Calculation results
Results appear here after calculation.
Feature contribution chart
Baseline-to-prediction waterfall
| Rank | Feature | Value | SHAP value | Absolute value | Contribution | Direction | Running output |
|---|---|---|---|---|---|---|---|
| No results yet. | |||||||
Formula used
SHAP explanations use an additive model. The prediction equals the expected output plus all feature contributions. This identity supports direct additivity validation.
f(x) = φ₀ + Σ φᵢ
φᵢ = Σ [ |S|! (M - |S| - 1)! / M! ] × [ f(S ∪ {i}) - f(S) ]
Here, φ₀ is the expected model output. Each φᵢ measures one feature’s marginal contribution. Coalition weights average effects across possible feature orders.
How to use
- Select regression or classification.
- Choose the model output scale.
- Enter the expected model output.
- Add each feature and SHAP value.
- Optionally add pairwise interactions.
- Set filtering and precision options.
- Calculate and review additivity.
- Export the table when needed.
Example data
| Feature | Observed value | SHAP value | Interpretation |
|---|---|---|---|
| Income | 62000 | 2.1 | Raises the prediction. |
| Age | 34 | 0.7 | Raises the prediction. |
| Debt ratio | 0.31 | -0.9 | Lowers the prediction. |
| Account tenure | 5.2 | 1.3 | Raises the prediction. |
Interpretation guide
A positive SHAP value moves the output above its baseline. A negative value moves the output below its baseline. Larger absolute values indicate stronger influence.
Probability explanations should remain within zero and one. Log-odds values require conversion before probability interpretation. Correlated features can divide or redistribute apparent influence.
Local explanations describe one observation. Global importance aggregates many observations using absolute SHAP values. Always compare explanations with model behaviour.
Frequently asked questions
What is a SHAP value?
It estimates how one feature changes a prediction from the expected model output.
Why must SHAP values add up?
The additive explanation reconstructs the prediction from the baseline and feature contributions.
What does a negative SHAP value mean?
It means the feature pushes the explained output lower than it otherwise would be.
Can SHAP explain classification probabilities?
Yes, when the explainer returns probability-scale values or values converted from log-odds.
What is the expected value?
It is the average model output over the selected background or reference dataset.
What is global SHAP importance?
It is commonly the mean absolute SHAP value across many observations.
Why do correlated features cause issues?
Their shared information can make attribution depend strongly on background assumptions.
When should exact coalitions be used?
Use them for small feature sets because subset counts grow exponentially.
Does this replace a model explainer library?
No. It validates, explores, and calculates supplied outputs without loading trained models.