Car-sharing issue is a popular research field in sharing economy. In this paper, we investigate the car-sharing relocation problem (CSRP) under uncertain demands. Normally, the real customer demands follow complicating probability distribution which cannot be described by parametric approaches. In order to overcome the problem, an innovative framework called Data-Driven Kernel Stochastic Programming (DDKSP) that integrates a non-parametric approach - kernel density estimation (KDE) and a two-stage stochastic programming (SP) model is proposed. Specifically, the probability distributions are derived from historical data by KDE, which are used as the input uncertain parameters for SP. Additionally, the CSRP is formulated as a two-stage SP model. Meanwhile, a Monte Carlo method called sample average approximation (SAA) and Benders decomposition algorithm are introduced to solve the large-scale optimization model. Finally, the numerical experimental validations which are based on New York taxi trip data sets show that the proposed framework outperforms the pure parametric approaches including Gaussian, Laplace and Poisson distributions with 3.72% , 4.58% and 11% respectively in terms of overall profits.