Get started
Work

Using Information on Influencing Factors to Partition Training Samples for Classification Under Changing Operating Conditions

Year: 2026
Type: article
Abstract: Training samples collected from operating systems are rarely homogeneous: seasons, working calendars, and scheduled events shift variable ranges, change class frequencies, and increase class overlap. The periods during which such factors act are often known in advance but are ignored when samples are formed. This paper formalises a factor-induced partition of the training sample with rules for selecting factors, shows that knowledge of the regime cannot increase the Bayes error, and derives a finite-sample condition under which partitioning pays, together with its relation to supplying the regime as an input feature. On two electricity streams with recovered calendars, seven classifiers, and protocols in which all methods predict identical records, partitioning beats the pooled model in 40 of 42 cells by up to 6.5 percentage points, but beats a pooled model given the regime as a feature by only 1.1 points for parametric and 0.1 for flexible classifiers. Where the calendar explains the change, partitioning outperforms DDM, ADWIN, and adaptive random forests in prequential evaluation; where unexplained drift dominates, reactive methods are better. A minimum subsample size prevents losses from over-refinement, and a pilot estimate of the heterogeneity gap predicts the benefit of partitioning better than the silhouette coefficient.
Source: Algorithms
DOI: 10.3390/a19110971
Language: English
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7225660566
FWCI:
Cites:
Cited by:
Related to: