We propose an end-to-end deep learning framework for age estimation using face images. Our key observation is that ranking face images by age plays an important role for learning features and estimating age. We thus exploit a ranking objective jointly with an age classification objective. In this joint configuration, the ranking objective provides relative information to a deep model, that produces higher accuracy. For the ranking objective, we use a triplet ranking strategy with two novel schemes: relative triplet selection and weighted triplet ranking loss. First, the relative triplet selection expands a pool of possible triplets, enabling effective learning for ranking. Second, the weighted triplet ranking loss reflects the relativeness of age and considers its varying importance for learning. We have applied our method to two famous age estimation benchmarks, Adience and FG-NET, and demonstrated that our approach achieves meaningful improvement over the state-of-the-art age estimation methods.