Crowdsourcing is a new distributed computing paradigm that leverages the wisdom of crowd and the voluntary human effort to solve problems or collect data. In this context, trustworthiness of user contributions is of crucial importance to the viability of crowdsourcing. Prior mechanisms either do not consider the trustworthiness or quality of contributions or have to assess it only after workers' submission of contributions, which results in irreversible effort expenditure and negative player utilities. In this paper, we propose a reputation system, EndorTrust, to not only assess but also predict the trustworthiness of contributions without wasting workers' effort. The key approach is to explore an inter-worker relationship called endorsement to improve trustworthiness prediction using machine learning methods, while also taking into account the heterogeneity of both workers and tasks.