Report

Learning under concept drift: an overview Zhimin He iTechs – ISCAS 2013-03-21 Agenda What’s Concept Drift Causes of a Concept Drift Types of Concept Drift Detecting and Handling Concept Drift Implications for Software Engineering Research Definitions Prediction p 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) instance. The task is to predict a label yt+1 for the target instance Xt+1. Definitions(cont.) 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: p i i 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 ))} 1 1 2 2 k k Causes of Concept Drift (cont.) p (c ) p ( X | c ) p (c | X ) p( X ) i i i 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 type. 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 predictions 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] References 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.” KDD,1999