Method | Regularizer (parameters) | Comments on computational efficiency |
---|---|---|
Explicit regularization | ||
Information criteria (4), e.g. AIC [24], BIC [25], EBIC [26] | \(\ell _0\)-penalty (\(\lambda\)) | Best subset selection not efficient for high-dimensional problems. Heuristic optimization [27,28,29] or mixed-integer optimization [30] can be used. |
Lasso [1] | \(\ell _1\)-penalty (\(\lambda\)) | Computationally efficient convex relaxation of \(\ell _0\)-type problem. |
\(\ell _1\)-penalty (\(\lambda , \gamma\)) | Combination of \(\ell _1\)-regularized and unregularized (restricted least squares) estimator. Computationally efficient but tuning more costly than for lasso. | |
Elastic net [4] | \(\ell_1\)-/\(\ell_2\)-penalty (\(\lambda , \alpha\)) | Combination of \(\ell _1\)- and \(\ell _2\)-penalties. Computationally efficient but tuning more costly than for lasso. |
Implicit regularization | ||
\(L_2\)Boosting [10] (Algorithm 1) | Early stopping (\(m_{{\text {stop}}}\)) | Tuning of stopping iteration \(m_{{\text {stop}}}\) via resampling leads to implicit regularization. |
Twin boosting [31] | Early stopping (\(m_1,m_2\)) | Two-stage approach using \(L_2\)Boosting estimates as weights in second stage of \(L_2\)Boosting. Tuning more costly than for single-stage \(L_2\)Boosting. |
Flexible (PFER) | Computationally intensive ensemble approach, applying e.g. lasso or \(L_2\)Boosting multiple times on subsamples. Provides control over false positives (PFER). | |
New Subspace Boosting: SubBoost (Algorithm 2), RSubBoost and AdaSubBoost (Algorithm 3) | Automatic stopping (\(\Phi\)) | Multivariable base-learners with double checking via selection criterion \(\Phi\) for automatic stopping. Randomized preselection of base-learners for scalability. For further hyper-parameters see Algorithms 2 and 3 and the Additional file 1 for their effects on the computational efficiency. |