Traffic prediction has drawn considerable attention due to its potential to optimize the operational efficiency of road networks. Existing methods commonly combine graph neural network (GNN) and recurrent neural network (RNN) to model spatio-temporal correlations. However, the above models still face challenges, including an inability to capture time-varying spatial correlations, inadequate consideration of spatio-temporal heterogeneity and inefficient iterative operations. To address the above challenges, in this paper, we propose a novel framework for traffic prediction, named time-based adaptive graph neural network (TAGNN). First, a novel graph learning module was developed to generate time-based adaptive graph dependency matrices, which capture hidden spatial correlations at different time steps. Second, two embedding matrices are proposed to assist the model in capturing spatio-temporal heterogeneity by attaching essential external features. Third, a temporal convolution module is proposed to capture temporal correlations by stacking grouped convolution. The receptive field expands exponentially with each additional layer, reducing parameters and improving prediction efficiency. Extensive experimental results demonstrate that our model adequately extracts the spatio-temporal correlation of nodes while ensuring prediction efficiency.