Feature Importance

This method helps you understand which features matter most for predicting a target variable. The engine trains several models, compares their performance, then explains the best model using SHAP to produce feature level importance.

What the engine does

  1. Firstly, the engine trains a set of tree-based models, namely:

    • Decision Tree

    • Random Forest

    • Extra Trees

    • LightGBM

  2. Then, we evaluate each model using 10 fold cross validation and rank them by mean accuracy

  3. The system keeps the best performing model and performs SHAP process generating feature importance results

How to run (UI)

  1. Open the link for the Feature Importance along with dataset-id parameter (e.g., https://data-valuation-ui.aidapt.s5projects.eu/feature-importance?dataset-id=XXX)

  2. Select the target metric

  3. Start the analysis and expect a short wait due to cross validation

  4. Review the outputs

    • Model leaderboard with the ranked models and their scores

    • SHAP results showing feature importance

      Feature Importance UI results
      Figure 1. Feature Importance results in the UI
  5. Optionally download the results as CSV for further analysis

Interpretation

The model leaderboard helps you compare model quality under the same evaluation setup. A higher mean accuracy indicates better performance on average across folds.

SHAP values explain how features influence the model prediction.

  • Larger absolute SHAP values indicate more influential features

  • Positive SHAP values generally push predictions toward the positive class while negative values push them away, the exact meaning depends on how the target is encoded

  • If two features are strongly correlated, importance may be shared between them so each one can appear less dominant on its own

Notes and good practice

  • Ensure the dataset includes a clearly defined target column

  • SHAP computation can be expensive for large datasets, reducing feature count or sample size can speed up analysis

  • Feature importance is an explanation of the trained model, not a proof of causality