Implicit feedback plays a critical role to construct recommender systems because this type of feedback is prevalent in the real-world. However, effectively utilizing implicit feedback is challenging because of positive-unlabeled or missing-not-at-random problems. To tackle these challenges, in this paper, we first show that existing approaches are biased toward the true metric. Subsequently, we provide a theoretically principled approach to handle the problems inspired by estimation methods in causal inference. In particular, we propose an unbiased estimator for the true metric of interest solving the above problems simultaneously. Experiments on two standard real-world datasets demonstrate the superiority of the proposed approach against state-of-the-art recommendation algorithms.