Recently, mining data streams has attracted significant attention and has been considered as a challenging task in supervised classification. Most of the existing methods dealing with this problem assume the availability of entirely labeled data streams. Unfortunately, such assumption is often violated in real-world applications given that obtaining labels is a time-consuming and expensive task, while a large amount of unlabeled instances are readily available. In this paper, we propose a new approach for handling concept-drifting data streams containing labeled and unlabeled instances. First, we use KL divergence and bootstrapping method to quantify and detect three possible kinds of drift: feature, conditional or dual. Then, if any occurs, a new classifier is learned using the EM algorithm; otherwise, the current classifier is kept unchanged. Our approach is general so that it can be applied with different classification models. Experiments performed with naive Bayes and logistic regression, on two benchmark datasets, show the good performance of our approach using only limited amounts of labeled instances.
展开▼
