Ensembling: train the pipelines and the metalearner on different data
オープン
enhancement
performance
- 主要言語
- Python
- スター
- 850
- フォーク
- 96
- PR マージ指標
- 30日以内にマージされた PR はありません
説明
This issue tracks:
* During automl search, first train each pipeline on full 80% automl training split, then generate predictions on 20% ensemble split, and use those predictions to train the metalearner on that 20% ensemble split using CV.
* For refitting best pipeline, if the best pipeline was non-ensemble, train on entire 100% (should already be the case). If best pipeline was ensemble, same as above: first train each pipeline on full 80% automl training split, then generate predictions on 20% ensemble split, and use those predictions to train the metalearner on that 20% ensemble split.
コントリビューションガイド
評価
この issue はまだ評価されていません。