Abstract
Group recommendation is a task to recommend items to groups such as households and communities. In this paper, we propose a non-linear matrix factorization method for group recommendation. The proposed method assumes that each member in groups has its own latent vector, and behavior of each group is determined by the probability distribution of the members' latent vectors. Recommending items is performed by using non-linear functions that map the distributions of the groups into scores for items. The non-linear functions are generated from Gaussian processes, which are defined by the similarities between distributions of the groups. We can efficiently calculate the similarities by embedding each distribution as an element in a reproducing kernel Hilbert space. We demonstrate the effectiveness of the method using two synthetic datasets and two real datasets in two prediction tasks.
