Consensus voting derives provisional inlier/outlier labels from the agreement of multiple outlier detectors. It avoids manually inspecting every record, but assumes most detectors perform reasonably well. section 14.4 section 14.4.2
A practical consensus workflow is:
- Choose a scaling method and a score-combining method for the detectors. section 14.4.2
- Run many detectors, preferably including diverse variants, and combine their scores into a final score. section 14.4.2
- Use the resulting consensus labels as pseudo-ground truth when selecting a smaller ensemble. section 14.4 section 14.4.2
- Balance similarity to the consensus against training time, prediction time, and robustness; favor diversity when robustness is difficult to test. section 14.4.2
Consensus is a practical shortcut rather than guaranteed ground truth: it may be weaker than a manually inspected superset of flagged records. Score-combination choices also affect whether the system emphasizes avoiding false positives or missing outliers; thresholds can shift that balance. section 14.4 section 14.6
