Analysis results
Ranked feature table
| Select | Rank | Feature | Type | MI score | Normalized | Entropy | Unique | Missing | Interpretation |
|---|
Selected feature tools
Copy-ready summary
No results yet.
Formula used
I(X;Y) = Σx,y p(x,y) log[p(x,y) ÷ (p(x)p(y))]
Mutual information measures shared information between a feature and target. It can detect linear and nonlinear statistical dependence. A high score does not establish causation.
Normalized scores divide mutual information by entropy-based quantities. Adjusted scores estimate chance agreement through target permutations. Continuous values may require discretization or neighbour estimation.
How to use this calculator
- Paste data, upload a file, or load an example.
- Parse the dataset and inspect the preview.
- Select the target and feature columns.
- Choose classification, regression, preprocessing, and estimator settings.
- Set ranking, filtering, chart, and comparison controls.
- Calculate, review warnings, and export the results.
Example output table
| Feature | Type | MI score | Normalized | Rank | Interpretation |
|---|---|---|---|---|---|
| contract_type | Categorical | 0.3912 | 0.5638 | 1 | Strong dependency |
| support_calls | Numeric | 0.2847 | 0.4415 | 2 | Moderate dependency |
| age | Numeric | 0.0631 | 0.1024 | 4 | Weak dependency |
Interpretation guidance
Scores depend on estimator, sample size, variable entropy, and preprocessing. Compare features calculated under identical settings. Treat practical labels as guidance rather than universal thresholds.
Check high-cardinality variables, identifiers, and leaked target information carefully. Recalculate scores inside cross-validation folds for predictive feature selection. Confirm important findings with validation performance and domain knowledge.
Frequently asked questions
What does mutual information measure?
It measures shared statistical information between two variables. Zero suggests no detected dependency under the estimator. Larger values indicate stronger measured dependence.
Can it detect nonlinear relationships?
Yes, mutual information can detect many nonlinear relationships. Correlation may miss curved or non-monotonic patterns. Estimation quality still depends on data and settings.
Is mutual information the same as correlation?
No, correlation mainly measures linear or monotonic association. Mutual information measures broader statistical dependence. Their scales and interpretations are different.
What does a zero score mean?
It means the estimator detected no shared information. Small samples can still produce unstable zero values. Try validation and alternative reasonable settings.
Should classification and regression use different methods?
Classification targets are naturally discrete. Regression targets are continuous and need estimation. This calculator supports binning and neighbour approximations.
How does binning affect results?
Too few bins can hide useful structure. Too many bins can exaggerate sampling noise. Compare several reasonable binning methods.
What is normalized mutual information?
It scales mutual information using entropy quantities. Normalization improves relative comparison in some datasets. The chosen formula should remain consistent.
What is adjusted mutual information?
It subtracts expected chance dependence estimated through permutations. More permutations improve stability but increase processing time. Results remain estimator dependent.
Can mutual information prove causation?
No, it measures statistical dependence rather than causality. Confounding and leakage can create large scores. Use experiments or causal methods separately.
How should missing data be handled?
Use a strategy appropriate for the dataset. Dropping rows is simple but may reduce power. Imputation can introduce its own assumptions.
Can I use scores for feature selection?
Yes, but calculate them without leaking validation targets. Combine scores with predictive validation and domain review. Reliable selection requires careful testing every time.