Preprocessing Cheat Sheet¶
This page summarizes how dataset defaults and model metadata combine in the current pipeline implementation.
Dataset-side defaults¶
| Dataset type | Default variance_threshold |
Default svd |
Notes |
|---|---|---|---|
descriptor |
disabled | disabled | Descriptor cleanup focuses on string parsing, dropping all-NaN columns, and median imputation. |
fingerprint |
enabled at 0.01 |
enabled at 10 components |
Fingerprints are coerced to numeric and cast to float32. |
Model-side switches¶
Model metadata comes from the selected hyperparameter YAML:
requires_scaler: controls whether the configured scaler is inserteduse_smote: controls whether the configured sampler is inserted
Current quick-grid models:
adaboostclassifiergradientboostingclassifierbernoullinbkneighborsclassifierlogisticregressionrandomforestclassifierextratreesclassifiersvcsgdclassifierxgbclassifierlgbmclassifier
Actual order reminder¶
The pipeline order is not the same as older docs in this repo. The current order is:
- polynomial features
- non-variance pre-model steps
- variance-threshold steps
- scaler when
requires_scaleris true - dimensionality-reduction pre-model steps such as SVD
- PCA
- sampler when enabled and
use_smoteis true - model
Practical examples¶
- Descriptor dataset +
logisticregression: scaler usually appears, sampler appears if enabled, SVD only appears if the dataset config turns it on. - Fingerprint dataset +
randomforestclassifier: variance threshold and SVD appear by default, scaler is skipped, sampler usually appears. - Any dataset +
extratreesclassifier: sampler is skipped by default becauseuse_smoteis false in the quick hyperparameter grid.
Common config combinations¶
Use these levers when you need to change preprocessing:
- set
training.scaler.enabled: falseto disable all scalers even for models that request one - set
variance_threshold: falseon a fingerprint dataset to disable the default filter - set
svd: falseon a fingerprint dataset to disable the default dimensionality reduction - use
training_overrideson a dataset entry when only one dataset should behave differently - set
training.pca.enabled: truewhen you want an additional PCA stage after any dataset-defined SVD
Sanity checks¶
Before a long run:
- confirm metadata columns like
Known,ID,SMILES, andkekule_smilesare dropped - confirm descriptor CSV paths use current
*_descriptors.csvnaming - confirm
training.grid_search.scoringmatches the metric you want to optimize - confirm the selected models actually exist in the referenced hyperparameter file