Learning under concept drift: an overview

Learning under concept drift: an overview
Zhimin He
iTechs – ISCAS
What’s Concept Drift
Causes of a Concept Drift
Types of Concept Drift
Detecting and Handling Concept Drift
Implications for Software Engineering Research
 Prediction
 X t  R is a vector in p-dimensional feature space
observed at time t and yt is the corresponding label.
 We call Xt an instance and a pair (Xt; yt) a labeled
instance. We refer to instances (X1; : : : ;Xt) as
historical data and instance Xt+1 as target (or testing)
 The task is to predict a label yt+1 for the target
instance Xt+1.
 Concept Drift
 Every instance Xt is generated by a source St.
 If all the data is sampled from the same source, i.e. S1
= S2 = : : : = St+1 = S we say that the concept is stable.
 If for any two time points i and j Si != Sj , we say that
there is a concept drift.
Causes of Concept Drift
 Let X  R is an instance in p-dimensional
feature space. X  c , where c1, c2,….ck is the
set of class labels.
 The optimal classier to classify X  c is
determined by a prior probabilities for the
classes P(ci) and the class-conditional
probability density functions p(X | ci), i = 1,….k.
 Concept /data source:
 a set of a prior probabilities of the classes and classconditional pdf's:
S  {( P (c ), P ( X | c )),( P(c ), P( X | c )),...( P(c ), P( X | c ))}
Causes of Concept Drift (cont.)
p (c ) p ( X | c )
p (c | X ) 
p( X )
 Concept drift may occur in three ways:
 Class priors P(c) might change over time.
 The distributions of one or several classes p(X|ci)
might change. (virtual drift)
 The posterior distributions of the class memberships
p(ci|X) might change.(real drift)
Types of Concept Drift
 Types:
 Sudden drift
 Gradual drift
 Incremental drift
 reoccurring contexts
Detecting and Handling Concept Drift
 Detecting
 Monitoring the raw data
 Monitoring parameters of learners
 Monitoring prediction errors of learners
 Handling
 Ensemble learning
 Instance selection
 Instance weights
 Training windows
 Training windows are naturally suitable for sudden concept
drift, while ensembles are more flexible in terms of change
Detecting and Handling Concept Drift (cont.)
 Overall solution for learning under concept drift
Implications for SE Research
 Concept drift is a fundamental issue for SE
 Cost estimation, defect prediction…
 Especially in the cross-company/cross-project context
 Be harmful to performance of prediction models
 Detecting and handling concept drift is a
challenging task!
 Quality problems of SE data, e.g., insufficient data
 Data generation context is highly unstable.
 Has become a increasingly popular research
topic in SE field!
 E.g., Burak Turhan [JESE 2012], Jayalath Ekanayake
[MSR 2009, JESE 2011]
1.Indre Zliobaite, “Learning under Concept Drift : an
Overview,” Tech-report, 2009
2.A. Dries and R. Ulrich, “Adaptive Concept Drift
Detection,” Journal of Statictical Analysis and Data
Mining, 2009
3.L. Minku, A. White, and X. Yao. “The impact of diversity
on on-line ensemble learning in the presence of concept
drift.” IEEE Transactions on Knowledge and Data
Engineering, 2009
4.M. Kelly, D. Hand, and N. Adams. “The impact of
changing populations on classier performance.”

similar documents