Lightweight Adaptation of Classifiers to Users and Contexts: Trends of the Emerging Domain

Table 1

Characteristics of the reviewed adaptation approaches. Notations: X: yes; A: depending on a chosen algorithm; for example, training data can be used instead of assumptions about new context and unlabelled data are used only as negative examples; MK: multiclass classification problems only; U: data are vectors of user preferences; base classifier: either ensemble member or the only context-specific model.

Method name	Applicability				Training		Additional data
	Suits dissimilar contexts	Suits large varieties of problems/algorithms	Most lightweight	Requires little assumptions about new contexts	For training of base classifier(s)	For classifier selection/combination	Uses contextual data	Uses raw primary data of other contexts	Uses models for other contexts	Uses unlabelled data for the target context
Model selection
Contextual weighting	X	X	A	A	X	X	A
Optimising utility function	X				A	A	A			A
Tuning classifiers for small datasets	X			X	X					A
Cascaded training	X	X		X	X					X
Learning context-specific relations between classifier outputs	X	MK	X	X		X				A
Optimising model parameters with evolutionary algorithms	X	X		X	X	X		A	X
Optimising model parameters with gradient descent	A	X		A	X			A	X
Algorithm-specific methods to shift a decision boundary	A			X	X			A	X	A
Adapting only selected parameters	X	X	X		X	X		A	X
Error weighting		X			X	X		X	A
The use of model parameters as training data	A				X				X
Vector modification	X	U	X	X	X			X		A
Modifying a similarity measure	X	U	A	A	X			A	A	A
Target context-specific combinations of cues, obtained in other contexts	X	U	A		X	X	A	A		A
Ensembles
Factor ensembles	X		X	X		X				A
Diversity-based ensembles	X	X		X	X					A
Optimising a pool of classifiers, trained on data for several contexts		X	X	X		X			X
Ensemble of generalisers			X			X	X		X
Knowledge transfer ensembles	X	X	X	X		X	A	X	A	A
Stacked ensembles	X	X	X			X
Dynamic selection of base classifiers	X	X	A	X		X	A	A
Sample-selecting ensembles	X	X		A	X			X	A	A
Context as a feature
Embedding contextual parameters as additional nodes into graphical models	X				X		X	X		A
Using historical contexts as nodes in graphical models	X				X		A	X		A
Using contextual parameters as input features	X	X		X	X	X	X	X		A
Including contextual similarity into a distance measure	X		X			X	X	X		A