Reflections on Machine Learning) nTool 1. Encoding causal assumptions: Transparency and testability nTool 2. Do-calculus and the control of confounding nTool 3. The algorithmitization of counterfactuals nTool 4. Mediation analysis and the assessment of direct and indirect effects nTool 5. Adaptability, external validity, and sample selection bias nTool 6. Recovering from missing data nTool 7. Causal discovery (因果探索) 5
(Pearl, 1995) • 操作変数法 (Wright, 1928) • フロントドア基準 (Pearl, 1993) • 代替変数 (Kuroki & Pearl, 2014) n十分条件の例 • 𝑥の親を全てで調整 24 𝐸 𝑦 𝑑𝑜 𝑥 = 𝐸 #の親(𝐸(𝑦|𝑥, 𝑥の親)) 介⼊後の期待値を条件付き期待値で推定可能: 回帰の問題へ (機械学習の利⽤) X Y U X Y U I M Y U X X Y W 誤差変数は省略 U
https://github.com/cdt15/lingam MITライセンス nCausalas by SCREEN AS nNode AI by NTT Communications nNTech Predict by neutral nCausal analysis by NEC nノンパラ: pcalg, causal-learn 42