A New On-line Learning Algorithm For Adaptive Text Filtering

Abstract

Much previous work on text filtering is developed for batch filtering. They may not perform effectively in adaptive text filtering which is a more realistic problem. We propose a new on-line learning algorithm, known as the ATF (Adaptive Text Filtering) algorithm, to tackle the adaptive filtering problem. Our approach maintains a pool of selective terms with potentially high predictive power. The documents are retrieved by considering both the predicted relevance and its value as a training observation. The experimental result on the FBIS document corpus shows that the ATF algorithm outperforms the pure EG (Exponentiated-gradient) algorithm.