Among various traffic data modeling and predicting methods, graph learning-based models attract more attention, because of their powerful representation ability for modeling spatial and temporal dependencies with graph neural networks. Despite their promising performance, several key problems have not been well addressed: 1) Sensed data are often noisy in many real transportation scenarios. 2) The spatio-temporal correlations of traffic data are complex and dynamic, especially for long-term modeling and predicting. In such cases, existing methods may not lead to satisfactory prediction results. In this paper, we aim at the above problems by exploring a Spatio-Temporal Heterogeneous Graph Neural Network With Multi-View Learning Framework(MVJGL) for traffic prediction. In particular, we first model different types of traffic features and construct multiple graph structures. Then, we design two parallel heterogeneous gated temporal convolution modules to extract long and short-term temporal dependencies from different traffic features, respectively. Moreover, we introduce parallel graph convolutions to cross-characterize the time-varying spatial dependencies of each view. Extensive experiment results on four real traffic datasets show the superior performance grain obtained by the proposed model.