6. Ideate Features and Models

Use Ideation to accelerate feature engineering through an automated process that generates feature and model candidates based on your data and use case.

This tutorial focuses on the Fully Automated Mode for a forecasting use case and walks through the following tasks:

• Run Ideation.
• Access the Ideation Report.
• Review the Best Performing Model.
• Examine the Feature Selection.
• Explore a Single Feature in detail.

Note

For guidance on manually creating features, refer to the SDK tutorials.

---

Step 1: Run a New Ideation Workflow

1. Navigate to Ideation from the 'Ideate' section of the menu.

2. Select the use case: "Store Daily Sales Amount Forecast for 28 Days".

   

   ---

3. Click to start the Ideation process.

4. Edit the Ideation name and description by clicking .

   

   ---

5. Configure the ideation by clicking . Set the training and validation observation tables in Modeling Setup:

   • Training Observation Table: Training
   • Validation Observation Table: Validation_eval

   Once done, click .

   

   ---

6. Begin the automated Ideation workflow by clicking .

Once the process is initiated, you'll see confirmation that the run has started:

After the process completes, a list of models will be displayed for your review.

How Ideation Works for Forecasting

Ideation dynamically tailors feature generation to your forecasting use case. The process includes:

1. Analyzing tables and relationships to identify relevant data.
2. Inferring missing semantic tags based on column metadata.
3. Recommending column transformations, such as time deltas, ratios, and differences.
4. Identifying key filters to isolate critical events.
5. Highlighting key columns for further feature engineering.
6. Proposing appropriate aggregation time windows based on the time series frequency.
7. Analyzing event frequency patterns to detect timing signals.
8. Leveraging calendar table attributes (events, SNAP flags) as forecast-point features.
9. Detecting existing features in the Catalog to promote feature reuse.
10. Conducting EDA on each feature and assigning an individual Predictive Score.
11. Selecting a feature set based on SHAP value analysis.
12. Running Machine Learning Models on the feature set.

Every step is transparently documented to ensure full traceability.

---

Step 2: Review the Ideation Report

1. Access the Detailed Report, describing each step of the ideation process, by clicking next to the Ideation name.

   

   ---

2. Visualize the full report with an indexed view in a new tab, by clicking .

   

---

Step 3: Review Best Performing Model

1. Go back to Model Training step to access the ideation leaderboard and click on for an overview.

   

   ---

2. Click on one Model to access more details.

   

   ---

3. Go to Evaluate tab for interactive plots and insights into model performance.

4. For a regression forecasting model, review the following evaluation plots:

   • Distribution of Predicted Scores: View how the model's predictions are distributed to check for skew or unexpected patterns.

     

     ---

   • Predicted vs Actual (scatter plot): Check how well individual predictions align with actual values. Points close to the diagonal indicate accurate predictions.

     

     ---

   • Avg Predicted vs Actual per Score-Ranked Bin: Assess calibration by comparing the average predicted value against the average actual value within bins ranked by predicted score.

     

   ---

5. Review Feature Importance to understand which features drive the predictions.

   

   ---

6. Go to Model Graph to review pre-processing steps and estimator.

   

---

Step 4: Review Feature Selection

Go to Feature Selection step to access suggested feature selection and click on a feature selection to get an overview of how this selection was generated, the signals captured and table columns used.

---

Step 5: Review a Single Feature

1. Select the Features tab of the Feature Selection step to review the suggested features.

   

   ---

2. Click on a feature to open its details. You can use the filter or the search , to find a specific feature.

   

   ---

3. Check Semantic Relevance in the About tab of the feature.

   

   ---

4. Explore Feature Lineage by going to the Lineage tab and click to trace the feature's origin and transformations.

   

   

   ---

5. Analyze Feature Distribution and its relationship with the Target in the EDA tab.

   Residual EDA

   When Ideation uses a naive prediction (e.g. STORE_Avg_of_Sales_records_sales_amounts_182cD — the store-level average of sales_amount over a 182-day window), the EDA shows the feature's relationship with the residual (actual / naive prediction) rather than the raw target. With a multiplicative structure, this ratio highlights whether the feature captures signal beyond what the naive baseline already explains, which is more informative for feature selection.

   

   ---

6. Go to the 'SDK Code' tab of the feature to see how to reproduce it programmatically.

   

---

Next Steps

• Continue to Predict and Evaluate to score models on holdout sets, compare them on the Leaderboard, and visualize forecasts.
• To refine ideation — column transforms, feature selection, or new models — see the Credit Default tutorials: Refine Ideation and Create New Feature Lists and Models.