Outlier Detector
Detect multivariate outliers using Isolation Forest and DBSCAN — two complementary methods that catch different kinds of anomalies. Unlike column-by-column checks, this analysis finds rows that are unusual in combination, even when each individual feature looks normal.
How it works: features are standardised, then each algorithm independently scores every row. In ensemble mode, rows flagged by either method are shown; rows flagged by both are marked as consensus outliers.
Drop a CSV file here or click to upload
No login required. Raw upload is processed for analysis and not retained by default.
Max 50MB · at least 10 rows · at least 2 numeric columns
Ensemble shows outliers flagged by either method; consensus = both.
Excluded from features; shown in outlier row details.
Expected proportion of outliers.
Neighbourhood radius (standardised space).
Minimum cluster size.
Running outlier detection…
Fitting Isolation Forest and DBSCAN on your data
Continue your pre-training checklist
Outlier detection is inherently domain-dependent. A row that is statistically anomalous may be valid data. Review flagged rows in context before removing them. Isolation Forest works best when outliers are sparse; DBSCAN is better at detecting cluster-based anomalies. Ensemble mode reduces false negatives from either method alone.
What you get
Health score
A 0–100 score based on the outlier rate, with a colour-coded verdict: Healthy, Needs attention, or Critical.
Method comparison
Side-by-side stats for Isolation Forest and DBSCAN — outlier counts, cluster breakdown, and anomaly score averages.
Feature contributions
Ranked list of features that most strongly separate outliers from inliers, using mean-shift Z-score.
Flagged rows
The top anomalous rows with their anomaly score, DBSCAN cluster, consensus flag, and full feature values.