Cost-sensitive learning for social network analysis and network measurement

Abstract

Recently, the application of data mining techniques to social network analysis and network measurement has received considerable attention. In this work, an aggregating N-dependence estimator (ANDE)-based cost-sensitive classification algorithm (CS_ANDE) was proposed for use with the unbalanced data commonly observed in social networks and network measurements. First, a one-dependence estimator was adopted to determine the approximate cost of misclassification. Second, multiple classifiers were constructed to minimize the misclassification cost. Subsequently, these classifiers were used to re-label samples. Ultimately, a CS_AODE classifier was obtained by learning these re-labeled samples. Consistent with two-dependence estimators, a CS_A2DE classifier was acquired. The CS_AODE and CS_A2DE classifiers were empirically evaluated against MetaCost and AODE for UCI datasets, and the results indicated that CS_AODE and CS_A2DE significantly outperformed the other classifiers and that the performance was stable under different parameters.