Online convex combination of ranking models

Abstract As a task of high importance for recommender systems, we consider the problem learning convex combination ranking algorithms by online machine learning. First, propose stochastic optimization algorithm that uses finite differences. Our new achieves close to optimal empirical performance two base rankers, while scaling well with an increased number models. In our experiments five real-world recommendation data sets, show offers significant improvement over previously known techniques. The proposed is first effective method combining ranked lists Secondly, exponentially weighted based on grid space weights. We has near-optimal worst-case bound. bound provides theoretical guarantee non-convex bandits using limited evaluations under very general conditions.