How yusrialsanaani/interpretability_of_machine_learning_models works
Main Components
• Data Source: Represent the dataset (the Parkinson’s Data Set from UCI) as the starting point.
• Data Preprocessing: Include a block for importing and cleaning the data from the dataset.
• Model Training: Create separate components (or a single “Model Training” block) that capture model building with multiple algorithms—most notably XGBoost (XGB) and Random Forest (RF).
• Prediction & Inference: Indicate where the trained models generate predictions.
• Interpretability Engine: Break this down into two subcomponents:
• SHAP: Highlight the use of SHAP to generate both global and local explanations based on Shap values.
• LIME: Indicate that LIME is used for producing local interpretability.
• Visualization & Error Analysis: A final component that represents the visualization outputs and the insights generated (e.g., interpretation plots) to aid in error analysis and build trust on model decisions.
Relationships and Data Flow
• From Model Training, indicate a flow towards the Prediction module.
• Finally, the results (explanatory visualizations and metrics) go to the Visualization & Error Analysis component.
Architectural Patterns and Principles
• That this setup follows a pipeline structure common in machine learning projects.
• The separation of concerns is clear: data ingestion, model execution, and post-hoc interpretation.
• That each module (especially within the Interpretability Engine) plays a distinct role in generating insights for either global or local model behavior.
• Indicators of dependencies like the use of SHAP’s dependency on Shap values and LIME’s focus on local approximations.