Training fits a model to data by identifying its predominant patterns and representing them. In outlier detection, this establishes a reference for what counts as normal; predicting is a distinct step that compares records against that reference. section 4.4 section 4.14
How fitting works depends on the model:
- Some models summarize the training data with statistics, clusters, or frequent item sets. section 4.4
- Distance- and density-based models such as KNN and Radius may retain the training points as their reference, making the choice of training data especially important. section 4.4
- Training can use a sample to save time, but sample size trades off speed and accuracy. Test with manipulated records to estimate a suitable size; prediction usually still needs to cover every record. section 10.2.1
Setting a sensor baseline
Fit a model to collected sensor readings believed to be reasonably normal, then compare incoming readings with that baseline. Refitting may be needed as conditions change, since older or unrepresentative data can make the reference misleading. section 4.14 section 4.4
