Derivative-Free Optimization for Hyper-Parameter Tuning in Machine Learning Problems

Transcription

1 Derivative-Free Optimization for Hyper-Parameter Tuning in Machine Learning Problems Hiva Ghanbari Jointed work with Prof. Katya Scheinberg Industrial and Systems Engineering Department Lehigh University ICCOPT Conference, Tokyo, Japan August 2016 ICCOPT 2016 Derivative Free AUC Optimization (1 of 30)

2 Outline Introduction AUC Optimization via DFO-TR Parameter Tuning of Cost-Sensitive LR Summary ICCOPT 2016 Derivative Free AUC Optimization (2 of 30)

3 Outline Introduction AUC Optimization via DFO-TR Parameter Tuning of Cost-Sensitive LR Summary ICCOPT 2016 Derivative Free AUC Optimization (3 of 30)

4 Learning From Imbalanced Data Sets Many real-world machine learning problems are dealing with imbalanced learning data (a) Balanced data set (b) Imbalanced data set ICCOPT 2016 Derivative Free AUC Optimization (4 of 30)

5 Learning From Imbalanced Data Sets ICCOPT 2016 Derivative Free AUC Optimization (5 of 30)

6 Learning From Imbalanced Data Sets Rare positive example, as the minority class, but numerous negative ones, as the majority class In many applications the minority class is the more interesting and important one The rare class has a much higher misclassification cost compared to the majority class ICCOPT 2016 Derivative Free AUC Optimization (6 of 30)

7 Learning From Imbalanced Data Sets Most common methods for dealing with imbalance data sets Methods Advantages Drawbacks Independent of underlying May remove significant Undersampling classifier patterns Independent of underlying classifier Time consuming Oversampling Can be easily implemented May lead to over-fitting Cost sensitive Minimizes the cost of misclassification The misclassification costs are unknown ICCOPT 2016 Derivative Free AUC Optimization (7 of 30)

8 Ranking Quality of Classifier f(x) Consider linear classifier f(x) = w T x + b, where { +1, if f(x i ) > 0, y i = 1, otherwise. Consider function f s(x) = 1 1+e f(x), where f(x) = wt x + b. 1 f s(x) f(x) = w T x + b ICCOPT 2016 Derivative Free AUC Optimization (8 of 30)

9 Ranking Quality of Classifier f(x) Consider linear classifier f(x) = w T x + b, where { +1, if f(x i ) > 0, y i = 1, otherwise. Consider function f s(x) = 1 1+e f(x), where f(x) = wt x + b. 1 f s(x) f(x) = w T x + b In perfect classification, all positive examples are ranked higher than the negative ones f s(x) ICCOPT 2016 Derivative Free AUC Optimization (8 of 30)

10 Receiver Operating Characteristic(ROC) Curve Sorted outputs based on descending value of f s(x) f s(x) Confusion Matrix of a binary classification problem, for a specific threshold to decide when negative turns into positive example +ve (Predicted Class) -ve (Predicted Class) +ve (Actual Class) True Positive (TP) False Negative (FN) -ve (Actual Class) False Positive (FP) True Negative (TN) Various thresholds result in different True Positive Rate = T P T P +F N False Positive Rate = F P F P +T N. and Definition 1 ROC curve presents the tradeoff between the true positive rate and the false positive rate, for all possible thresholds. ICCOPT 2016 Derivative Free AUC Optimization (9 of 30)

11 Area Under ROC Curve (AUC) How we can compare ROC curves? ICCOPT 2016 Derivative Free AUC Optimization (10 of 30)

12 Area Under ROC Curve (AUC) How we can compare ROC curves? Higher AUC = Better classifier ICCOPT 2016 Derivative Free AUC Optimization (10 of 30)

13 Measuring AUC How can we measure the value of AUC of a classifier? Lemma 2 Consider linear classifier f(x) = w T x + b, which has classified positive set S + := {x + i : i = 1,..., N + } from negative set S := {x j : j = 1,..., N }. Then, the associated AUC can be obtained by WMW result where F AUC (w) = I w[w T x i > w T x j ] = N+ N i=1 j=1 Iw[wT x + i > w T x j ]. N + N { +1, if w T x + i > w T x j, 0, otherwise. ICCOPT 2016 Derivative Free AUC Optimization (11 of 30)

14 Probabilistic Interpretation of AUC We can measure the ranking quality of the classifier f(x) through a probability value. F AUC (w) = P (w T X + > w T X ). where X + and X are randomly sampled from S + and S. If two sets S + and S are randomly drawn from distributions D 1 and D 2, then E(F AUC (w)) = P (w T ˆX+ > w T ˆX ), (1) where ˆX + and ˆX are randomly chosen from distributions D 1 and D 2, respectively. ICCOPT 2016 Derivative Free AUC Optimization (12 of 30)

15 Ranking Loss maximizing AUC function F AUC (w) is equivalent to minimizing ranking loss 1 F AUC (w) N+ N i=1 j=1 I[wT x + i wt x j 0] 1 F AUC (w) = 1 +, N + N where as is shown below I[w T x + i wt x j 0] = { +1, if w T x + i wt x j 0, 0, otherwise. I( ) 1 0 w T (x + i x j ) ICCOPT 2016 Derivative Free AUC Optimization (13 of 30)

16 Surrogate Convex Losses The main challenge of minimizing ranking loss 1 F AUC (w) is in its non-smoothness and discontinuousness owned from indicator function I( ). ICCOPT 2016 Derivative Free AUC Optimization (14 of 30)

17 Surrogate Convex Losses The main challenge of minimizing ranking loss 1 F AUC (w) is in its non-smoothness and discontinuousness owned from indicator function I( ). Exponential Loss: I exp(w, x + i x j ) = e wt (x + i x j ) w T (x + i x j ) ICCOPT 2016 Derivative Free AUC Optimization (14 of 30)

18 Surrogate Convex Losses The main challenge of minimizing ranking loss 1 F AUC (w) is in its non-smoothness and discontinuousness owned from indicator function I( ). Logistic Loss: w T (x + i x j ) I log (w, x + i x j ) = log(1 + e wt (x + i x j ) ), ICCOPT 2016 Derivative Free AUC Optimization (15 of 30)

19 Surrogate Convex Losses The main challenge of minimizing ranking loss 1 F AUC (w) is in its non-smoothness and discontinuousness owned from indicator function I(.). Hinge Loss: I hinge (w, x + i x j ) = max{0, 1 wt (x + i x j )} w T (x + i x j ) ICCOPT 2016 Derivative Free AUC Optimization (16 of 30)

20 Outline Introduction AUC Optimization via DFO-TR Parameter Tuning of Cost-Sensitive LR Summary ICCOPT 2016 Derivative Free AUC Optimization (17 of 30)

21 Algorithmic Framework of DFO-TR Algorithm 1 Trust Region Based Derivative Free Optimization 0. Initializations. Define interpolation set X and compute corresponding function values F X. 1. Build the model. Build Q k (x) as a quadratic approximation of function f. 2. Minimize the model within the trust region. Find ˆx k such that Q k (ˆx k ) := min x Bk Q k (x), where B k = {x : x x k k }. Compute function f(ˆx k ) and the ratio ρ k = f(x k) f(ˆx k ) Q k (x k ) Q k (ˆx k ). 3. Update the interpolation set. 4. Update the trust region radius. ICCOPT 2016 Derivative Free AUC Optimization (18 of 30)

22 AUC, a Smooth Function of w in Expectation Theorem 3 If two d dimensional random vectors ˆX 1 and ˆX 2 have a joint multivariate normal distribution, such that ( ) ( ) ( ) ˆX1 µ1 Σ11 Σ N (µ, Σ), where, µ =, and Σ = 12, ˆX 2 µ 2 Σ 21 Σ 22 then the marginal distributions of ˆX1 and ˆX 2 are normal distributions with the following properties ˆX 1 N (µ 1, Σ 11 ), ˆX2 N (µ 2, Σ 22 ). ICCOPT 2016 Derivative Free AUC Optimization (19 of 30)

23 AUC, a Smooth Function of w in Expectation Theorem 4 Consider two random vectors ˆX 1 and ˆX 2, then for any vector w R d, we have Z = w T ( ˆX 1 ˆX 2 ) N (µ Z, σ 2 Z ), where µ Z = w T (µ 1 µ 2 ), σ 2 Z = wt (Σ 11 + Σ 22 Σ 12 Σ 21 ) w. Theorem 5 If two random vectors ˆX 1 and ˆX 2, respectively from the positive and the negative classes, have a joint normal distribution, then the expected value of AUC function can be defined as ( ) µz E(F AUC (w)) = φ, σ Z where φ is the cumulative function of the standard normal distribution, so that φ(x) = e 1 2 x2 /2π, for x R. ICCOPT 2016 Derivative Free AUC Optimization (20 of 30)

24 Computational Result (DFO-TR vs. Hinge) Performance of algorithms based on the average value of AUC and number of function evaluations, Algorithm sonar(d = 60, N = 208) fourclass(d = 2, N = 862) AUC fevals AUC fevals Hinge ± ± DFO-TR ± ± Algorithm svmguide1(d = 4, N = 3089) magic04(d = 10, N = 4020) AUC fevals AUC fevals Hinge ± ± DFO-TR ± ± Algorithm svmguide3(d = 22, N = 1243) shuttle(d = 9, N = 3500) AUC fevals AUC fevals Hinge ± ± DFO-TR ± ± Algorithm segment(d = 19, N = 2310) ijcnn1(d = 22, N = 4691) AUC fevals AUC fevals Hinge ± ± DFO-TR ± ± Algorithm letter(d = 16, N = 5000) poker(d = 10, N = 5010) AUC fevals AUC fevals Hinge ± ± DFO-TR ± ± ICCOPT 2016 Derivative Free AUC Optimization (21 of 30)

25 Computational Result (Stochastic DFO-TR) Stochastic vs Deterministic DFO-TR ICCOPT 2016 Derivative Free AUC Optimization (22 of 30)

26 Outline Introduction AUC Optimization via DFO-TR Parameter Tuning of Cost-Sensitive LR Summary ICCOPT 2016 Derivative Free AUC Optimization (23 of 30)

27 Cost-Sensitive Logistic Regression Problem Biasing the classifier toward the minority class where F (w) := 1 N N C(y i ) log(1 + exp( y i w T x i )) + λ 2 w 2, i=1 C(y i ) = { C +, if y i > 0, C, otherwise. ICCOPT 2016 Derivative Free AUC Optimization (24 of 30)

28 Cost-Sensitive Logistic Regression Problem Biasing the classifier toward the minority class where GOAL F (w) := 1 N N C(y i ) log(1 + exp( y i w T x i )) + λ 2 w 2, i=1 C(y i ) = { C +, if y i > 0, C, otherwise. Find the best value of costs C + and C and regularization parameter λ, to achieve a linear classifier f(x) = w T x + b with maximum value of AUC. ICCOPT 2016 Derivative Free AUC Optimization (24 of 30)

29 AUC as a Function of {C +, C, λ} AUC is a function of w F AUC (w) = N+ N i=1 j=1 Iw[wT x i > w T x j ]. N + N The vector w is a function of the parameters {C +, C, λ}, so that F (w) := 1 N N C(y i ) log(1 + exp( y i w T x i )) + λ 2 w 2, i=1 ICCOPT 2016 Derivative Free AUC Optimization (25 of 30)

30 AUC as a Function of {C +, C, λ} AUC is a function of w F AUC (w) = N+ N i=1 j=1 Iw[wT x i > w T x j ]. N + N The vector w is a function of the parameters {C +, C, λ}, so that F (w) := 1 N N C(y i ) log(1 + exp( y i w T x i )) + λ 2 w 2, i=1 F AUC (w) is a non-smooth function of the parameters {C +, C, λ}. ICCOPT 2016 Derivative Free AUC Optimization (25 of 30)

31 w a Smooth Function of {C +, C, λ} Theorem 6 In the cost sensitive logistic regression, the vector w is a smooth function of non-zero parameters C +,C, and λ if matrix M and added matrix M v have the same rank (one special case which satisfies this condition is when the matrix M is a full rank matrix) N + exp(y i w T x i ) M = (1 + exp(y i w T x i )) 2 (x ix T i ), i=1 v = 1 C + N + i=1 y i x i 1 + exp(y i w T x i ). (2) Similar condition is required when M and v are constructed based on the negative class. ICCOPT 2016 Derivative Free AUC Optimization (26 of 30)

32 Computational Result In this section we compare the following algorithms, Cost-Sensitive Logistic Regression based on Class Distribution Ratio Cost-Sensitive Logistic Regression based on Optimized Ratio Algorithm sonar fourclass svmguide1 magic04 CSLR-CDR CSLR-OR Algorithm diabetes german a9a svmguide3 CSLR-CDR CSLR-OR Algorithm connect-4 shuttle HAPT segment CSLR-CDR CSLR-OR Algorithm mnist ijcnn1 satimage vowel CSLR-CDR CSLR-OR Algorithm poker letter w1a w8a CSLR-CDR CSLR-OR performance of algorithms based on average value of AUC ICCOPT 2016 Derivative Free AUC Optimization (27 of 30)

33 Outline Introduction AUC Optimization via DFO-TR Parameter Tuning of Cost-Sensitive LR Summary ICCOPT 2016 Derivative Free AUC Optimization (28 of 30)

34 Summary and Future Study The main challenge of optimizing AUC is in its non-smoothness and discontinuousness Directly Optimizing AUC function via trust region based derivative free optimization Tuning parameters of cost-sensitive logistic regression to obtain a classifier with maximum AUC value ICCOPT 2016 Derivative Free AUC Optimization (29 of 30)

35 References A. R. Conn, K. Scheinberg, and L. N.Vicente, Introduction To Derivative-Free Optimization, SIAM J, M. Menickelly R. Chen and K. Scheinberg, Stochastic optimization using a trust-region method and random models, C. X. Ling, J. Huang, and H. Zhang, AUC: a Statistically Consistent and more Discriminating Measure than Accuracy, J.A. Hanley, B.J. McNeil, The meaning and use of the area under a receiver operating characteristic (ROC) curve, Radiology, 143:29?36, ICCOPT 2016 Derivative Free AUC Optimization (30 of 30)