Accurate traffic forecasting is of vital importance for the management and decision in intelligent transportation systems . Indeed, it is a nontrivial endeavor to predict future traffic conditions due to the complexity of spatial relationships and temporal dependencies. Recent research developed Spatio-Temporal Graph Neural Networks (ST-GNNs) to capture the spatio-temporal correlations and achieved superior performance. However, the graph adjacency matrices that most ST-GNNs use are either pre-defined by heuristic rules or directly learned with trainable parameters. While node attributes, which record valuable information of traffic conditions, have not been fully exploited to guide the learning of better graph structure. In this paper, we propose an Ada ptive S patio- T emporal graph neural Net work, namely Ada-STNet, to first derive optimal graph structure with the guidance of node attributes and then capture the complicated spatio-temporal correlations via a dedicated spatio-temporal convolution architecture for multi-step traffic condition forecasting. Specifically, we first propose a graph structure learning component to obtain an optimal graph adjacency matrix from both macro and micro perspectives. Next, we design a dedicated spatio-temporal convolution architecture to learn spatial relationships and temporal dependencies. Moreover, we present a two-stage training strategy to improve the model performance. Extensive experimental results on real-world datasets demonstrate the effectiveness and interpretability of our approach.