Collaborative filtering (CF) is one of the most efficient ways for recommender systems. Typically, CF-based algorithms analyze users' preferences and items' attributes using one of two types of feedback: \textit{explicit feedback} (e.g., ratings given to item by users, like/dislike) or \textit{implicit feedback} (e.g., clicks, views, purchases). Explicit feedback is reliable but is extremely sparse; whereas implicit feedback is abundant but is not reliable. To leverage the sparsity of explicit feedback, in this paper, we propose a model that efficiently combines explicit and implicit feedback in a unified model for rating prediction. This model is a combination of \textit{matrix factorization} and \textit{item embedding}, a similar concept with word-embedding in natural language processing. The experiments on three real-datasets (\textit{Movilens 1M}, \textit{Movielens 20M}, and \textit{Bookcrossing}) demonstrate that our method can efficiently predict ratings for items even if the ratings data is not available for them. The experimental results also show that our method outperforms competing methods on rating prediction task in general as well as for users and items which have few ratings.