NSF-ITR/IM PROJECT From Bits to Information: Statistical Learning Technologies for Digital Information Management Search
چکیده
Project Title: Polycategorical Categorization for Personalized Information Filtering PI: T. Hofmann Participants: Ioannis Tsochandaritis and Thomas Hofmann Abstract: Polycategorical categorization is an extension of standard classification in which items are labeled by multiple binary labels. We are particularly interested in cases with large numbers of overlapping categories and a priori unknown dependencies between labels. The main application of this approach is in personalized information filtering. We assume that a population of users rate subsets of documents from a shared repository according to their interests and preferences. Each user may label/rate documents in a different manner, but users will typically have interests in common. Currently there are two methods available: (i) collaborative filtering, which ignores the content and models the dependency structure between user ratings to make predictions, (ii) information filtering which solves the classification for each user independently by ignoring dependencies between user ratings. We propose to combine both methods by learning a prior dependency model for ratings and then use this model to generate probabilistic labels for missing ratings. The labeled data as well as the probabilistically labeled data are then used as input for a kernelbased classifier.