Model Comparison
The Model Comparison panel sits at the top of every session and gives you a unified view of all models in the session and their performance metrics. It updates automatically as models finish training.
The Comparison Table
By default the panel shows a TABLE view with one row per model. Columns include:
| Column | Description |
|---|---|
| Model Name | The name assigned when the model was added. |
| Model | The algorithm type (e.g., Support Vector Machine, Decision Tree). |
| Metric columns | One column per evaluation metric. For classification: Accuracy, F1, Precision, Recall, ROCAUC, LogLoss, and more. For regression: RMSE, MAE, and others. Metrics shown vary by task. |
| Actions | ▶ run, 👁 view details, 🗑 delete. |
Models that have not been trained yet show - in the metric columns.
Switching Between Splits
Three buttons above the table let you choose which data split the metrics reflect:
| Button | Description |
|---|---|
| TRAINING | Metrics calculated on the training subset. |
| VALIDATION | Metrics calculated on the validation subset. Used during development to tune hyperparameters. |
| TEST | Metrics calculated on the held-out test subset. The most reliable indicator of real-world performance. |
Switch between splits to understand whether a model is overfitting (high training metrics but low test metrics) or generalizing well.
Charts View
Click CHARTS in the top right of the comparison panel to switch to a visual view.
Chart Types
Two chart types are available, selectable with the BARRA (Bar) and RADAR toggles:
Bar Chart
Displays a grouped bar chart where each group represents a metric and each bar within the group represents a model. This makes it easy to compare models on any individual metric at a glance.
Radar Chart
Displays a spider/radar chart where each axis represents a metric. Each model is drawn as a polygon — a model with uniformly good performance across all metrics appears as a larger, more balanced polygon. Useful for identifying trade-offs: a model may score high on Accuracy but low on Recall, visible as an irregular shape.
Metric Selection
On the left side of the charts view, a Metrics panel lists all available metrics with checkboxes. Two quick-select buttons are available:
- ALL — enables all metrics.
- NONE — clears all selections.
Select only the metrics relevant to your analysis to reduce visual clutter. The chart updates immediately as you check or uncheck metrics.
Tips
- Always compare models on the TEST split for final evaluation — training and validation metrics can be misleading if the model overfit.
- Use the Radar Chart to quickly spot models with balanced performance across all metrics vs. models that excel in one area but underperform in others.
- Add a baseline model (e.g., Dummy Classifier for classification tasks) to the session before training other models — it sets a minimum performance floor that all other models should exceed.
- The comparison table is scrollable horizontally when many metrics are shown — scroll right to see all columns.
Troubleshooting
| Symptom | Likely cause | Solution |
|---|---|---|
All metric columns show - | No models have been trained yet | Train at least one model using the TRAIN button on its card |
Some models show - while others show values | Those models have not been trained | Click TRAIN on each untrained model card, or use RUN ALL |
| Charts view is empty | No trained models in the session | Train at least one model before switching to charts view |
| Radar chart is hard to read | Too many metrics selected | Deselect less relevant metrics using the Metrics panel on the left |