| Run | Trees | Sample size | Contamination | Threshold | Anomalies | Mean score | Seed |
|---|---|---|---|---|---|---|---|
| No completed runs. | |||||||
Isolation Forest isolates observations through repeated random feature splits. Anomalies usually require fewer splits than typical observations. Their average path lengths therefore produce larger anomaly scores.
The contamination setting determines the automatic classification threshold. A manual threshold can replace that quantile when needed. Positive decision scores indicate points below the anomaly threshold.
- Paste data or upload a CSV or TXT file.
- Parse the data, then choose an ID column and numeric features.
- Set tree count, sample size, contamination, scaling, and threshold options.
- Select a plot type and the X and Y features.
- Run the calculator, inspect anomalies, and export the required results.
| ID | Feature 1 | Feature 2 | Feature 3 | Expected pattern |
|---|---|---|---|---|
| P01 | 2.1 | 1.9 | 2.3 | Typical |
| P08 | 2.5 | 2.3 | 2.2 | Typical |
| P13 | 7.8 | 8.4 | 7.5 | Likely anomaly |
| P14 | -3.2 | -2.8 | -3.5 | Likely anomaly |
| P15 | 5.9 | -4.7 | 6.2 | Likely anomaly |
- Higher anomaly scores indicate observations isolated more quickly.
- Negative decision scores indicate values beyond the selected anomaly threshold.
- Borderline observations lie within the chosen tolerance around the threshold.
- Extreme anomalies appear near the top of the anomaly-score ranking.
- Detected anomalies may be valid rare events, measurement errors, or distribution shifts.
- Small datasets, constant columns, duplicates, and heavy missingness can weaken conclusions.
What does Isolation Forest detect?
It detects observations that are isolated with unusually short random-tree paths. These points differ from common data patterns. Detection does not prove that a record is wrong.
What contamination value should I use?
Use an estimate of the expected anomaly proportion. Lower values produce fewer flagged observations. Compare several settings before making operational decisions.
Why are scores different after changing the seed?
Isolation Forest uses randomized samples, features, and splits. Different seeds generate different tree ensembles. Stable anomalies should remain highly ranked across repeated runs.
Does feature scaling matter?
Random split trees are less scale-sensitive than distance methods. Scaling still improves comparability and interpretation. It also helps when inspecting standardized feature deviations.
What does a negative decision score mean?
This calculator defines decision score as threshold minus anomaly score. Negative values exceed the selected anomaly threshold. Positive values remain on the normal side.
Can I use more than two features?
Yes, the model can use many selected numeric features. Two features are displayed in contour and scatter views. Other features remain fixed at representative values.
Why is the contour only an approximation?
The contour evaluates a grid across two selected dimensions. Remaining features use their transformed means. The full model still scores every observation with all features.
How should duplicate rows be handled?
Duplicates can represent valid repeated events or accidental copies. Review their context before removal. Many duplicates can distort anomaly proportions and ranking.
Can results replace expert review?
No, unsupervised anomaly detection provides screening evidence. Domain context remains essential for interpretation. Confirm important findings before taking consequential action.