Recovery Analysis for Adaptive Learning from
Non-stationary Data Streams
AmmarShakerandEykeHüllermeier
Abstract. The extension of machine learning methods from static to dynamic en-
vironments has received increasing attention in recent years; in particular, a large
numberofalgorithmsforlearningfromso-called data streams hasbeendeveloped.
An important property of dynamic environments is non-stationarity, i.e., the as-
sumptionofanunderlyingdatageneratingprocessthatmaychangeovertime.Cor-
respondingly,theabilitytoproperlyreacttoso-called concept change isconsidered
asanimportantfeatureoflearningalgorithms.Inthispaper,weproposeanewtype
ofexperimentalanalysis,called recovery analysis,whichisaimedatassessingthe
ability of a learner to discover a concept change quickly, and to take appropriate
measurestomaintainthequalityandgeneralizationperformanceofthemodel.
1 Introduction
The development of methods for learning from so-called data streams hasbeena
topicofactiveresearchinrecentyears[6,9].Roughlyspeaking,thekeyideaisto
haveasystemthatlearnsincrementally,andmaybeeveninreal-time,onacontinu-
ousandpotentiallyunboundedstreamofdata,andwhichisabletoproperlyadapt
itself to changes of environmental conditions or properties of the data generating
process. Systems with these properties have already been developed for different
machinelearninganddataminingtasks,suchasclusteringandclassification[7].
Anextensionofdataminingandmachinelearningmethodstothesettingofdata
streamscomeswithanumberofchallenges.Inparticular,thestandard“batchmode”
of learning, in which the entire data as a whole is to provided as an input to the
learningalgorithm(or“learner”forshort),isnolongerapplicable.Correspondingly,
the learner is not allowed to make several passes through the data set, which is
commonly done by standard methods in statistics and machine learning. Instead,
AmmarShaker · EykeHüllermeier
DepartmentofMathematicsandComputerScience,UniversityofMarburg,Germany
e-mail: {shaker,eyke}@mathematik.uni-marburg.de
R.Burduketal.(Eds.): CORES 2013,AISC226,pp.289–298.
DOI:10.1007/978-3-319-00969-8_ 28 © Springer International Publishing Switzerland 2013