Skip to content

data.md

Latest commit

 

History

History
91 lines (65 loc) · 4.12 KB

data.md

File metadata and controls

91 lines (65 loc) · 4.12 KB
Raw

Datasets

Input Data Requirements

STREAMLINE expects tabular supervised learning datasets.

  1. Use .csv, .tsv, or .txt files with a header row.
  2. Rows are instances and columns are variables.
  3. Include one outcome column specified by outcome_label.
  4. Include an optional instance identifier column specified by instance_label.
  5. Include optional feature-type files listing categorical and quantitative feature names.
  6. Encode missing values as blank cells, NA, NaN, or ?; avoid numeric placeholders such as -99 unless they are real values.
  7. Keep replication datasets aligned with the original dataset's feature names and outcome/instance labels.

Recommended preparation before P1:

  • Remove free-text, image, raw time-series, or other unstructured columns unless you have already converted them to tabular features.
  • Decide whether high-cardinality identifier-like fields should be instance_label, ignored, or removed before modeling.
  • Check that the outcome column has the interpretation you expect. Binary labels should be documented so downstream metrics are not ambiguous.
  • Keep a copy of the raw source data outside the STREAMLINE output directory.

The current code supports:

Task Example outcome_type Notes
Binary classification Binary Standard classification metrics, ROC, PRC, and calibration outputs.
Multiclass classification Multiclass Macro/micro summaries where applicable.
Regression Continuous Regression metrics and residual-style outputs.

Feature Types

You can provide feature type files through:

categorical_features = path/to/categorical_features.csv
quantitative_features = path/to/quantitative_features.csv

Each file should contain feature names, one per row or in a single column. If feature type files are omitted, P1 can infer categorical features using categorical_cutoff.

Supplying feature-type files is recommended for real analyses because numeric codes are common in biomedical and administrative data. A coded feature such as 1, 2, 3 may be categorical even though it looks numeric to Python.

One-Hot Encoding And Native Categorical Models

By default, P1 one-hot encodes non-binary categorical features. Set one_hot_encoding = False when you want later models to handle raw categorical columns directly.

When one_hot_encoding is false, P6 only runs models listed in native_categorical_models by default. Unsupported explicitly requested models raise an error instead of silently changing the feature representation.

Included UCI Demo Datasets

The repository includes deterministic 80:20 train/replication splits for three UCI-derived demos:

Task Training folder Replication folder Outcome
Binary classification data/UCIBinaryClassification data/UCIRepBinaryClassification Class
Multiclass classification data/UCIMulticlassClassification data/UCIRepMulticlassClassification Class
Regression data/UCIRegression data/UCIRepRegression MPG

The demo sources are:

Feature type files for the demos live in data/UCIFeatureTypes/.

The normal demo folders contain training/development splits. The matching UCIRep* folders contain held-out replication splits created for STREAMLINE demonstration and testing.

Replication Data

Replication datasets are external validation inputs for P10. They are not assumed to be official test splits from UCI. The included demo replication folders are deterministic held-out 20% splits made to exercise P10 and replication reporting.

Ignored Columns

Identifier-like fields should be supplied as instance_label or ignored before modeling. For Auto MPG, the car name is treated as an identifier-like field in the demo preparation rather than a normal predictive feature.