Sparse linear models

Transcript

1. Sparse linear models Daisuke Yoneoka October 20, 2014 Daisuke Yoneoka Sparse linear models October 20, 2014 1 / 21
2. Notations γ ͸ bit vector Ͱ, ಛ௃ྔ j ͕ؔ࿈͋Δ৔߹͸ γj = 1, ͦΕҎ֎͸ 0. ∥γ∥0 = D j=1 γj ͸ l0 pseudo-norm. ∥γ∥1 = D j=1 |γj| ͸ l1 norm. ∥γ∥2 = ( D j=1 γ2 j )1/2 ͸ l2 norm. subderivative (ྼඍ෼): ತؔ਺ f : I → R ͷ θ0 Ͱྼඍ෼ͱ ͸,f(θ) − f(θ0) ≥ g(θ − θ0) θ ∈ I Λຬ଍͢Δ g ͷू߹ NLL: negative log likelihood, NLL(θ) ≡ − N i=1 log p(yi|xi, θ) Daisuke Yoneoka Sparse linear models October 20, 2014 2 / 21
3. l1 regularization: basics l0 (i.e.∥w∥0) ͸ತؔ਺Ͱͳ͍, ࿈ଓͰ΋ͳ͍! → ತؔ਺ۙࣅ! p(γ|D) ΛٻΊΔ͜ͱͷ೉͠͞ͷ͍͘Β͔͸ γ ∈ {0, 1} ͱ཭ࢄͰ͋Δ͜ͱ Prior p(w) Λ࿈ଓͳ෼෍ (ϥϓϥε෼෍) Ͱۙࣅ͢Δ. p(w|λ) = D j=1 Lap(wj |0, 1/λ) ∝ D j=1 e−λ∥wj ∥ േଇ෇͖໬౓͸ f(w) = log p(D|w) − log p(w|λ) = NLL(w) + λ∥w∥1 . ͜Ε͸ argminw NLL(w) + λ∥w∥0 ͱ͍͏ non-convex ͳ l0 ͷ໨తؔ਺ͷತؔ਺ۙࣅ ͱߟ͑ΒΕΔ Linear regression ͷ৔߹ (Known as BPDN (basis pursuit denoising)) f(w) = N i=1 − 1 2σ2 (yi − (wT xi))2 + λ∥w∥1 = RSS(w) + λ′∥w∥1 ͨͩ͠,λ′ = 2λσ2 Prior ʹ̌ฏۉϥϓϥε෼෍Λ͓͍ͯ,MAP ਪఆ͢Δ͜ͱΛ l1 ਖ਼ଇԽͱݺͿ Daisuke Yoneoka Sparse linear models October 20, 2014 3 / 21
4. ͳͥ l1 ਖ਼ଇԽ͸εύʔεͳͷ͔ʁ Linear regression ʹݶఆ͢Δ͕ GLM Ұൠʹ֦ுՄೳ ໨తؔ਺͸ minwRSS(w) + λ∥w∥1 ⇔ LASSO: minwRSS(w)s.t. λ∥w∥1 ≤ B B খˠ λ େ ͜Ε͸, Quadratic program (QP) ͱͳ͍ͬͯΔ. ͪͳΈʹ minw RSS(w) + λ∥w∥2 2 ⇔ RIDGE: minw RSS(w)s.t. λ∥w∥2 2 ≤ B Figure: 13.3; l1 (left) vs l2 (right) regularization Daisuke Yoneoka Sparse linear models October 20, 2014 4 / 21
5. Optimality conditions for lasso Lasso ͸ non-smooth optimization (ඍ෼ෆՄೳ࠷దԽ) ͷྫ. ໨తؔ਺͸ minw RSS(w) + λ∥w∥1 ୈҰ߲ͷඍ෼͸ ∂ ∂wj RSS(w) = ajwj − cj . ͨͩ͠ aj = 2 n i=1 x2 ij , cj = 2 n i=1 xij(yi − wT −j xi,−j) j ͱ j ͳ͠ͷ࢒ࠩͷ಺ੵ cj ͸ j ൪໨ͷಛ௃ྔ͕ y ͷ༧ଌʹͲΕ͚ͩؔ࿈͍ͯ͠Δ͔Λදݱ શମͷྼඍ෼͸ ∂wj f(w) = (aj wj − cj ) + λ∂wj ∥w∥1 = ⎧ ⎪ ⎨ ⎪ ⎩ {ajwj − cj − λ} if wj < 0 [−cj − λ, −cj + λ] if wj = 0 {ajwj − cj + λ} if wj > 0 Matrix form Ͱॻ͘ͱ, XT (Xw − y)j RSS ͷඍ෼ͷ෦෼ ∈ ⎧ ⎪ ⎨ ⎪ ⎩ {−λ} if wj < 0 [−λ, λ] if wj = 0 {λ} if wj > 0 Daisuke Yoneoka Sparse linear models October 20, 2014 5 / 21
6. Optimality conditions for lasso (Cont. 2) cj ͷ஋ʹΑͬͯ ∂wj f(w) = 0 ͷղͱͯ͠ఆٛ͞ΕΔ ˆ wj ͷ஋͸̏ύλʔϯ cj < −λ: ಛ௃ྔ͸࢒ࠩͱڧ͘ෛͷ૬ؔ, ྼඍ෼͸ ˆ wj = cj + λ aj < 0 ʹ͓͍ͯ 0. cj ∈ [−λ, λ]: ಛ௃ྔ͸࢒ࠩͱऑ͘૬ؔ, ྼඍ෼͸ ˆ wj = 0 ʹ͓͍ͯ 0. cj > −λ: ಛ௃ྔ͸࢒ࠩͱڧ͘૬ؔ, ྼඍ෼͸ ˆ wj = cj − λ aj > 0 ʹ͓͍ͯ 0. ͭ·Γɺ ˆ wj(cj) = ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ cj + λ aj if cj < −λ 0 if cj ∈ [−λ, λ] cj − λ aj if cj > λ ⇔ ˆ wj(cj) = soft( cj aj ; λ aj ) ͨͩ͠ɺsoft ͸ soft thresholding Ͱఆٛ͸ soft(a; δ) ≡ sign(a)(|a| − δ)+ Daisuke Yoneoka Sparse linear models October 20, 2014 6 / 21
7. Optimality conditions for lasso (Cont. 3) LASSO (Tibshirani, 1996) ͸݁ہ, λ = 0 ͷͱ͖ ˆ w ͸ OLS ͱҰॹ λ > λmax ͷͱ͖ ˆ w = 0 (ͨͩ͠ λmax = ∥XT y∥∞ = max|yT x:,j |) ͜ͷܭࢉํ๏͸,(XT y)j ∈ [−λ, λ] ͳΒ͹ 0 ͕࠷దͰ͋Δ͜ͱΛར༻ Ұൠతʹ͸ λmax = max|∇jNLL(0)| Daisuke Yoneoka Sparse linear models October 20, 2014 7 / 21
8. LS, lasso (l1), ridge (l2), subset selection (l0) ͷൺֱ X ͸ਖ਼ن௚ަ, ͭ·Γ XT X = I ΛԾఆ͓ͯ͘͠ͱ, RSS(w) = ∥y − Xw∥2 = yT y + wT XT Xw − 2wT XT y = const + k w2 k − 2 k i wk xi yi OLS ղ͸ ˆ wOLS k = xT :k y Ridge ղ͸ ˆ wridge k = ˆ wOLS k 1 + λ Lasso ղ͸ sign( ˆ wOLS k ) | ˆ wOLS k | − λ 2 + subset selection ղ͸ ˆ wSS k = ˆ wOLS k if rank(| ˆ wOLS k |) ≤ K 0 otherwise Daisuke Yoneoka Sparse linear models October 20, 2014 8 / 21
9. ਖ਼ଇԽύε ಛ௃ྔ͝ͱʹ ˆ w(λ) ͱ λ ͷ஋Λϓϩοτͨ͠΋ͷ Lasso ͸ D > N ͷ৔߹Ͱ΋ N ·Ͱ͔͠ม਺બ୒Ͱ͖ͳ͍ Elastic net ͳΒ ͹ D ·Ͱͷ਺ͷม਺બ୒Մೳ Daisuke Yoneoka Sparse linear models October 20, 2014 9 / 21
10. Ϟσϧબ୒ Ϟσϧબ୒ͷҰகੑʹ͍ͭͯ (cf. AIC, BIC, MDL ͳͲͷ৘ใྔج४࿦૪) ఆٛ: (ਖ਼͍͠Ϟσϧؚ͕·Ε͍ͯΔͱ͍͏લఏͷԼͰ) N → ∞ Ͱਖ਼͍͠Ϟσϧͷύϥϝʔληοτ ͕બ୒͞ΕΔ͜ͱ debiasing: Lasso Ͱ non-zero ͱਪఆ͞Εͨಛ௃ྔΛ༻͍ͯ࠶౓ OLS (ඞཁ. ͳͥͳΒ,Lasso Ͱ͸ؔ܎͋Δ܎਺΋ͳ͍΋ͷ΋ ॖখਪఆ͍ͯ͠Δ͔Β) ΫϩεόϦσʔγϣϯͰ༧ଌਫ਼౓Ͱ λ ܾఆ. ͜Ε͸,true ϞσϧΛબ୒Ͱ͖Δ஋ʹͳΔͱ͸ݶΒͳ͍. (ͳͥͳΒ Lasso ͸ॖখਪఆʹͳ͍ͬͯΔͷͰ, ॏཁ ͳಛ௃ྔΛ࢒ͨ͢Ίʹ͸ λ ͸গ͠େ͖ΊʹऔΔඞཁ͕͋Δ͔Β) ؔ܎ͳ͍ಛ௃ྔ΋ؚΊΔͷͰ false positive ͕ଟ͘ͳΔ Ϟσϧબ୒ͷҰகੑ͕ͳ͍! (Meinshausen, 2006) Ch.13.6.2 Ͱ per-dimension ʹΑΔ λ ͷνϡʔχϯάΛ঺ հ (બ୒ͷҰகੑ͋Γ) ܽ఺: σʔλ͕গ͠มΘ͚ͬͨͩͰ݁Ռ͕มΘΔ (Bayesian approach ͷํ͕ robust) Bolasso (Bach, 2008): Bootstrap Ͱղܾ: stability selection of inclusion probability (Meinshausen, 2010) Λܭࢉඞཁ Daisuke Yoneoka Sparse linear models October 20, 2014 10 / 21
11. ϥϓϥε෼෍Λࣄલ෼෍ʹ࣋ͭ sparse linear model ͷ Bayes ਪଌ ͜Ε·Ͱͷྫ͸ॴҦ MAP ਪఆ posterior ͷ mode ͸ sparse ͕ͩ, mean ΍ median ͸ͦ͏Ͱͳ͍ posterior ͷ mean ΛೖΕͨ΄͏͕༧ଌೋ৐ޡࠩΛখ͘͞Ͱ͖Δ Elad, 2009 ͸ spike-slab model Ͱ posterior mean ͷํ͕༧ଌੑೳ͕͍͍ࣄΛূ໌ ͨͩ͠, ܭࢉྔ͸ߴՁ Daisuke Yoneoka Sparse linear models October 20, 2014 11 / 21
12. l1 ਖ਼ଇԽͷΞϧΰϦζϜ ೋ৐ϩεؔ਺ͷ࠷దԽʹݶఆ͢Δ. (ͦͷଞͷϩεؔ਺΁΋֦ுՄೳ) Coordinate descent: Ұؾʹ࠷దԽͰͳ͘, ͦͷଞશͯΛ fix ͯ͠ 1 ͚ͭͩ࠷దԽ w∗ j = argminz f(w + zej ) − f(w) (z ͸ j ൪໨͕ 1 ͷ unit ϕΫτϧ) Ұ࣍ݩͷ࠷దԽ͕ղੳతʹղ͚Δ৔߹ʹ༗ޮ 1 ͔ͭͮͭ͠࠷దԽͰ͖ͳ͍ͷͰऩଋ͕஗͍ shooting ΞϧΰϦζϜ (Fu, 1998, Wu, 2008) (ex. logit ͷ৔߹͸ Yaun, 2010): Daisuke Yoneoka Sparse linear models October 20, 2014 12 / 21
13. l1 ਖ਼ଇԽͷΞϧΰϦζϜ (Cont. 2) Active set ๏ Coordinate descent ͷز͔ͭ·ͱΊͯ࠷దԽ͢Δόʔδϣϯ ͨͩ͠, ͲΕΛݻఆ͠, ͲΕΛ update ͢Δ͔ܾఆ͠ͳ͚Ε͹ͳΒͳ͍ͷͰେม warm starting: ΋͠ λk ≈ λk−1 ͳΒ͹, ˆ w(λk) ͸ ˆ w(λk−1) ͔Β؆୯ʹܭࢉͰ͖Δ Ծʹ͋Δ஋ λ∗ ͷ࣌ͷղ͕஌Γ͍ͨͱ͢Δͱ,warm starting Λ࢖͏ͱ λmax ͔Β୳࢝͠Ίͯ λ∗ ·ͰࢸΔΞϧΰϦζϜͱͳΔ. (Continuation method or homotopy method) ͜Ε͸͍͖ͳΓ λ∗ Λܭࢉ͢Δ (cold starting) ΑΓ λ∗ ͕খ͍͞৔߹, ޮ཰తͳ৔߹͕ଟ͍! LARS (least angle regression and shrinkage): homotopy method ͷҰछ Step 1: λ ͸ y ͱ࠷΋ڧ͘૬ؔ͢Δ̍ͭͷಛ௃ྔ͚͔ͩΒܭࢉͰ͖Δ΋ͷΛॳظ஋ʹ͢Δ Step 2: rk = y − X:,Fk wk Ͱఆٛ͞ΕΔ࢒ࠩʹର͢Δ࠷ॳͷಛ௃ྔͱಉ͚ͩ͡ͷ૬ؔΛ΋ ͭ̎ͭ໨ͷಛ௃ྔ͕ݟ͔ͭΔ·Ͱ λ ΛݮΒ͍ͯ͘͠. (Fk ͸ k ൪໨ͷ active set) least angle Λߟ͑Δ͜ͱͰղੳతʹ࣍ͷ λ ΛܭࢉՄೳ Step 3: શͯͷม਺͕௥Ճ͞ΕΔ·Ͱ܁Γฦ͢ ͜ͷͱ͖,Lasso ͷ solution path Έ͍ͨͳ΋ͷΛඳͨ͘Ίʹ͸ಛ௃ྔΛ ʡ औΓআ͘ ʡ ͜ͱ͕ ՄೳͰ͋Δ͜ͱ͕ඞཁ LAR: LARS ʹࣅ͍ͯΔ͕ಛ௃ྔΛ ʡ औΓআ͘ ʡ ͜ͱΛڐ͞ͳ͍৔߹. (ͪΐ ͬͱ଎ ͘,OLS ͱಉ͡ίετͰ O(NDmin(N, D)) greedy forward search ΍ least square boosting ͱ΋ݺ͹ΕΔ Daisuke Yoneoka Sparse linear models October 20, 2014 13 / 21
14. Proximal and gradient projection methods ತͳ໨తؔ਺ f(θ) = L(θ) + R(θ) Λߟ͑Δ. (L(θ) ͸ϩεؔ਺ͰತͰඍ෼Մೳ, R(θ) ͸ਖ਼ ଇԽ߲Ͱತ͕ͩඍ෼Մೳͱ͸ݶΒͳ͍) ྫ͑͹,f(θ) = R(θ) + 1/2∥θ − y∥2 2 ͷΑ͏ͳͱ͖ (L(θ) = RSS(θ) Ͱܭըߦྻ͕ X = I ͷͱ͖) ತؔ਺ R ͷ proximal operator ͷಋೖ: proxR (y) = argminz R(z) + 1/2∥z − y∥2 2 ௚ײతʹ͸ z Λ y ʹ͚ۙͮͳ͕Β R Λখ͍ͯ͘͘͞͠ iterative ͳ࠷దԽͷதͰ࢖͏৔߹͸, y Λ θk ʹͯ͠࢖͏ Ex. Lasso ໰୊ͷͱ͖ L(θ) = RSS(θ), R(θ) = IC (θ) ͱͰ͖Δ. (ͨͩ͠,C = θ : ∥θ∥1 ≤ B ͔ͭ IC (θ) ≡ 0 if θ ∈ C +∞ otherwise ) ҎԼ, ͲͷΑ͏ʹͯ͠ R ͷ proximal operator Λܭࢉ͢Δ͔Λݟ͍ͯ͘. Daisuke Yoneoka Sparse linear models October 20, 2014 14 / 21
15. Proximal operator Proximal operator ͸ҎԼͷΑ͏ʹදݱՄೳ. (ܭࢉ࣌ؒ͸ O(D) (Duchi, 2008)) R(θ) = λ∥θ∥1 ͷͱ͖: proxR (θ) = soft(θ, λ) (soft-thresholding) R(θ) = λ∥θ∥0 ͷͱ͖: proxR (θ) = hard(θ, √ 2λ) (hard-thresholding) ͨͩ͠,hard(u, a) ≡ uI(|u| > a) R(θ) = IC (θ) ͷͱ͖: proxR (θ) = argminz∈C ∥z − θ∥2 2 = projC (θ) (C ΁ͷࣹӨ) C ͕௒ཱํମͷͱ͖ (i.e., C = θ : lj ≤ θj ≤ uj): projC (θ)j = ⎧ ⎪ ⎨ ⎪ ⎩ lj if θj ≤ lj θj if lj ≤ θj ≤ uj uj if θj ≥ uj C ͕௒ٿͷͱ͖ (i.e., C = θ : ∥θ∥2 ≤ 1): projC (θ)j = ⎧ ⎨ ⎩ θ ∥θ∥2 if ∥θ∥2 > 1 θ otherwise C ͕ 1-norm ٿͷͱ͖ (i.e., C = θ : ∥θ∥1 ≤ 1): projC (θ)j = soft(θ, λ) ͨͩ͠, λ ͸ ∥θ∥1 ≤ 1 ͷͱ͖ 0. ͦΕҎ֎ͷ࣌͸ j=1 −Dmax(|θj | − λ, 0) = 1 ͷղͰఆٛ ͞ΕΔ Daisuke Yoneoka Sparse linear models October 20, 2014 15 / 21
16. Proximal gradient method Proximal operator ΛͲ͏΍ͬͯޯ഑๏ͷͳ͔Ͱ࢖͏͔Λࣔ͢. θ ͷߋ৽ΞϧΰϦζϜ͸ೋ࣍ۙࣅ θk+1 = argminz R(z) + L(θk) + gT k (z − θk) + 1 2tk ∥z − θk∥2 2 (ͨͩ͠,gk = ∇L(θk),tk ͸͜ͷԼ, ࠷ޙͷ߲͸ L ͷϔγΞϯͷۙࣅ ∇2L(θk) ≈ 1 tk I) ⇔ θk+1 = argminz tkR(z) + 1 2 ∥z − uk∥2 2 = proxtkR (uk). (where uk = θk − tkgk) R(θ) = 0 ͷͱ͖: gradient descent ͱ͓ͳ͡ R(θ) = IC (θ) ͷͱ͖: projected gradient descent ͱ͓ͳ͡ R(θ) = λ∥θ∥1 ͷͱ͖: iterative soft thresholding ͱ͓ͳ͡ tk ΋͘͠͸ αk = 1/tk ͷબͼํʹ͍ͭͯ αkI ͕ ∇2L(θ) ͷྑ͍ۙࣅʹͳ͍ͬͯΔͱԾఆ͢Δͱ,αk(θk − θk−1 ≈ gk − gk−1) ͕੒ཱ ͕ͨͬͯ͠ αk = argminα ∥α(θk − θk−1 − (gk − gk−1))∥2 2 = (θk − θk−1)T (gk − gk−1) (θk − θk−1)T (θk − θk−1) Λղ͚͹ྑ͍. (Barzilai-Borwein (BB) or Spectral stepsize) BB stepsize ͱ iterative soft thresholding ͱ homotopy method Λ߹ΘͤΔͱ BPDN (basis pursuit denoising) Λ଎͘ղ͚Δ (SpaRSA ΞϧΰϦζϜ) Daisuke Yoneoka Sparse linear models October 20, 2014 16 / 21
17. Nesteov’s method θk ͷपΓͰ͸ͳ͘ผͷॴͰೋ࣍ۙࣅͯ͠΍Δͱ΋ͬͱ଎͍ proximal gradient descent ͕ಘ ΒΕΔ. θk+1 = proxtkR (φk − tk gk ) gk = ∇L(φk ) φk = θk + k − 1 k + 2 (θk − θk−1 ) Nester’s method ͱ iterative soft thresholding ͱ homotopy method Λ߹ΘͤΔͱ BPDN (basis pursuit denoising) Λ଎͘ղ͚Δ. (FISTA ΞϧΰϦζϜ (fast iterative shrinkage thresholding algorithm)) Daisuke Yoneoka Sparse linear models October 20, 2014 17 / 21
18. Lasso ͷ EM ΞϧΰϦζϜ Laplace ෼෍Λ Gaussian scale mixture (GSM) Ͱදݱ͢Δ. Lap(wj|0, 1/γ) = γ 2 e−γ|wj | = N(wj|0, τ2 j )Ga(τ2 j |1, γ2 2 )dτ2 j ͜ΕΛ༻͍Ε͹, ಉ࣌෼෍͸ p(y, w, τ, σ2|X) = N(y|Xw, σ2IN )N(w|0, Dτ )IG(σ2|aσ, bσ) ⎡ ⎣ j Ga(τ2 j |1, γ2/2) ⎤ ⎦ ∝ (σ2)−N/2 exp − 1 2σ2 ∥y − Xw∥2 2 |D−1/2 τ exp − 1 2 wT Dτ w (σ2)aσ+1 exp(−bσ/σ) j exp(− γ2 2 τ2 j ) ͨͩ͠,Dτ = diag(τ2 j ) Ͱ X ͸ඪ४Խ,y ͸ centered ͞Ε͍ͯΔͷͰ offset ߲͸ແࢹՄೳ. EM ΞϧΰϦζϜͰߟ͑Δ (Figueiredo, 2003) E step: τ2 j , σ2 Λਪఆ͢Δ M step: w ʹؔͯ͠࠷దԽ͢Δ ࣮͸͜ͷ ˆ w ͸ Lasso ਪఆྔͱಉ͡ʹͳΔ Daisuke Yoneoka Sparse linear models October 20, 2014 18 / 21
19. Why EM? l1 ͷ MAP ਪఆͷΞϧΰϦζϜ͸୔ࢁ͋Δͷʹ, ͳΜͰ͋͑ͯ EM ͳͷ͔ʁ probit ΍ robust linear model ͳͲͷਪఆྔΛܭࢉ͠΍͍͢ ෼ࢄʹؔͯ͠ Ga(τ2 j |1, γ2/2) Ҏ֎ͷ prior ΋ߟ͑΍͍͢ Bayesian lasso Λ࢖͑͹ full posterior p(w|D) Λܭࢉ͠΍͍͢ Daisuke Yoneoka Sparse linear models October 20, 2014 19 / 21
20. ໨తؔ਺, E/M step േଇ෇͖ର਺໬౓ؔ਺͸ lc(w) = − 1 2σ2 ∥y − Xw∥2 2 − 1 2 wT Λw + const. (ͨͩ͠,λ = diag( 1 τ2 j ) Ͱਫ਼౓ߦྻ) E step ·ͣ͸ E[ 1 τ2 j |wj] ͷܭࢉΛߟ͑Δ E[ 1 τ2 j |wj ] = − log N(wj |0, τ2 j )p(τ2 j )dτ2 j |wj | Λ௚઀ܭࢉ͢Δ ΋͘͠͸, p(1/τ2 j |w, D) = InverseGaussian γ2 w2 j , γ2 ͱ͢Δͱ,E[ 1 τ2 j |wj ] = γ |wj | ݁ہ,¯ Λ = diag(E[1/τ2 1 ], . . . , E[1/τ2 D ]) ࣍ʹ σ2 ͷਪఆΛߟ͑Δ. posterior ͸ p(σ2|D, w) = IG(aσ + (N)2, bσ + 1 2 (y − X ˆ w)T (y − X ˆ w)) = IG(aN , bN ) ͕ͨͬͯ͠ E[1/σ2] = an bN ≡ ¯ ω Daisuke Yoneoka Sparse linear models October 20, 2014 20 / 21
21. ໨తؔ਺, E/M step (Cont.) M step ˆ w = argmaxw − 1 2 ¯ ω∥y − Xw∥2 2 − 1 2 wT Λw Λܭࢉ͍ͨ͠ ͜Ε͸Ψ΢γΞϯ prior ͷ΋ͱͰ MAP ਪఆ: ˆ w = (σ2 ¯ Λ + XT X)−1XT y ஫ҙ: Sparse ੑΛߟ͍͑ͯΔͷͰ wj ͷ΄ͱΜͲ͕ 0 ⇔ τ2 j ͷ΄ͱΜͲ͕ 0. ͜ͷͱ͖, ¯ Λ ͷٯߦྻͷܭࢉ͕ෆ҆ఆ SVD ෼ղ͕࢖͑Δ! (i.e., X = UDV T ): ˆ w = ΨV (V T ΨV + 1 ¯ ω D−2)−1D−1UT y ɹͨͩ͠, Ψ = ¯ Λ−1 = diag( 1 E[1/τ2 j ] ) = diag( |wj | − log N(wj |0, τ2 j )p(τ2 j )dτ2 j ) Note; Lasso ͷ໨తؔ਺͸ತͳͷͰৗʹ global optim ʹཧ࿦తʹ͸౸ୡՄೳ. ͕ͩ, ਺஋ܭࢉతʹෆՄೳͳ͜ͱ͕ଟ͍! ྫ͑͹,M step Ͱ ˆ wj = 0 ͱͨ͠ͱ͖, E step Ͱ͸ τ2 j = 0 ͱਪఆ݁͠Ռͱͯ͠ ˆ wj = 0 ͱͯ͠ ͠·͍, ͜ͷؒҧ͍͸मਖ਼ෆՄೳʹͳΔ! (Hunter, 2005) Daisuke Yoneoka Sparse linear models October 20, 2014 21 / 21