Analyzing and Securing Social Networks

492 Analyzing and Securing Social Networks

the distributions of training and test data are the same. However, in data streams, this assump-

tion is violated because of concept drift. Simple majority voting is therefore a better alternative.

Our experiments conrm this in practice, obtaining better results with simple rather than weighted

majority voting.

We have shown in Masud et al. (2011) that EMPC can further reduce the expected error in clas-

sifying concept-drifting data streams compared with SPC approaches, which use only one data

chunk for training a single classier (i.e., r = v = 1). Intuitively, there are two main reasons for the

error reduction. First, the training data per classier is increased by introducing the multichunk

concept. ...