build-pipeline

name	build-pipeline
description	Generate a complete scikit-learn pipeline with GAFIME feature interaction mining. Use when the user wants to create an ML pipeline, integrate GAFIME with sklearn, build a classification or regression workflow, or says things like "build a pipeline", "create a model with GAFIME", "sklearn integration", "churn prediction pipeline", "set up classification", or "generate training script".

Build ML Pipeline

Generate a complete, ready-to-run Python script that integrates GAFIME into a scikit-learn pipeline.

Instructions

1. Ask the user for:

   • Task type: classification or regression
   • Data source: CSV/Parquet path, or they'll use in-memory data
   • Target column: name of the target variable
   • Model preference: LogisticRegression, RandomForest, XGBoost, CatBoost, or auto
   • Number of top interactions (k): how many GAFIME features to add (default: 10)
   • Whether to include v0.4.x discrete threshold/rectangle candidates

2. Run the pipeline generator:

   python .claude/skills/build-pipeline/scripts/generate_pipeline.py \
       --task classification \
       --data "data.parquet" \
       --target "churn" \
       --model auto \
       --k 10 \
       --output "gafime_pipeline.py"
   

3. The script generates a complete Python file with:

   • Data loading (Polars or Pandas)
   • Train/test split
   • GafimeSelector in an sklearn Pipeline
   • Cross-validation evaluation
   • Results reporting
   • Feature importance analysis

   Note: GafimeSelector is the sklearn transformer for pair interaction augmentation. If the user wants v0.4.x discrete functions, generate an explicit GafimeEngine training-fold step with enable_discrete_functions=True, then materialize selected discrete candidates with evaluate_discrete_candidate. Do not fit thresholds on the test fold.

4. Review the generated script with the user and customize if needed:

   • Adjust k (number of interactions)
   • Change operator (multiply, add, subtract, divide)
   • Add enable_discrete_functions=True in an engine feature-generation stage if threshold/rectangle signals are needed
   • Modify evaluation metric
   • Add preprocessing steps

v0.4.x Discrete Engine Snippet

from gafime import ComputeBudget, EngineConfig, GafimeEngine

config = EngineConfig(
    backend="auto",
    metric_names=("pearson", "r2"),
    mi_bins=96,
    enable_discrete_functions=True,
    discrete_mode="soft",
    discrete_ranking="split_aware",
    budget=ComputeBudget(
        max_discrete_candidates=20_000,
        top_k_features_for_discrete=24,
    ),
)
report = GafimeEngine(config).analyze(X_train, y_train)

On GPU backends, keep discrete_mode="soft". Hard discrete mode is C++ Core only. CUDA, Metal, and ROCm/HIP use soft/vectorized discrete paths. Keep the default discrete_ranking="split_aware" for threshold, interval, and rectangle candidates unless the user explicitly asks to rank by the report metrics.

In current v0.4.x releases, mi_bins=96 is an adaptive maximum and split-aware ranking uses soft-binary mask MI, so it is a better default than Pearson-only ranking for discrete candidates.

For explicit GPU runs, verify the backend's report-metric support. Use backend="core" when the user requires a metric that a GPU payload does not report natively.

5. Explain each section of the generated pipeline so the user understands what's happening.

Example

User says: "Create a churn prediction pipeline using GAFIME and CatBoost"

Actions: Run generator with --task classification --model catboost --target churn

Result: A complete gafime_pipeline.py that loads data, discovers top 10 feature interactions with GAFIME, trains CatBoost, and reports AUC-ROC with cross-validation.