Scalable text classification as a tool for personalization

We consider scalability issues of the text classification problem where by using (multi)-labeled training documents, we try to build classifiers that assign documents into classes permitting classification in multiple classes. A new class of classification problems; called 'scalable', is introduced, with applications on web mining. Scalable classification utilizes newly classified instances in order to improve the accuracy of future classifications and capture changes in semantic representation of different topics. In addition, definition of different similarity classes is allowed, resulting in a 'per-user' classification procedure. Such an approach provides a new methodology for building personalized applications. This is due to the fact that the user becomes a part of the classification procedure. We explore solutions for the scalable text classification problem and introduce an algorithm that exploits a new text analysis technique that decomposes documents into the vector representation of their sentences according to the user expertise. Finally, a web-based personalized news categorization system that bases upon this approach is presented.