Monitor data quality, investigate anomalies detected by the ML ensemble, and move from findings to recommended actions.
Historical demonstration runs are synthetic; current KPIs come from Collibra DQ.
Live Data Quality from Collibra combined with record-level and run-level ML analysis.
Overall quality score from Collibra DQ.
Classic DQ runs with retrievable scan results.
Dimension health derived from Classic DQ category impact scores.
Current distribution of completed, failed and cancelled DQ job runs.
Historical Data Quality trend for the selected date range. The main line represents the overall DQ score; available quality dimensions can be displayed as supporting series.
Synthetic test-run history used for the ML demonstration. The chart compares DQ signals with run-level anomaly risk.
Distribution of records across LOW, MEDIUM, HIGH and CRITICAL anomaly-risk levels.
Number of observations classified as anomalous by each member of the ensemble.
Click a record to open its detailed ensemble analysis in the ML Details tab.
| Transaction | Selected value | Risk | Votes | Score | |
|---|---|---|---|---|---|
| Loading… | |||||
Record-level ensemble combining Isolation Forest, One-Class SVM, Mahalanobis Distance and LDOF. The final risk score is an anomaly-prioritisation score, not a probability.
—
Number of ensemble models voting for anomaly.
Fraction of models supporting the winning class.
Weighted average of normalized model anomaly scores.
Model-specific anomaly scores normalized to a common 0–1 scale. Raw scores are available in the tooltip but are not directly comparable across models.
Absolute z-score for each model feature. A higher value means the feature is farther from the dataset mean in standard-deviation units. This is a descriptive deviation metric, not model feature importance.
Signed z-scores for the selected transaction and the nearest comparison records across the same model features. Each row is a transaction and each column is a feature. Values close to zero are typical; large absolute values indicate stronger deviation.
Each chart shows the model-specific raw anomaly-score distribution for the full analysed population. The yellow line marks the currently selected transaction. Its percentile shows where the record sits within that model's population. Higher raw score means more anomalous in this project, but raw values must not be compared between different models because every detector uses a different scoring scale.
Risk score = 65% × anomaly strength + 35% × anomaly vote ratio.
Anomaly strength is the weighted average of normalized anomaly scores returned by the models. The anomaly vote ratio is the weighted fraction of models classifying the record as anomalous.
Risk thresholds: LOW < 0.35, MEDIUM ≥ 0.35, HIGH ≥ 0.60, CRITICAL ≥ 0.80.
The final risk score is an anomaly-prioritisation score, not a probability.
Isolation Forest isolates unusual observations through random feature-space partitions.
One-Class SVM learns a boundary around typical observations and flags records outside that region.
Mahalanobis Distance measures covariance-aware multivariate distance from the data centre.
LDOF compares local record distance with the density of neighbouring observations.
Findings are grouped by transaction. Generate a remediation recommendation with the dedicated Ollama recommender, then use Data Steward Copilot separately for governance questions.
The fine-tuned Ollama recommender receives an already detected DQ issue and returns business-oriented remediation plus next steps. It is separate from Data Steward Copilot and does not calculate ML or DQ scores.
