<p>Accurate and real-time short-term urban traffic flow prediction is vital for traffic planning and resident commuting. To explore the complex internal and external attributes of traffic flow and establish their dependencies, this paper proposes a multi-attribute-based short-term traffic flow prediction mode. First, considering the heterogeneity of connection strengths between adjacent nodes in a road network, a node influence assessment method based on network representation learning (SDNE) is introduced to depict these connections. Second, addressing the complex spatio-temporal correlations of traffic flow, a spatio-temporal feature mining component based on node connection strength is constructed. It combines graph convolutional neural networks, which uncover hidden spatial relationships in road networks, with dilated causal convolutional neural networks, capturing long-term sequential dependencies. Finally, considering the significant impact of external attributes on traffic flow, a NCSMGCN model is proposed, leveraging multi-head attention mechanisms for efficient integration of spatio-temporal and external attribute features, further improving prediction accuracy. Experimental results show that the proposed model effectively mines the complex internal and external attributes and their dependencies, improving traffic flow prediction accuracy.</p>