Assisting cluster coherency via n-grams and clustering as a tool to deal with the new user problem

Collaborative filtering systems typically need to acquire some data about the new user in order to start making personalized suggestions, a situation commonly referred to as the "new user problem". In this work we attempt to address the new user problem via a unique personalized strategy for prompting the user with articles to rate. Our approach makes use of hypernyms extracted from the WordNet database and proves to be converging fast to the actual user interests based on minimal user ratings, which are provided during the registration process. In addition, we explore the possible enhancement of the document clustering results, and in particular clustering of news articles from the web, when using word-based n-grams during the keyword extraction phase. We present and evaluate a weighting approach that combines clustering of news articles derived from the web, using n-grams that are extracted from the articles at an offline stage. This technique is then compared with the single minded "bag-of-words" representation that our clustering algorithm, W-kmeans, previously used. Our experimentation reveals that via fine tuning the weighting parameters between keyword and n-grams, as well as the n value itself, a significant improvement regarding the clustering results metrics can be achieved.