Analysis result
The chart and statistics update after valid data is loaded.
Interpretation
Selected observations
| ID | Feature | SHAP | Interaction |
|---|---|---|---|
| Click or box-select chart points. | |||
Formula used
A SHAP value measures one feature’s contribution to a prediction relative to a baseline prediction. Positive values push predictions upward, while negative values push them downward.
f(x) = E[f(X)] + Σ φᵢ Dependence plot point: (xᵢⱼ, φᵢⱼ) Pearson correlation: r = Σ[(x - x̄)(y - ȳ)] / √(Σ(x - x̄)² Σ(y - ȳ)²) Linear trend: y = a + bx
The calculator also reports Spearman correlation, descriptive statistics, outliers, group differences, and a fitted linear trend when numeric data is available.
How to use
- Paste CSV or JSON data, or upload a compatible file.
- Load the data and confirm the previewed rows and columns.
- Map the main feature, SHAP value, interaction, target, and group columns.
- Choose filtering, comparison, trend, and chart appearance options.
- Generate the plot and inspect statistics, outliers, and selected observations.
- Export the chart, filtered data, settings, or complete PDF report.
Example data
| ID | Age | Age_SHAP | Income | Actual | Predicted | Group |
|---|---|---|---|---|---|---|
| 1 | 22 | -0.61 | 24000 | 0 | 0.18 | Test |
| 6 | 37 | 0.05 | 59000 | 1 | 0.55 | Train |
| 12 | 59 | 0.91 | 108000 | 1 | 0.92 | Test |
Interpretation guide
The horizontal axis contains observed feature values. The vertical axis contains SHAP contributions for that feature. Color can reveal interactions with another feature.
Upward patterns suggest larger feature values generally increase predictions. Downward patterns suggest they reduce predictions. Curves, bands, or abrupt changes can indicate nonlinearity and thresholds.
Separated color bands may indicate feature interactions. Dense vertical spreads show differing effects at similar feature values. Always validate patterns with domain knowledge and model diagnostics.
Common mistakes
- Using SHAP values from a different model or dataset version.
- Treating correlation as proof that the feature causes the prediction.
- Ignoring preprocessing, encoding, missing values, and background data choices.
- Comparing multiclass SHAP values without selecting the same output class.
- Overinterpreting sparse regions with very few observations.
- Assuming a smooth trend accurately represents every local prediction.
Frequently asked questions
What does a SHAP dependence plot show?
It shows how a feature’s observed values relate to that feature’s contribution to model predictions.
What does a positive SHAP value mean?
It means the feature pushes the prediction above the selected baseline for that observation.
What does a negative SHAP value mean?
It means the feature pushes the prediction below the selected baseline for that observation.
Why are points colored?
Color represents another feature and helps reveal possible interaction effects between two features.
Can categorical features be plotted?
Yes. The calculator supports jittered scatter, box, and violin views for categorical values.
How are outliers detected?
Numeric SHAP outliers use the interquartile range rule unless outlier detection is disabled.
What is automatic interaction selection?
It selects the numeric column with the strongest absolute correlation to the plotted SHAP values.
Does correlation prove causation?
No. Dependence patterns describe model behavior and associations, not real-world causal relationships.
Can groups or model versions be compared?
Yes. Map a group column, then overlay groups or display separate plot panels.
Are uploaded files sent to a server?
No. This implementation processes uploaded data inside the browser after the page loads.
Can the chart be exported?
Yes. Export PNG, SVG, PDF, filtered CSV data, selected observations, and settings JSON.