Variational auto-encoder (VAE) is an efficient non-linear latent factor model that has been widely applied in recommender systems (RS). However, a drawback of VAE for RS is their inability of exploration. A good RS is expected to recommend items that are known to enjoy and items that are novel to try. In this work, we introduce an exploitation-exploration motivated VAE (XploVAE) to collaborative filtering. To facilitate personalized recommendations, we construct user-specific subgraphs, which contain the first-order proximity capturing observed user-item interactions for exploitation and the higher-order proximity for exploration. We further develop a hierarchical latent space model to learn the population distribution of the user subgraphs, and learn the personalized item embedding. Empirical experiments prove the effectiveness of our proposed method on various real-world data sets.