Active Learning with Boosting for Spam Detection

Spam detection algorithms have been developed to train in a large enough set of labeled data and predict with a high accuracy of 95% if an email is spam or not. A problem that arises in this setting is that labeling examples is a costly process. It requires humans to read them one by one and classify them. Active learning is a learning approach developed to address this problem. It learns a small set of labeled examples and implements strategies to improve learning by labeling additional examples. Its goal is to provide a good classifier with as small as possible a number of examples. Boosting algorithms intuitively fit well into the active learning framework since they are good at distinguishing significant examples. In this report, I study performance and behavior of active learning in combination with four boosting algorithms: AdaBoost, AdaBoost*,LPBoost, ERLPBoost.