From collaborative filtering to multitask learning

Recent work on collaborative filtering has led to a large number of both scalable and theoretically well founded algorithms. In this paper, we show that collaborative filtering and multitask learning are innately closely connected. In particular, the 'learning the kernel' paradigm in multitask learning turns out to be identical to a Ky-Fan norm minimization. This allows us to “import” collaborative filtering techniques into multitask learning and vice versa; in particular, we solve a multitask learning problem where the tasks also have features. We show the feasibility of our approach on two large real-world multitask learning applications. Joint work with Markus Weimer, Wei Chu, Deepayan Chakrabarti.