<p>Session-based recommendation (SBR) aims to predict the future action based on the user’s continuous click history. Category information can effectively improve the SBR performance. Therefore, researchers have been attempting to fuse category attribute in SBR. However, existing methods directly connect item nodes to category nodes for capturing long distance relationships in the global graph, which can make the propagation path of item information longer, causing difficulty in capturing higher-order hidden relationships between items; meanwhile, they can also indirectly connect to noisy items with the same category, resulting in inaccurate target embedding. In addition, existing researches split the learning of item and category embedding in local graphs. They neither distinguish the category similarity between neighbors, nor capture the interest transition information that is jointly determined by items and categories. This leads to indistinguishable information about neighbors of different categories and fails to accurately reflect users’ interest change. To address these challenges, we propose a novel SBR model, Category-aware Dual Channel Graph Neural Networks (CDC-GNN), which collaboratively guides the generation of recommendation results by category-aware global and local channels. Specifically, the proposed model first constructs a global category-item co-occurrence graph to directly capture global information between items. Then a novel category embedding connection strategy(CEC) is proposed in the local session graph, which can help item to distinguish the importance of their neighbors. Finally, the recommendation list is generated based on the co-guidance of global and local information on three publicly available datasets. The experimental results show that our CDC-GNN outperforms state-of-the-art models. Meanwhile, the universality test results prove that the proposed category embedding connection strategy is a general approach for existing SBR models.</p>