¨ããç¹å¾´ãããã ãåèªæ°(ãããã¯super-pixels)Kç¨åº¦ã ãã¨ãããã¨ã示ãã¦ããã£ã½ãã
å©ç¨ããç¹å¾´ã®é¸æã¯ãæ¹ç¨å¼ããç´æ¥è§£ããã¨ã§å®ç¾ãããã¨ã¯é£ããã
ãã®ãããã¾ãèè
ããK-Lassoã¨å¼ãã§ãããLassoã§æ£ååãã¹ã使ç¨ãã¦å©ç¨ããç¹å¾´ãKåé¸æããæå°äºä¹æ³ãä»ãã¦éã¿ãå¦ç¿ããæ¹æ³ã«ãã£ã¦ãå©ç¨ããç¹å¾´ã®é¸æã«ã¤ãã¦ã®è§£ã¨è¿ä¼¼ãããã
ããã«ãããæ¹ç¨å¼ã解ããã¨ãã§ããããã«ãªããããç·å½¢ã¢ãã«(Githubã®ã³ã¼ããèªãéãã§ã¯Ridgeå帰)ã§å¦ç¿ãè¡ãã
ãã®å¦ç¿ããç·å½¢ã¢ãã«ã®åå帰ä¿æ°ã確èªãããã¨ã§ãé¸æãããç¹å¾´ã«ã¤ãã¦ãã©ãã ãåé¡ã«è²¢ç®ãã¦ãããã®èª¬æãè¡ããã¨ãã§ããã
ããã¾ã§ã説æããå
容ãä¸è¨ã®å³ã®Algorithm 1 ã§ããã
(åºå
¸: https://arxiv.org/pdf/1602.04938.pdf)
Algorithm 1 ã¯åã ã®äºæ¸¬ã«ã¤ãã¦ã®èª¬æãçæããã®ã§ããã®è¤éãã¯ãã¼ã¿ã»ããã®ãµã¤ãºã«ä¾åããã®ã§ã¯ãªãããè¨ç®ããæéã¨ãµã³ãã«æ°ã«ä¾åãããããã
æ¤è¨¼ã¨èå¯
æ¤è¨¼ãã©ã
ä»åã¯ãã«ã¦ã§ã¢ã¨æ£å¸¸ãªããã°ã©ã ã®APIã³ã¼ã«ã®ãã¼ã¿ã»ãããæå
ã«ãã£ãã®ã§ãèè
ãã®LIMEããã±ã¼ã¸ã使ã£ã¦ã¿ããã¨ã«ããã
ãã¼ã¿ã»ããã¯ä¸è¨ããã¨ã£ã¦ãããã®ã ã
Malicious datasets * - Csmining Group
ãã¼ã¿ã»ããã®å 容ã¯ãã¡ã¤ã«å½¢å¼ãcsvããã«ã¦ã§ã¢ã®æ°ã320æ¤ä½ãæ£å¸¸ãªããã°ã©ã ã®æ°ã68æ¤ä½ã¨ããå¾®å¦ãªæ°ã¨ãªã£ã¦ããã
ç°¡åãªæ¤è¨¼ã®çµæã¯ä¸è¨ã®éãã ã£ãã
github.com
æ¤è¨¼ç¨ã«ã©ã³ãã ãã©ã¬ã¹ãã使ã£ã¦ãã«ã¦ã§ã¢ã¨æ£å¸¸ãªããã°ã©ã ãåé¡ããã
ã©ã³ãã ãã©ã¬ã¹ããé¸ãã çç±ã¯ãç´ç·ã§ã¯ãªãåé¢å¢çãå¼ãã¦ããã¦ãã¤ããã®ã¢ãã«èªä½ãéè¦ãã¦ããç¹å¾´ãåºããããã ãä»ã«ããåé¡å¨ãããã°æãã¦ããã ãããã¨ããã
è³ç´°èãæ»ãã§ããã®ã§ããã¼ã¿ãå¦ç¿ç¨(ãã«ã¦ã§ã¢ã310æ¤ä½ã¨æ£å¸¸ãªããã°ã©ã 68æ¤ä½)ã¨ãã¹ãç¨(ãã«ã¦ã§ã¢ã10æ¤ä½)ã«æã§åããã
ãã®ãã¹ãç¨ã®äºæ¸¬çµæã¯f1_scoreã1.0ã¨ãªã£ãããã«ã¦ã§ã¢ã¨æ£å¸¸ãªããã°ã©ã ã®APIã³ã¼ã«ãç¨ããåé¡ã¯å²ã¨ç·å½¢åé¢å¯è½ãªãã®ãå¤ãå°è±¡ãªã®ã§ã交差æ¤è¨¼ã¨ããã¦ããªãä¸ã«ããã¹ãæ°ãå°ãªãã®ã§ãããªããã§ã¯ããã¨ã¯æãã
èå¯ãã©ã
çµæã®èå¯ã ããLIMEã§åºåããã "GetFileAttributesW"ã"GetFullPathNameW"ã"GetLongPathNameW"
ãªã©ã®ç¹å¾´ããã©ã³ãã ãã©ã¬ã¹ãã®ç¹å¾´ã©ã³ãã³ã°ã®ä¸ä½ã«é£ãè¾¼ãã§ãããã¨ããããã
LIMEã§åºåãããã®ã¯ããã®ãã¼ã¿åä½ã®ã©ã®ç¹å¾´ãéè¦ãã¦åé¡ãããã§ããexplaining predictionsã«ããããã©ã³ãã ãã©ã¬ã¹ãã®ç¹å¾´ã©ã³ãã³ã°ã¯å¤åexplaining modelsãªã®ã§ãå³å¯ã«æ¯è¼ãã¹ãã§ã¯ãªããããããªãã
ããããã¢ãã«èªä½ãéè¦ãã¦ããç¹å¾´ã¯explaining predictionsã®ä¸ä½ã«æ¥ã¦ãç´æçã«ã¯ãããããªãã¨æãã®ã§ãããæãã«ãªãã説æã§ãã¦ããæ°ãããã
åèæç®
LIMEè«æï¼
âWhy Should I Trust You?â Explaining the Predictions of Any Classifier
https://arxiv.org/pdf/1602.04938.pdf