Training time is the duration required to fit a model or detector to data. For outlier detection, it should be considered separately from inference time: fitting may happen infrequently outside production, while repeated retraining can matter when new data continually arrives. section 3.3.5
Use training time as a practical scalability and deployment consideration:
- Benchmark both fitting and prediction as row counts and feature counts grow; a detector that works on small data may become impractical at production scale. section 3.3.5
- Compare algorithms rather than assuming faster fitting means faster inference: KNN fits quickly but performs many distance calculations during prediction, whereas Isolation Forest and CBLOF do more work during fitting. section 8.7
- Prefer simple univariate aggregation or scalable multivariate detectors when workload size makes slower methods impractical. section 3.3.5
Timing detectors across growing workloads
A timing experiment duplicates an existing dataset to create six workload sizes, then measures fit and prediction durations for 12 detectors. On the abalone data, OCSVM and KDE become especially slow as rows increase, while many other detectors remain short; KNN and ABOD reach about 12 seconds for the largest prediction workload shown. section 8.7
