An effective model-free Gaussian Process based online social media recommendation

Nowadays, a variety of social media platforms acts as the major information portals, frequently updates the news and media pool and recommends them to dynamically arriving individuals, which requires real-time performance and immediate response to user feedback. Traditional offline/static recommendation approaches assume there exist large amount of historical interactive data, and build models from these data, which can neither deal with cold-start issue and nor fully incorporate individuals' feedbacks. Reinforcement learning based online recommendation system such as multi-armed bandit (MAB) and contextual MAB, attempts to maximize the long-term learning gains over time through exploring and exploiting the inherently interactive nature of online learning processes. However, most of work adopt the model-based paradigm, in which the feedback/reward function of individual is statically set as a specific format (e.g., linear in the features of the recommended contents and individual), which can only tailor to some specific learning objective and individual model. Without assuming the fixed reward model for each individual, this paper utilizes Gaussian process (GP) to characterize the expected feedback/reward of any recommended social media. Then, targeting at relatively accurate and noisy feedbacks of individual users, two MAB based algorithms are designed to select recommended social media to balance the effect of exploration and exploitation. The experimental results on real social media dataset Movielens demonstrate the effectiveness of our proposed methods compared with other model-based recommendation schemes.