Python
ã ãã¶é£å·æ°å³ã®æ¹ãå¤ãããããã¾ããããæ²ããã«ä»åãMMM (Marketing/Media Mix Modeling)ãã¿ãããã¾ãããã®åº¦åãä¸ããã®ã¯ãã¡ãã§ãããããPyMC Marketing MMMã§ããããã¾ã§ãåã®è¦³æ¸¬ç¯å²ã§ãããåºãæ®åãã¦ããPyMCãã¼ã¹ãªã®ããã£ã¦ãâ¦
éºã ããèªç¤¾ãããã¯ãã«ã¤ãã¦ã®ã¾ã¨ãè¨äºãæ¸ãã®ã¯è¥å¹²æ°ãå¼ããã®ã§ãããå æ¥Gemini 2.5 Proã«vibe codingã¨ãããã¨ã§MMMã®ãµã³ãã«ã³ã¼ããæ¸ããã¦ã¿ããæ¢ã«sunsetæ¸ã¿ã®Ligthweight MMMã使ã£ãã³ã¼ããè¿ã£ã¦ããä¸ã«ãããã«ããã£ã¦ãã®ããâ¦
å ææ¨è«: åºç¤ããæ©æ¢°å¦ç¿ã»æç³»åè§£æã»å ææ¢ç´¢ãç¨ããæææ±ºå®ã®ã¢ããã¼ãä½è :鿬 æãªã¼ã 社Amazonèè ã®éæ¬ãããããæåã§ãæµè´ããã ããã®ãããã¡ãã®ãå ææ¨è« âåºç¤ããæ©æ¢°å¦ç¿ã»æç³»åè§£æã»å ææ¢ç´¢ãç¨ããæææ±ºå®ã®ã¢ããã¼ãâãâ¦
æ¯å¹´åã®äºã®è¨ããªããæ¸ãã¦ããæ¨è¦æ¸ç±ãªã¹ãè¨äºã§ãããä½ã ããã ã§ä»å¹´ãæ¸ããã¨ã«ãã¾ããããªãæ¨å¹´åº¦çã®è¨äºãä¸ã«ãªã³ã¯ãã¦ããã¾ããã®ã§ã以åã®ãã¼ã¸ã§ã³ãèªã¾ãããæ¹ã¯ãã¡ãããèªã¿ãã ããã ä»åã®ãã¼ã¸ã§ã³ã§ã¯ãè¶£åãã¡ãã£ã¨â¦
CausalImpactã«ã¤ãã¦ã¯ãéå»ã«ãã®ããã°ã§ãä½åº¦ã話é¡ã«ãã¦ãããã¨æãã¾ãã端çã«è¨ãã°ãseasonalityã«ãããã¤ã¢ã¹ãè£æ£ããããã®å®é¨è¨ç»ã§ããDIDï¼Difference in Differencesï¼å·®åã®å·®åæ³ï¼ã«ãã£ã¦å¾ãããtest/controlã°ã«ã¼ãã®æç³»åãâ¦
(Image by wal_172619 from Pixabay)å»å¹´ã§æä¾ã®æ¨è¦æ¸ç±ãªã¹ãã®æ´æ°ã¯ä¸æ¦çµäºããã¤ããã ã£ããã§ãããè¨äºãå ¬éãã¦ä»¥éã«ãããã¯æ°ãã«ãªã¹ãå ¥ããããªãããã«ã¯ãããªãï¼ãã¨ããæ¸ç±ãå¹¾ã¤ãç¾ããäºæ ã«ãªãã¾ããã®ã§ãåè¨æ¤åãã¦ä»å¹´ãâ¦
以åãAds carryover & shape effectsä»ãã®Media Mix Modelingãã¨ããè¨äºã§åãä¸ãããã¤ã¸ã¢ã³MMMã®technical report (Jin et al., 2017)ã§ããã彿RStanã§å®è£ ããã¦ãããã®ã4å¹´ã®æãçµã¦æä»£ã®è¶¨å¢ã«æ²¿ãå½¢ã§Pythonãã¼ã¹ã®OSSã¨ãã¦ãªãªã¼ã¹ãâ¦
2å¹´ãããåã«å¿ è¦ããã£ã¦çTensorFlowã¨TensorFlow-Hubã«ããæ§ã ãªã¢ãã«ããã¬ã¼ã ã¯ã¼ã¯ä¸¦ã³ã«äºåå¦ç¿æ¸ã¿ã¢ãã«ã®å®è£ ã試ãã¦ããã®ã§ãããTF2ã®æµ¸éã«ä¼´ããããã®å¤ãã®ä»æ§ã夿´ã«ãªã£ã¦ãããä¸ã«ã¯åããªããªã£ã¦ããã³ã¼ãããã£ãã®ã§ããâ¦
å æ¥ã®è¨äºã§ãæ¸ããããã«ãã©ããããæè¿RStanå¨ãã®ç°å¢ãè²ã å³ãããªã£ã¦ããæ°ããã¦ãã¦ãä»®ã«RStanãä»å¾ç°å¢é¢ã§ã®ä¸å ·åããã¹ããããªã©ã§ä½¿ããªããªã£ãããã¤ã¸ã¢ã³ã¢ããªã³ã°ãããªããªã£ã¦å°ããããããã¨ãã屿§ãæè¿æ±ãã¤ã¤ããã¾ãâ¦
(MNIST database - Wikipedia) åã¯ç»åèªèåéã¯é夿¼¢ãªã®ã§ãããããæè¿åå¿è åãã«CNNã®ãã¬ã¼ãã³ã°ãè¡ããã¨ãä¼ç»ãã¦ãã¦ããã®ç®çã«é©ããç»åèªèã®ãªã¼ãã³ãã¼ã¿ã»ãããæ¢ãã¦ããã®ã§ããã ã¨ããã¨èª°ããæãã¤ãã®ãMNISTã§ã¯ãªããâ¦
ããã¼ã¿åæããããªãRã¨Pythonã®ã©ã¡ãã使ãã¹ãï¼ãã¸ã®å人çãªåç
(Background image by Pixabay)æè¿ã¾ãããã¼ã¿åæããããªãRã¨Pythonã®ã©ã¡ãã§ããã¹ãããã¨ãã話é¡ãåºã¦ããããã§ãã è¨èªä»æ§ããã®ä»ã®ä½¿ãåæã¨ããç¹ã§ã¯ã大ä½ãã®è¨äºã«æ¸ãã¦ãããã¨ãåèã«ããã°è¯ãã¨æãã¾ãããã®ä¸ã§ã人ã«ã¯å½ç¶â¦
å æ¥ã®è¨äºã§ã¯AutoML Tablesã試ãã¦ã¿ã¾ãããã調åã«ä¹ã£ã¦ããã¾ã§è§¦ã£ã¦ããªãã£ãAutoML Natural Languageã試ãã¦ã¿ããã¨æã£ãã®ã§ããã 以åã®è¨äºã«ãæ¸ããããã«ãåã¯å ã èªç¶è¨èªå¦çãè¦æã§TensorFlow Hubã®pre-trained modelã«ããfine-â¦
以åããGoogleã§ã¯Cloud AutoMLã¨ãã"Learning to learn"ãã¬ã¼ã ã¯ã¼ã¯ã«ããã人æå®å ¨ä¸è¦ã®å ¨èªåæ©æ¢°å¦ç¿ã¢ããªã³ã°ï¼API使ããµã¼ãã¹ãå±éãã¦ãã¦ãã¾ãããããããã¯ç»åèªèãå忍è¦ã¯ãã¾ãèªç¶è¨èªå¦çãã¡ã¤ã³ã§ãæããªã¼ã½ããã¯ã¹ãªâ¦
TensorFlow Probability (TFP)ããªãªã¼ã¹ããã¦ãããã°ããçµã¡ã¾ãããæè¿ã«ãªã£ã¦ãããªã¢ã¸ã¥ã¼ã«ãå ¬éãããã¨ç¥ãã¾ããã Framework for Bayesian structural time series modelsã¨é¡ããã¦ããéãã§ããºããªTFPã§ãã¤ãºæ§é æç³»åã¢ãã«ãæ¨å®ãâ¦
ä»åããã ã®åå¿é²ã§ãããã©ã¡ããã¨ããã¨çªå¤ç·¨ã§ããTensorFlowé¨åã¯ããã¾ã§ãTF-Hubã§ããã¹ããã¼ã¿ãfeature vectorã«ç´ãã¨ããã¾ã§ã®ã¿ã§ãããããå ã¯ä»ã¾ã§ãã¾ã試ãã¦ããªãã£ãt-SNE (t-distributed stochastic neighbor embedding)ã使ã£â¦
ããã¾ãå°ãã¿ã§ãã大ãããã¨ã¯ãã¦ããªãã®ã§ãèå³ã®ãªãæ¹ã¯èªã¿é£ã°ãã¦ãã ãã£ã¦çµæ§ã§ããä»åã®ãé¡ã¯ãææ åæ(sentiment analysis)ã§ãã顿ã¨ãã¦é¸ãã ã®ã¯ãä¸è¨ã®ãªã¼ãã³ãã¼ã¿ã»ããã§ãã ä¸è¬ã«ãææ åæèªä½ã¯NLPãè¦æãªåããè¦â¦
æ©æ¢°å¦ç¿ã®ã¨ãã»ã³ã¹ -å®è£ ããªããå¦ã¶Python,æ°å¦,ã¢ã«ã´ãªãºã - (Machine Learning)ä½è : å è¤å ¬ä¸åºç社/ã¡ã¼ã«ã¼: SBã¯ãªã¨ã¤ãã£ãçºå£²æ¥: 2018/09/21ã¡ãã£ã¢: åè¡æ¬ãã®ååãå«ãããã°ãè¦ãçºå£²ããã¦ããã ãã¶çµã¡ã¾ãããæ§æ³æ®µéã®é ããèâ¦
ãã£ããã®é±æ«ã«ããããããå°é¢¨ãæ¥ã¦ãã¾ããããã¹ãåºæ¥ãªããã°è¡æ©ããåºæ¥ãæã極ãããã¨ã«ãªã£ã¦ãã¾ã£ãã®ã§ãUCI ML repositoryãçºãã¦ãã¦è¦ã¤ããé©å½ãªãã¼ã¿ã»ããã«å¯¾ãã¦TensorFlowã§éã¶ã¨ãããã¨ããã¦ã¿ã¾ããã åºæ¬çã«ã¯ãã®ã·â¦
ããã¯ååã®è¨äºã®ç¶ãã§ãã å°ãã¿ã«ãã¦ãã ã®åå¿é²ã§ãã®ã§ãäºããäºæ¿ãã ããã ååã®è¨äºã§å ã åèã«ããã¦ããã ãã以ä¸ã®ããã°è¨äºãªãã§ãããããã¯åºæ¬çã«è±èªNNLMã®128次å embeddingã§è©¦ãããã®ãªãã§ããããã®ã¾ã¾ããã¨ACC 0.965ãâ¦
ã ãã¶ä¹ ãã¶ãã®çTensorFlowä¸è»¢å «åè¨ã§ããä»åããã ã®åå¿é²ã«ã¤ããä½ä¸ã¤æ°ãããã¨ãåèã«ãªããã¨ãæ¸ãã¦ããã¾ããããã¯ã½ã³ã¼ãã®ç¾ åã§ããããã¾ããã®ã§ãä½ã調ã¹ç©ã§ãã©ãçãããæ¹ã¯ãã®è¨äºã®ãªã³ã¯å ãªã©ãªã©ãã覧ãã ããã ä»åâ¦
ãã°ããåã«ãããªè¨äºãåºã¦ããã®ããè¦ãããã¾ããã æããã«ããã¯åãææï¼ç¬ï¼ã§é©å½ã«æ¾è¨ãããã¨ããã£ããã§å·»ãèµ·ãã£ãè°è«ã®ä¸æ¦ãªã®ã§ã¯ãªããã¨æãã®ã§ãããå人çã«ã¯ãã¡ãã®@yohei_kikutaããã®ä»°ãéãã§å¤§ä½è¯ãã®ã§ã¯ãªããã¨èâ¦
ä»å¹´ãä½åãæ¸ç±ããæµè´ããã ãããã§ãããå¿ããã£ããä»ã«æ¸ãããè¨äºããã£ããã§ãã®ã¿ã¤ãã³ã°ã§æ¸è©ãæ¸ããã³ãããã®ããå¤§å¤æç¸®ã§ãããã®ã¿ã¤ãã³ã°ã§ã¾ã¨ãã¦æ¸ããã¦ããã ãããã¨æãã¾ãã
ååã®è¨äºã§ããããDeepã£ã½ãã¨ããã¾ã§æ¥ãã®ã§ããã®ã¾ã¾ãã£ãã¨Deepãããã®è±¡å¾´ã§ãããCNN (Convolutional Neural Network)ã«ãã£ã¦ãã¾ããã¨æãã¾ããã¡ãªã¿ã«ä»åã大ãã¦åç §ãã¦ãã¾ããããåèæç®ã¨ãã¦æ·±å±¤å¦ç¿éæ¬ãæ²ãã¦ããã¾ããæ·±â¦
åºå¼µãã¦ãããè¶³åºçèçã«ããã£ãããã¦ãããã¡ã«ããã£ããçTensorFlowã®åå¼·ãæ»ã£ã¦ãã¾ã£ã¦ãã¾ãããã¨ããããå ã«é²ã¿ã¾ãããã¤ãéãã§ãããç¹ã«æå³ã¯ãªããã®ã®æç§æ¸ã¨ãã¦ãã¡ããæãã¦ããã¾ããæ·±å±¤å¦ç¿ (æ©æ¢°å¦ç¿ãããã§ãã·ã§ãã«â¦
ä½ã¨ãªãTFã®ã使³ãåãã£ã¦ããã®ã§ã©ãã©ãå ã«è¡ãã¾ããããè¨ãã°ãã ã®åå¿é²ãªã®ã§ä½ä¸ã¤åºå ¸ã¨ãåèæç®ã¨ãæ¸ãã¦ã¾ãããããã®ã·ãªã¼ãºã§ãã£ã¦ãããã¨ã®çè«çåºç¤ã¯æ·±å±¤å¦ç¿éæ¬ãã»ã¼å ¨ã¦ã«ãã¼ãã¦ãã¾ããæ·±å±¤å¦ç¿ (æ©æ¢°å¦ç¿ãããã§ãâ¦
ååã¯ãã¸ã¹ãã£ãã¯å帰ããã£ããã§ãããããè¨ãã°æ®éã®ç·å½¢å帰ãã£ã¦ãªãã£ãã®ã§ãã£ã¦ããã¾ããç·å½¢åå¸°ã¯æ¼«ç¶ã¨OLSã§ãããªãæ®éã«éè¡åè¨ç®ãã¦ããã¾ããªãã§ãããããã ã¨é¢ç½ããªãã®ã§ããã¦å¾é æ³ã§ããã¨ããéæããªãã¨ããããã¨æâ¦
ããããèãã¦ã¿ãããTensorFlowãããã¯ã¨ã³ãã¨ãã¦Kerasãåãããããã«Rä¸ã§åãããããããã¨ã¯ãããã®ã®ãèå¿ã®TensorFlowãçã§æ¸ãããã¨ã£ã¦æ°ããããããããªãã£ãã®ã§ãããããã§ã¯ãããã¨æãã®ã§ã仿´ãªããã§ããTensorFlowãçã§â¦
ï¼å·¦ï¼Kerasãå³ï¼MXnetï¼Kaggle Masterã®éã§ã¯MXnetããããã«äººæ°ãªDeep Learningãã¬ã¼ã ã¯ã¼ã¯ã¨ãããã©ããã¼ãã@fcholletæ°ã®æã«ããKerasã Keras Documentation çµæ§è¦å¿ããã®ã§ãããããããæå ã®Pythonç°å¢ã§èµ°ãããã«ãªã£ãã®ã§ã試ãã¦â¦
è¦ç¯2å¹´ãã¨ãã¨ã宿ãã¾ãããæ©æ¢°å¦ç¿ã®ãã©ã¡ã¼ã¿ãã¥ã¼ãã³ã°ã«æ©ããçããã®ããã«ãã³ã¼ãã£ã³ã°ãæ°å¦ã大ã®è¦æãªåãé å¼µã£ã¦ä½ãã¾ããããããPTGH (Parameter Tuning by God's Hand)ãã¬ã¼ã ã¯ã¼ã¯ã§ããRã§ãPythonã§ãåãã¾ããä¸èº«ã¨ãã¦â¦
ããè¨ãã°3å¹´åã«ãããªã¾ã¨ãçã¨ã³ããªãæ¸ããã®ã§ããããã®å 容ã¯ãã®ã¾ã¾ããªãã®é¨åã2å¹´åã«åè¡ããæèã®åæ¡ã«ããªã£ãã¨ãããã¨ã§ãè²ã æãåºæ·±ãã¨ã³ããªã§ãããªã®ã§ãããã»ã»ã»ãã®3å¹´ã®éã«çµ±è¨å¦ã»æ©æ¢°å¦ç¿ã»ãã¼ã¿ãã¤ãã³ã°ã®è«¸ææ³â¦