Hide/Show Apps

Using ontology based web usage mining and object clustering for recommendation

Yılmaz, Hakan
Many e-commerce web sites such as online book retailers or specialized information hubs such as online movie databases make use of recommendation systems where users are directed to items of interests based on past user interactions. Keyword-based approaches, collaborative and content filtering techniques have been tried and used over the years each having their own shortcomings. While keyword based approaches are naive and do not take content or context into account collaborative and content filtering techniques suffer from biased ratings, first item and first-rater problems. Recent approaches try to incorporate underlying semantic properties of data by employing ontology based usage mining. This thesis aims to design a recommendation system based on ontological data where web pages are seen as objects with attributes and relations. Instead of relying on users’ content ratings, user sessions are clustered on a iv semantic level to capture different behavioral groups. Since semantic information is used for the clustering distance function, each cluster represents a behavior group instead of simpler data groups. New users are then assigned to individual clusters that best represent their behavior and recommendations are generated accordingly. In this thesis we use the recommendation results as a means for measuring the effectiveness of the clusters we have generated. We have compared the results obtained using the ontological data and the results obtained without using it and shown that semantic integrating semantic knowledge increases both precision and recall.