This paper proposes an Adaptive Graph Convolutional Recurrent Network (AGCRN) for traffic forecasting. The authors argue that learning node-specific patterns is essential for traffic forecasting, and that pre-defined graphs are unnecessary. To achieve this, they introduce two adaptive modules: Node Adaptive Parameter Learning (NAPL) and Data Adaptive Graph Generation (DAGG). NAPL captures node-specific patterns by learning parameters specific to each node, while DAGG infers interdependencies among traffic series automatically. AGCRN combines these modules with recurrent networks to capture fine-grained spatial and temporal correlations in traffic data. Experiments on two real-world traffic datasets show that AGCRN outperforms state-of-the-art methods significantly without pre-defined graphs. The results demonstrate that AGCRN can accurately capture spatial and temporal correlations in traffic series and achieve promising predictions. The paper also includes ablation studies showing the effectiveness of both NAPL and DAGG. The proposed method is applicable to a wide variety of multivariate/correlated time series forecasting tasks.This paper proposes an Adaptive Graph Convolutional Recurrent Network (AGCRN) for traffic forecasting. The authors argue that learning node-specific patterns is essential for traffic forecasting, and that pre-defined graphs are unnecessary. To achieve this, they introduce two adaptive modules: Node Adaptive Parameter Learning (NAPL) and Data Adaptive Graph Generation (DAGG). NAPL captures node-specific patterns by learning parameters specific to each node, while DAGG infers interdependencies among traffic series automatically. AGCRN combines these modules with recurrent networks to capture fine-grained spatial and temporal correlations in traffic data. Experiments on two real-world traffic datasets show that AGCRN outperforms state-of-the-art methods significantly without pre-defined graphs. The results demonstrate that AGCRN can accurately capture spatial and temporal correlations in traffic series and achieve promising predictions. The paper also includes ablation studies showing the effectiveness of both NAPL and DAGG. The proposed method is applicable to a wide variety of multivariate/correlated time series forecasting tasks.