DeepFuseLoc: A Deep Neural Framework with KAN-based Feature Crossing for Semantic and Structural Bug Localization

Accurately locating buggy source code files from textual bug reports remains a challenging task in software maintenance, as it requires capturing complex semantic and structural correspondences between natural language and code representations. To address this challenge, we propose DeepFuseLoc , a novel deep neural network-based approach that fuses multi-level textual and structural semantics through adaptive feature interaction. DeepFuseLoc introduces a KAN-based Feature Crossing Network (KCN) to model high-order nonlinear relationships among features, and employs a dynamic weighting mechanism to adaptively balance information matching scores and snippet matching scores, allowing the model to automatically focus on salient features for different bug types. Comprehensive experiments conducted on four real-world Java projects show that DeepFuseLoc consistently outperforms state-of-the-art baselines in both accuracy and robustness, achieving substantial improvements in Top- k accuracy, MAP, and MRR. Ablation studies and component analysis further validate that both the hierarchical information matching mechanism and the KCN-based nonlinear fusion module contribute substantially to the performance improvements, operating in a complementary manner to enhance model effectiveness.Overall, DeepFuseLoc provides an effective solution for automated bug localization, advancing the integration of deep semantic modeling and structural analysis toward higher software quality and lower maintenance effort.