Research map / Learning theory and algorithms / Matrix and tensor methods
A Two-Level Decomposition Framework Exploiting First and Second Order Information for SVM Training Problems
Giulio Galvan, Matteo Lapucci, Chih-Jen Lin, Marco Sciandrone
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use two-level decomposition (inner SMO on working set q=4) instead of plain SMO | SMO-WSS1/WSS2 · 17 data sets, 25 problems each | total runtime: a9a 7708.22 -> 3003.09; covtype 2213224.58 -> 518847.84; cod-rna 26490.30 -> 1999.17 | Improved |
| use TLD-ISMO on HIGGS subset | SMO · HIGGS-subset | completion within time limit: 1011667 (3) vs 1166775 (5) | Improved |
| add second-order WSS2 instead of WSS1 (ablation) | SMO · iterations | number of iterations | Improved |
| add second-order WSS2 instead of WSS1 (ablation) | SMO · computing time | runtime: in about a half of the situations WSS1 led to a faster training | No clear effect |
| increase working set size q (ablation) | Algorithm 1 with extended-WSS1 · 150 SVM problems | number of iterations | Improved |
| increase working set size q (ablation) | Algorithm 1 with extended-WSS1 · 150 SVM problems | kernel evaluations | Hurt |
| increase working set size to non-extreme values (ablation) | Algorithm 1 · 150 SVM problems | runtime | Improved |
| double working set size (ablation) | TLD-ISMO · small data sets | number of iterations: slightly less than a half | Improved |
| use WSS-MIX (mixed first and second order selection) (ablation) | TLD-ISMO q=4 · 175 Gaussian kernel SVM problems | runtime | Improved |
| use second-order information on large data sets (ablation) | TLD-ISMO · rcv1, real-sim | runtime | Improved |
| replace SMO-MVP with Dai-Fletcher subproblem solver at q=50 (ablation) | Two-level decomposition · SVM problems | runtime | Improved |
| replace GVPM/Dai-Fletcher by SMO-MVP as inner solver for small q (ablation) | Two-level decomposition · SVM problems | runtime: limited q | Improved |
Closest papers by meaning
- Solving L1-regularized SVMs and Related Linear Programs: Revisiting the Effectiveness of Column and Constraint Generation (JMLR 2022 · similarity 0.88)
- Faster Algorithms for Structured Linear and Kernel Support Vector Machines (ICLR 2025 · similarity 0.88)
- An exact solver for the Weston-Watkins SVM subproblem (ICML 2021 · similarity 0.88)
- Fast Algorithms for Linear and Kernel SVM+ (CVPR 2016 · similarity 0.88)
- Scaling Up Sparse Support Vector Machines by Simultaneous Feature and Sample Reduction (ICML 2017 · similarity 0.87)
- Local Duality for Sparse Support Vector Machines (TPAMI 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
