<p>Dynamic community detection is essential for uncovering hierarchical structures and evolutionary patterns in complex networks. However, mainstream approaches based on evolutionary nonnegative matrix factorization (NMF) fail to explicitly model and track the temporal evolution of inter-community interactions. This limitation compromises interpretability and hinders the accurate capture of complex dynamic features. To address these issues, this paper proposes a dynamic community detection method based on evolutionary nonnegative matrix tri-factorization (Evo-NMTF). By decomposing the network adjacency matrix into a node-community affiliation matrix and a community association matrix, our approach enables explicit modeling and quantitative analysis of both intra- and inter-community interactions. A dual-path smoothness constraint mechanism is introduced to preserve temporal continuity at the node and community levels. Additionally, sparse regularization is incorporated to enhance robustness and partition clarity. Experimental results on multiple synthetic and real-world dynamic networks show that Evo-NMTF outperforms state-of-the-art baselines across various evolution patterns and noise conditions, as evaluated by normalized mutual information (NMI) and modularity. Ablation studies further verify that the dual-path constraints and sparsity strategy are key contributors to the performance gains. This work presents a reliable, interpretable, and robust framework for dynamic network analysis.</p>