Guy Blanc Neha Gupta Jane Lange Li-Yang Tan ! x f(x) 000010101 - PowerPoint PPT Presentation

Guy Blanc Neha Gupta Jane Lange Li-Yang Tan

𝑦 ! x f(x) 000010101 -1 𝑦 " 𝑦 # 011011010 1 100100111 1 101001000 1 𝑦 " -1 1 1 001010010 -1 −1 1 ID3, C4.5, CART

1) Choose a good root 𝑦 $ [variable influence, information gain, Gini impurity] 2) Recurse on subtrees 𝐶𝑣𝑗𝑚𝑒(𝑔 𝑦 $ < 0 ) 𝐶𝑣𝑗𝑚𝑒(𝑔 𝑦 $ ≥ 0 ) “Good” root = relevant, important, influential

Heuristics do poorly Monotone Noise-stable functions [already well-known] functions [new in BLT20] [new in BGLT20] Heuristics do well New algorithm: provable guarantees for all functions [BGLT20]

𝑦 ! Stochastic node 1 𝒚 𝟐 𝒚 𝟑 𝒚 𝟒 𝒚 𝟓 𝒚 𝟔 $2 $0 $5 $1200 $45 -1 1 Query cost vector Randomized DT