A false positive occurs when an outlier detector flags an issue that is not actually present. For a labeled record, it means predicting “outlier” when the record is an inlier. This differs from a false negative, which misses an issue that is present. A statistically unusual record is not necessarily a useful finding: whether a flag is false depends on the purpose of the investigation. section 1.1.5 section 14.0
Sensor readings
A temperature-monitoring system compares current readings with data representing normal operation. It may flag unusually high or low fluctuations for investigation, but a flag alone does not establish a process fault. Deciding how many readings to investigate means balancing false positives against missed issues. section 1.1.5
False positives become more likely or more consequential in several settings:
- Examining many combinations of features can surface entities or time periods that are unusual by chance rather than interesting anomalies. section 12.6.5
- With very few rows, a record may look unusual simply because similar examples are absent from the available data; many features increase the risk of spurious findings. section 10.3
Evaluate flags in the context of the available investigation effort. Precision at k asks how many of the top k records investigated are actual outliers; an alternative counts how many records must be examined to find k true outliers. For systems that need binary labels, first seek good score ranking, then choose a threshold suited to the relative importance of false positives and false negatives. Ensembles of outlier detectors offer another way to reduce both kinds of error, though useful ground-truth labels are difficult to establish consistently. section 8.5.2 section 14.0

























