Classification. Jordan Boyd-Graber University of Maryland WEIGHTED MAJORITY. Slides adapted from Mohri. Jordan Boyd-Graber UMD Classification 1 / 13

Size: px

Start display at page:

Download "Classification. Jordan Boyd-Graber University of Maryland WEIGHTED MAJORITY. Slides adapted from Mohri. Jordan Boyd-Graber UMD Classification 1 / 13"

Chrystal Paul
5 years ago
Views:

1 Classification Jordan Boyd-Graber University of Maryland WEIGHTED MAJORITY Slides adapted from Mohri Jordan Boyd-Graber UMD Classification 1 / 13

2 Beyond Binary Classification Before we ve talked about combining weak predictor (boosting) Jordan Boyd-Graber UMD Classification 2 / 13

3 Beyond Binary Classification Before we ve talked about combining weak predictor (boosting) What if you have strong predictors? Jordan Boyd-Graber UMD Classification 2 / 13

4 Beyond Binary Classification Before we ve talked about combining weak predictor (boosting) What if you have strong predictors? How do you make inherently binary algorithms multiclass? How do you answer questions like ranking? Jordan Boyd-Graber UMD Classification 2 / 13

5 General Online Setting For t = 1 to T : Get instance x t X Predict ŷ t Y Get true label y t Y Incur loss L(ŷ t, y t ) Classification: Y = {0,1}, L(y, y ) = y y Regression: Y, L(y, y ) = (y y ) 2 Jordan Boyd-Graber UMD Classification 3 / 13

6 General Online Setting For t = 1 to T : Get instance x t X Predict ŷ t Y Get true label y t Y Incur loss L(ŷ t, y t ) Classification: Y = {0,1}, L(y, y ) = y y Regression: Y, L(y, y ) = (y y ) 2 Objective: Minimize total loss t L(ŷ t, y t ) Jordan Boyd-Graber UMD Classification 3 / 13

7 Prediction with Expert Advice For t = 1 to T : Get instance x t X and advice a t,i Y,i [1,N ] Predict ŷ t Y Get true label y t Y Incur loss L(ŷ t, y t ) Jordan Boyd-Graber UMD Classification 4 / 13

8 Prediction with Expert Advice For t = 1 to T : Get instance x t X and advice a t,i Y,i [1,N ] Predict ŷ t Y Get true label y t Y Incur loss L(ŷ t, y t ) Objective: Minimize regret, i.e., difference of total loss vs. best expert Regret(T ) = L(ŷ t, y t ) min L(a t,i, y t ) (1) i t t Jordan Boyd-Graber UMD Classification 4 / 13

9 Mistake Bound Model Define the maximum number of mistakes a learning algorithm L makes to learn a concept c over any set of examples (until it s perfect). M L (c ) = max x 1,...,x T mistakes(l, c ) (2) For any concept class C, this is the max over concepts c. M L (C ) = max c C M L(c ) (3) Jordan Boyd-Graber UMD Classification 5 / 13

10 Mistake Bound Model Define the maximum number of mistakes a learning algorithm L makes to learn a concept c over any set of examples (until it s perfect). M L (c ) = max x 1,...,x T mistakes(l, c ) (2) For any concept class C, this is the max over concepts c. M L (C ) = max c C M L(c ) (3) In the expert advice case, assumes some expert matches the concept (realizable) Jordan Boyd-Graber UMD Classification 5 / 13

11 Halving Algorithm H 1 H ; for t 1...T do Receive x t ; ŷ t Majority(H t, a t, x t ); Receive y t ; if ŷ t y t then H t +1 {a H t : a (x t ) = y t }; return H T +1Algorithm 1: The Halving Algorithm (Mitchell, 1997) Jordan Boyd-Graber UMD Classification 6 / 13

12 Halving Algorithm Bound (Littlestone, 1998) For a finite hypothesis set M Halving(H ) lg H (4) After each mistake, the hypothesis set is reduced by at least by half Jordan Boyd-Graber UMD Classification 7 / 13

13 Halving Algorithm Bound (Littlestone, 1998) For a finite hypothesis set M Halving(H ) lg H (4) After each mistake, the hypothesis set is reduced by at least by half Consider the optimal mistake bound opt(h ). Then VC(H ) opt(h ) M Halving(H ) lg H (5) For a fully shattered set, form a binary tree of mistakes with height VC(H ) Jordan Boyd-Graber UMD Classification 7 / 13

14 Halving Algorithm Bound (Littlestone, 1998) For a finite hypothesis set M Halving(H ) lg H (4) After each mistake, the hypothesis set is reduced by at least by half Consider the optimal mistake bound opt(h ). Then VC(H ) opt(h ) M Halving(H ) lg H (5) For a fully shattered set, form a binary tree of mistakes with height VC(H ) What about non-realizable case? Jordan Boyd-Graber UMD Classification 7 / 13

15 Weighted Majority (Littlestone and Warmuth, 1998) for i 1...N do w 1,i 1; for t 1...T do Receive x t ; ŷ t 1 a t,i =1 w t a t,i =0 w t Receive y t ; if ŷ t y t then for i 1...N do if a t,i y t then w t +1,i β w t,i ; else wt +1,i w t,i ; Weights for every expert Classifications in favor of side with higher total weight (y {0,1}) that are wrong get their weights decreased (β [0,1]) If you re right, you stay unchanged return w T +1 Jordan Boyd-Graber UMD Classification 8 / 13

16 Weighted Majority (Littlestone and Warmuth, 1998) for i 1...N do w 1,i 1; for t 1...T do Receive x t ; ŷ t 1 a t,i =1 w t a t,i =0 w t Receive y t ; if ŷ t y t then for i 1...N do if a t,i y t then w t +1,i β w t,i ; else wt +1,i w t,i ; Weights for every expert Classifications in favor of side with higher total weight (y {0,1}) that are wrong get their weights decreased (β [0,1]) If you re right, you stay unchanged return w T +1 Jordan Boyd-Graber UMD Classification 8 / 13

17 Weighted Majority (Littlestone and Warmuth, 1998) for i 1...N do w 1,i 1; for t 1...T do Receive x t ; ŷ t 1 a t,i =1 w t a t,i =0 w t Receive y t ; if ŷ t y t then for i 1...N do if a t,i y t then w t +1,i β w t,i ; else wt +1,i w t,i ; Weights for every expert Classifications in favor of side with higher total weight (y {0,1}) that are wrong get their weights decreased (β [0,1]) If you re right, you stay unchanged return w T +1 Jordan Boyd-Graber UMD Classification 8 / 13

18 Weighted Majority (Littlestone and Warmuth, 1998) for i 1...N do w 1,i 1; for t 1...T do Receive x t ; ŷ t 1 a t,i =1 w t a t,i =0 w t Receive y t ; if ŷ t y t then for i 1...N do if a t,i y t then w t +1,i β w t,i ; else wt +1,i w t,i ; Weights for every expert Classifications in favor of side with higher total weight (y {0,1}) that are wrong get their weights decreased (β [0,1]) If you re right, you stay unchanged return w T +1 Jordan Boyd-Graber UMD Classification 8 / 13

19 Weighted Majority Let mt be the number of mistakes made by WM until time t Let m t be the best expert s mistakes until time t N is the number of experts m t logn + m t log 1 β log 2 1+β (6) Thus, mistake bound is O (logn ) plus the best expert Halving algorithm β = 0 Jordan Boyd-Graber UMD Classification 9 / 13

20 Proof: Potential Function Potential function is the sum of all weights Φ t w t,i (7) We ll create sandwich of upper and lower bounds i Jordan Boyd-Graber UMD Classification 10 / 13

21 Proof: Potential Function Potential function is the sum of all weights Φ t w t,i (7) We ll create sandwich of upper and lower bounds For any expert i, we have lower bound i Φ t w t,i = β m t,i (8) Jordan Boyd-Graber UMD Classification 10 / 13

22 Proof: Potential Function Potential function is the sum of all weights Φ t w t,i (7) We ll create sandwich of upper and lower bounds For any expert i, we have lower bound i Φ t w t,i = β m t,i (8) Weights are nonnegative, so i w t,i w t,i Jordan Boyd-Graber UMD Classification 10 / 13

23 Proof: Potential Function Potential function is the sum of all weights Φ t w t,i (7) We ll create sandwich of upper and lower bounds For any expert i, we have lower bound i Φ t w t,i = β m t,i (8) Each error multiplicatively reduces weight by β Jordan Boyd-Graber UMD Classification 10 / 13

24 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 (9) Jordan Boyd-Graber UMD Classification 11 / 13

25 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 (9) Half (at most) of the experts by weight were right Jordan Boyd-Graber UMD Classification 11 / 13

26 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 (9) Half (at least) of the experts by weight were wrong Jordan Boyd-Graber UMD Classification 11 / 13

27 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 = 1 + β 2 Φ t (9) Jordan Boyd-Graber UMD Classification 11 / 13

28 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 = 1 + β 2 Φ t (9) Initially potential function sums all weights, which start at 1 Φ 1 = N (10) Jordan Boyd-Graber UMD Classification 11 / 13

29 Proof: Potential Function (Upper Bound) If an algorithm makes an error at round t Φ t +1 Φ t 2 + βφ t 2 = 1 + β 2 Φ t (9) Initially potential function sums all weights, which start at 1 Φ 1 = N (10) After mt mistakes after T rounds 1 + β mt Φ T N (11) 2 Jordan Boyd-Graber UMD Classification 11 / 13

30 Weighted Majority Proof Put the two inequalities together, using the best expert β m T 1 + β mt ΦT N (12) 2 Jordan Boyd-Graber UMD Classification 12 / 13

31 Weighted Majority Proof Put the two inequalities together, using the best expert β m T 1 + β mt ΦT N (12) 2 Take the log of both sides 1 + β m T logβ logn + m T log 2 (13) Jordan Boyd-Graber UMD Classification 12 / 13

32 Weighted Majority Proof Put the two inequalities together, using the best expert β m T 1 + β mt ΦT N (12) 2 Take the log of both sides 1 + β m T logβ logn + m T log 2 (13) Solve for mt m T logn + m T log 1 β (14) log 2 1+β Jordan Boyd-Graber UMD Classification 12 / 13

33 Weighted Majority Recap Simple algorithm No harsh assumptions (non-realizable) Depends on best learner Downside: Takes a long time to do well in worst case (but okay in practice) Solution: Randomization Jordan Boyd-Graber UMD Classification 13 / 13

Online Learning. Jordan Boyd-Graber. University of Colorado Boulder LECTURE 21. Slides adapted from Mohri

Online Learning. Jordan Boyd-Graber. University of Colorado Boulder LECTURE 21. Slides adapted from Mohri Online Learning Jordan Boyd-Graber University of Colorado Boulder LECTURE 21 Slides adapted from Mohri Jordan Boyd-Graber Boulder Online Learning 1 of 31 Motivation PAC learning: distribution fixed over