Model Selection via Bilevel Optimization Kristin P Bennett

Model Selection via Bilevel Optimization Kristin P. Bennett, Jing Hu, Xiaoyun Ji, Gautam Kunapuli and Jong-Shi Pang Department of Mathematical Sciences Rensselaer Polytechnic Institute Troy, NY

Convex Machine Learning Convex optimization approaches to machine learning has been major obsession of machine learning for last ten years. But are the problems really convex?

Outline n n The myth of convex machine learning Bilevel Programming Model Selection n n Regression Classification Extensions to other machine learning tasks Discussion

Modeler’s Choices Data CONVEX! Function Loss/Regularization Optimization Algorithm w

Many Hidden Choices n Data: n n n Function Family: n n Variable Selection Scaling Feature Construction Missing Data Outlier removal linear, kernel (introduces kernel parameters) Optimization model n n n loss function regularization Parameters/Constraints

Data Function Loss/Regularization Cross-Validation Strategy Generalization Error NONCONVEX Cross-Validation: C, ε, [X, y] Optimization Algorithm w

How does modeler make choices? n n n Best training set error Experience/policy Estimate of generalization error n n n Cross-validation Bounds Optimize generalization error estimate n n Fiddle around. Grid Search Gradient methods Bilevel Programming

Splitting Data for T-fold CV

CV via Grid Search For every C, ε • For every validation set, Solve model on corresp. training set, and to estimate loss for ε C • Estimate generalization error for C, ε Return best values for C, ε Make final model using C, ε

CV as Continuous Optimization Problem n Bilevel Program for T folds Outer-level validation problem T inner-level training problems n Prior Approaches: Golub et al. , 1979, Generalized Cross-Validation for one parameter in Ridge Regression

Benefit: More Design Variables Add feature box constraint: inner-level problems. in the

-insensitive Loss Function

Inner-level Problem for t-th Fold

Optimality (KKT) conditions for fixed

Key Transformation n KKT for the inner level training problems are necessary and sufficient Replace lower level problems by their KKT Conditions Problem becomes an Mathematical Programming Problem with Equilibrium Constraints (MPEC)

Bilevel Problem as MPEC Replace T inner-level problems with corresponding optimality conditions

MPEC to NLP via Inexact Cross Validation Relax “hard” equilibrium constraints to “soft” inexact constraints tol is some user-defined tolerance.

Solvers Strategy: Proof of concept using nonlinear general purpose solvers from NEOS on NLP FILTER, SNOPT Sequential Quadratic Programming Methods FILTER results almost always better. Many possible alternatives: Integer Programming Branch and Bound Lagrangian Relaxations

Computational Experiments: DATA Synthetic n (5, 10, 15)-D Data with Gaussian and Laplacian noise and (3, 7, 10) relevant features. n NLP: 3 -fold CV n Results: 30 to 90 train, 1000 test points, 10 trials QSAR/Drug Design n 4 datasets, 600+ dimensions reduced to 25 top principal components. NLP: 5 -fold CV n Results: 40 – 100 train, rest test, 20 trials

Cross-validation Methods Compared n n n Unconstrained Grid: Try 3 values each for C, ε Constrained Grid: Try 3 values each for C, ε, and {0, 1} for each component of Bilevel/FILTER: Nonlinear program solved using off-the-shelf SQP algorithm via NEOS

15 -D Data: Objective Value

15 -D Data: Computational Time

15 -D Data: TEST MAD

QSAR Data: Objective Value

QSAR Data: Computation Time

QSAR Data: TEST MAD

Classification Cross Validation Given sample data from two classes. Find classification function that minimizes out-ofsample estimate of classification error 1 -1

Lower level - SVM n n n Define parallel planes Minimize points on wrong side Maximize margin of separation

Lower Level Loss Function: Hinge Loss Measures distance of points that violate the appropriate hyperplane constraints,

Lower Level Problem: SVC with box

Inner-level KKT Conditions

Outer-level Loss Functions n Misclassification Minimization Loss (MM) n n n Loss function used in classical CV Loss = 1, if validation pt misclassified, 0, otherwise (computed using step function, ) Hinge Loss (HL) n n Both inner and outer levels use same loss function Loss = distance from (computed using max function, )

Hinge Loss is Convex Approx. of Misclassification Minimization Loss

Hinge Loss Bilevel Program (Bilevel. HL) n Replace max in outer level objective with convex constraints n Replace inner-level problems with KKT conditions

Hinge Loss MPEC

Misclassification Min. Bilevel Program (Bilevel. MM) Misclassifications are counted using the step function, defined component wise for a n-vector as

The Step Function Mangasarian (1994) showed that any solution, , to the LP satisfies

Misclassifications in the Validation Set n Validation point misclassified when the sign of is negative i. e. , n This can be recast for all validation points (within the t-th fold) as

Misclassification Minimization Bilevel Program (revisited) Outer-level average misclassification minimization Inner-level problems to determine misclassified validation points Inner-level training problems

Misclassification Minimization MPEC

Inexact Cross Validation NLP n n n Both Bilevel. HL and Bilevel. MM MPECs are transformed to NLP by relaxing equilibrium constraints (inexact CV) Solved using FILTER on NEOS These are compared with classical cross validation: unconstrained and constrained grid.

Experiments: Data sets 3 -fold cross validation for model selection Average results for 20 train test splits

Computational Time

Training CV Error

Testing Error

Number of Variables

Progress n n n Cross Validation is a bilevel problem solvable by continuous optimization methods Off-the-shelf NLP algorithm – FILTER solved classification and regression Bilevel Optimization extendable to many Machine Learning problems

Extending Bilevel Approach to other Machine Learning Problems n n n Kernel Classification/Regression Variable Selection/Scaling Multi-task Learning Semi-supervised Learning Generative methods

Semi-supervised Learning n n n Have labeled data, and unlabeled data Treat missing labels, , as design variables in the outer level Lower level problems are still convex

Semi-supervised Regression Outer level minimizes error on labeled data to find optimal parameters and labels -insensitive loss on labeled data in inner level -insensitive loss on unlabeled data in inner level Inner level regularization

Discussion n New capacity offers new possibilities: Outer level objectives? Inner level problem? classification, ranking, semi-supervised, missing values, kernel selection, variable selection, … n Need special purpose algorithms for greater efficiency, scalability, robustness This work was supported by Office of Naval Research Grant N 00014 -06 -1 -0014.

Experiments: Bilevel CV Procedure n n Run Bilevel. MM/Bilevel. HL to compute optimal parameters, Drop descriptors with small Create model on all training data using Compute test error on hold-out set

Experiments: Grid Search CV Procedure n n Unconstrained Grid: Try 6 values for on a log 10 scale Constrained Grid: Try 6 values for and {0, 1} for each component of (perform RFE if necessary) Create model on all training data using optimal grid point Compute test error on hold-out set

Extending Bilevel Approach to other Machine Learning Problems n n n Kernel Classification/Regression Different Regularizations (L 1, elastic nets) Enhanced Feature Selection Multi-task Learning Semi-supervised Learning Generative methods

Enhanced Feature Selection n n Assume at most descriptors allowed Introduce outer-level constraint (with n n counting the non-zero elements of Rewrite constraint, observing that Get additional conditions for )

Kernel Bilevel Discussion n Pros n n n Performs model selection in feature space Performs feature selection in input space Cons n n Highly nonlinear model Difficult to solve

Kernel Classification (MPEC form)

Is it okay to do 3 -folds

Applying the “kernel trick” n Drop the box constraint, Eliminate from the optimality conditions n Replace n with an appropriate

Feature Selection with Kernels n Parameterize the kernel with such that if the n-th descriptor vanishes from Linear kernel n Polynomial kernel n Gaussian kernel n

Kernel Regression (Bilevel form)