ifcfill¶
ifcfill prepares tabular data for synthetic-data generation. It transforms real data into Integer, Float, and Categorical (IFC) variables, fills missing values using fast NumPy operations, and keeps the metadata needed to map generated synthetic data back to the original table structure.
ifcfill is intentionally unsupervised. It does not require a target variable;
instead, it learns transformation and inverse-transformation rules from the real
data so the same inverse rules can be applied to synthetic outputs from any
tabular generator.
Key Features¶
| Feature | Description |
|---|---|
| Flexible input | pandas.DataFrame or CSV file path |
| Automatic type inference | Detects integer, float, categorical, and datetime columns |
| Per-column type overrides | Force a specific type for any column |
| Configurable imputation | Independent strategy per type (mean, median, mode, zero, constant) |
| Categorical missingness | Uses a synthesis-time sentinel, then applies IFCFill's authoritative reconstructed mask |
| Missingness reconstruction | Always restores fitted marginal P(M_j) or empirical joint P(M_1, ..., M_p) missingness |
| Categorical label encoding | Optionally encode filled categories as integer codes with a separate inverse-compatible encoder layer |
| Datetime conversion | Date/time → integer relative to a configurable anchor |
| Parallel processing | Use n_jobs to process columns concurrently during fit and transform |
| Constant column removal | Drops true constants while preserving learnable categorical missing categories |
| Missing value tracking | Records count and fraction per column via missing_report_ |
| Inverse transform | Restores semantic dtypes, constants, column order, and missingness across all original columns |
| Portable fitted state | Save learned transformations to JSON and load them later on another machine |
| Generator support | Fit on real data, transform for a generator, inverse-transform generated synthetic data |
Installation¶
Quick example¶
from ifcfill import IFCTransformer
tf = IFCTransformer()
real_ifc = tf.fit_transform("real_data.csv")
print(tf.missing_report_)
tf.save("ifcfill-state.json")
# synthetic_ifc is produced by your tabular synthetic-data generator
loaded_tf = IFCTransformer.load("ifcfill-state.json")
synthetic_restored = loaded_tf.inverse_transform(synthetic_ifc)
Get started →{ .md-button .md-button--primary } Examples →{ .md-button } API Reference →{ .md-button }