Skip to main content
The Encoding Phase transforms stable features into ML-ready signals. This is where raw business variables become the mathematical inputs your models actually learn from. AIDflow’s encoding suite is target-aware and stability-verified — every transformation is validated against the binary target and screened for temporal drift before it reaches an agent.

​
The Objective: Stable → Encoded → ML-Ready

The Encoding Phase follows a strict linear progression. WOE-ML steps depend on the previous ones.

​
1. Numerical Discretization (num_encoder)

The Binning Engine. Continuous features are transformed into discrete bins using the selected fitting strategy. This creates the foundation for WOE encoding and enables interpretable bin-level analysis.
  • Bin boundaries are learned from the training distribution.
  • Each bin is validated for minimum population share and target separation.
  • The resulting encoding is locked — the same boundaries apply at inference time.

​
2. Categorical Encoding (cat_encoder)

Category Normalization. Object and categorical features are mapped into a consistent encoded representation.
  • Low-frequency categories are handled (merged).
  • The category → code mapping is stored as an immutable artifact.
  • Ensures the model sees a stable representation regardless of original category order.

​
3. Interaction Discovery (ixn_encoder)

Pattern Amplifier. Generates new interaction variables by combining stable numerical and categorical features. Captures relationships that individual features cannot express alone.
  • Uses a tree-based search to identify high-signal feature pairs.
  • Interactions are screened for target relevance before inclusion.
  • Optional — skip if your features already capture the relevant relationships.

​
4. WOE Encoding (woe_encoder)

The Core Transformation. Converts all encoded features (numerical bins, categorical codes, interactions) into Weight of Evidence values relative to the binary target.
  • Each bin’s WOE = ln(Bad Rate / Good Rate) for that segment.
  • The result is a feature that is approximately linear in log-odds space — ideal for GLM agents.
  • WOE values are locked at training time. New observations map to the same WOE values at inference.

​
5. WOE Stability Filter (woe_stabilizer)

The Second Guardrail. Screens WOE-encoded features for temporal stability. A feature whose WOE distribution shifts significantly across time folds is flagged as unstable and pruned.
  • Complements the Data Phase stability check (which operates on raw features distribution).
  • This operates on the encoded representation — verifies time stability of target variable across feature bins.
  • Result: The final feature set is stable both in its distribution AND in its target behavior.

​
6. ML Sample Synchronization (data_encoder)

The Handoff Point. Applies all previously trained encodings to produce the final train.ml and test.ml datasets. This is the exact input that agents will train on.
  • Ensures the same encoding pipeline is applied consistently to both train and test.
  • Prevents the “train on one version, test on another” class of errors.
  • The train.ml / test.ml datasets are registered into the context — agents consume them directly.

​
Practical Implementation

Execute the full Encoding Phase:
Or run the entire phase at once:
Once complete, the ML-ready datasets are available:

​
Deep Dive: Technical Reference

For detailed information on binning strategies, WOE computation, and stability thresholds: View NumEncoder API | View WoeEncoder API | View WoeStabilizer API