Direct Convex Relaxations of Sparse SVM Antoni B. Chan, Nuno Vasconcelos, and Gert R. G. Lanckriet The 24th International Conference on Machine Learning (ICML 2007) Presented by Shuiwang Ji
Outline Introduction; Quadratically Constrained Quadratic Programming (QCQP) formulation; Semidefinite Programming (SDP) formulation; Experiments;
Sparsity of SVM x1, …, xd SVM is sparse w.r.t. data points, but not sparse w.r.t. features.
Motivations & Related Work Features may be noisy, redundant; Sparsity enhance interpretability; Sparse PCA (Zou et al. & d'Aspremont et al.); Sparse Eigen Methods by D.C. Programming (ICML07);
An Example
Vector Norm Number of nonzero entries in x
C-SVM Primal and Dual 2-norm
LP-SVM Primal and Dual 1-norm
Convex QCQP Relaxation
Interpretations of QCQP-SSVM Problem 6 and 7 are equivalent; QCQP-SSVM is a combination of C-SVM and LP-SVM, 1-norm encourages sparsity and 2-norm encourages large margin;
QCQP-SSVM Dual
QCQP-SSVM QCQP-SSVM automatically learns an adaptive soft-threshold on the original SVM hyperplane.
SDP Relaxation
SDP-SSVM Dual The optimal weighting matrix increases the influence of the relevant features while demoting the less relevant features; SDP-SSVM learns a weighting on the inner product such that the hyperplane in the feature space is sparse.
Results on Synthetic Data
Results on 15 UCI data sets