Skip to content

Advanced Features# Advanced features# Advanced features# Advanced Features# Advanced Features

This page summarizes the notebook's advanced techniques: explainability

with SHAP, enhanced visualizations, and production-ready pipeline tips.

This page summarizes the notebook's advanced techniques: explainability

SHAP Explainability

with SHAP, enhanced visualizations, and production-ready pipeline tips.

SHAP (SHapley Additive exPlanations) attributes model predictions to

features. For tree-based models use TreeExplainer (fast). For otherThis page summarizes the notebook's advanced techniques: explainability

models KernelExplainer can be used but is much slower.

SHAP explainability (concise)

Key visualizations:

with SHAP, enhanced visualizations, and production-ready pipeline tips.

  • Summary plot — global feature importance

  • Dependence plots — relationship between a feature and its SHAP valueSHAP (SHapley Additive exPlanations) attributes model predictions to

  • Waterfall/force plots — single-prediction breakdown

features. For tree-based models use TreeExplainer (fast). For otherAdvanced capabilities beyond the baseline ML pipeline.Advanced capabilities beyond the baseline ML pipeline.

Example:

models KernelExplainer can be used but is much slower.

explainer = shap.TreeExplainer(model)## SHAP explainability (concise)

shap_values = explainer.shap_values(X_sample)

shap.summary_plot(shap_values, X_sample)Key visualizations:

Feature Engineering Highlights

  • Summary plot — global feature importance

Notebook-derived engineered features used in the pipeline:

  • Dependence plots — relationship between a feature and its SHAP valueSHAP (SHapley Additive exPlanations) attributes model predictions to

  • cabin_multiple — number of cabins

  • name_title — extracted title from passenger name- Waterfall/force plots — single-prediction breakdown

  • norm_fare — log- or scaled fare

features. For tree-based models use TreeExplainer (fast). For other## SHAP Explainability## SHAP Explainability

Keep transformations reproducible: implement them as transformers in

the production Pipeline (see titanic_ml/pipeline/train_pipeline.py).Example:

Visualizationsmodels KernelExplainer can be used but is much slower.

  • plot_cv_score_distribution() — boxplots of CV scores to inspect```python

stability

  • plot_precision_recall_curve() — preferred for imbalanced datasetsexplainer = shap.TreeExplainer(model)

  • plot_model_comparison() — horizontal bar chart of model scores

shap_values = explainer.shap_values(X_sample)

Production Notes

shap.summary_plot(shap_values, X_sample)Key visualizations:

  • Use the pipeline for production training and inference to avoid data

leakage (preprocessing fit only on training data)```

  • Serialize pipelines with pickle (or ONNX for cross-platform

serving) and store in models/SHAP (SHapley Additive exPlanations) explains model predictions by computing feature contributions.SHAP (SHapley Additive exPlanations) explains model predictions by computing feature contributions.

Troubleshooting## Feature engineering highlights

  • SHAP memory issues: reduce sample size- Summary plot — global feature importance

  • Feature name mismatches: ensure the same FeatureEngineer is used

during training and inference- cabin_multiple: number of cabins

  • Pickling errors: avoid local (nested) class definitions

  • name_title: title extracted from passenger name- Dependence plots — relationship between a feature and its SHAP value

  • norm_fare: normalized or log-transformed fare

Use the notebook for exploration and the pipeline module for training

and deployment. Keep configuration and constants in- Waterfall/force plots — single-prediction breakdown

notebooks/utils/config.py to avoid hardcoded values in cells.

Implement these as Transformers in the production pipeline to avoid

leakage.### Available Visualizations### Implementation

For more details see the notebook Section 9 (SHAP) and

titanic_ml/pipeline/train_pipeline.py.

VisualizationsExample:

  • plot_cv_score_distribution() — boxplots of CV scores to inspect# Advanced Features

model stability

  • plot_precision_recall_curve() — useful for imbalanced datasets```python

  • plot_model_comparison() — horizontal bar chart of model scores

explainer = shap.TreeExplainer(model)This page summarizes the notebook's advanced techniques: explainability

Production notes

shap_values = explainer.shap_values(X_sample)with SHAP, enhanced visualizations, and production-ready pipeline tips.

  • Use titanic_ml/pipeline/train_pipeline.py for production training to ensure

consistent preprocessing and avoid leakageshap.summary_plot(shap_values, X_sample)

  • Serialize pipelines (pickle/ONNX) and store under models/

SHAP Explainability (Concise)

Troubleshooting

  • SHAP memory issues: reduce sample size

  • Feature name mismatches: ensure same feature engineering at train## Feature engineering highlightsSHAP (SHapley Additive exPlanations) attributes model predictions to

and inference

  • Pickling errors: define classes at module levelfeatures. Use TreeExplainer for tree models (fast) and KernelExplainer

---- cabin_multiple: number of cabinsonly when necessary (slow).

For more details see the notebook Section 9 (SHAP) and- name_title: title extracted from passenger name

titanic_ml/pipeline/train_pipeline.py.

  • norm_fare: normalized or log-transformed fareKey visualizations:

Implement these as Transformers in the production pipeline to avoid- Summary plot — global feature importance

leakage.- Dependence plots — feature value vs. SHAP value

  • Waterfall / force plots — single-prediction explanations

Visualizations

Best practices:

  • plot_cv_score_distribution() — boxplots of CV scores to inspect

model stability- For large datasets sample 100–500 rows for SHAP summaries

  • plot_precision_recall_curve() — useful for imbalanced datasets- Focus on the positive class for binary classification (Survived=1)

  • plot_model_comparison() — horizontal bar chart of model scores- Use TreeExplainer for tree-based models

Production notesExample:

  • Use titanic_ml/pipeline/train_pipeline.py for production training to ensure```python

consistent preprocessing and avoid leakageexplainer = shap.TreeExplainer(model)

  • Serialize pipelines (pickle/ONNX) and store under models/shap_values = explainer.shap_values(X_sample)

shap.summary_plot(shap_values, X_sample)

Troubleshooting```

  • SHAP memory issues: reduce sample size## Feature Engineering Highlights

  • Feature name mismatches: ensure same feature engineering at train

and inferenceNotebook-derived engineered features used in the pipeline:

  • Pickling errors: define classes at module level

  • cabin_multiple — number of cabins

---- name_title — extracted title from passenger name

  • norm_fare — log- or scaled fare

For more details see the notebook Section 9 (SHAP) and

titanic_ml/pipeline/train_pipeline.py.Keep transformations reproducible: implement them as transformers in

the production Pipeline (see titanic_ml/pipeline/train_pipeline.py).

Production Notes

  • Use the pipeline for production training and inference to avoid data leakage (preprocessing fit only on training data).
  • Serialize pipelines with pickle (or ONNX for cross-platform serving) and store in models/.

Visualizations

  • plot_cv_score_distribution() — boxplots of CV scores to inspect stability
  • plot_precision_recall_curve() — preferred for imbalanced datasets
  • plot_model_comparison() — horizontal bar chart of model scores

Troubleshooting

  • SHAP memory issues: reduce sample size
  • Feature name mismatches: ensure the same FeatureEngineer is used during training and inference
  • Pickling errors: avoid local (nested) class definitions

Use the notebook for exploration and the pipeline module for training and deployment. Keep configuration and constants in notebooks/utils/config.py to avoid hardcoded values in cells.

For more details see the notebook Section 9 (SHAP) and titanic_ml/pipeline/train_pipeline.py.