Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Incomplete
-
2.2.0
-
None
Description
Currently LinearSVC in Spark only supports OWLQN as the optimizer ( check https://issues.apache.org/jira/browse/SPARK-14709). I made comparison between LBFGS and OWLQN on several public dataset and found LBFGS converges much faster for LinearSVC in most cases.
The following table presents the number of training iterations and f1 score of both optimizers until convergence
| Dataset | LBFGS with hinge | OWLQN with hinge | LBFGS with squared_hinge |
|---|---|---|---|
| news20.binary | 31 (0.99) | 413(0.99) | 185 (0.99) |
| mushroom | 28(1.0) | 170(1.0) | 24(1.0) |
| madelon | 143(0.75) | 8129(0.70) | 823(0.74) |
| breast-cancer-scale | 15(1.0) | 16(1.0) | 15 (1.0) |
| phishing | 329(0.94) | 231(0.94) | 67 (0.94) |
| a1a(adult) | 466 (0.87) | 282 (0.87) | 77 (0.86) |
| a7a | 237 (0.84) | 372(0.84) | 69(0.84) |
data source: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html
training code: new LinearSVC().setMaxIter(10000).setTol(1e-6)
LBFGS requires less iterations in most cases (except for a1a) and probably is a better default optimizer.
Attachments
Issue Links
- relates to
-
SPARK-20503 ML 2.2 QA: API: Python API coverage
-
- Resolved
-
-
SPARK-20348 Support squared hinge loss (L2 loss) for LinearSVC
-
- Resolved
-
- links to
