Enhancing Architecture Recovery for Mixed-Source Software

Mixed-source software combines autonomous code with reused open-source modules, posing new challenges for architecture recovery. Existing architecture recovery research largely assumes provenance-neutral settings, and mixed-source software architecture recovery has not yet been systematically addressed. As a result, directly applying conventional methods to mixed-source systems exposes two key limitations and calls for enhancing conventional recovery methods for mixed-source software. First, reused open-source code carries implicit component priors that conventional methods do not exploit. In mixed-source systems, open-source functionality is typically reused as cohesive units (e.g., libraries or submodules) whose internal modular organization has been shaped by long-term community evolution, suggesting that entities within the same reused unit should remain largely co-located in the recovered architecture. Source-agnostic recovery may ignore this prior and fragment reused modules, thereby reducing recovery precision. Second, conventional methods often produce mixed components that entangle autonomous and reused entities. This cross-source entanglement blurs reuse boundaries and obscures how open-source modules are introduced and interfaced with autonomous code, reducing the recovered architecture's usefulness. Preserving appropriate cross-source boundaries retains mixed-source semantics and better supports refactoring and autonomy strengthening. To address these issues, we propose REFINE-Mixed, a mixed-source-aware architecture recovery approach that extends a conventional architecture recovery method with two mixed-source enhancements: incorporating explicit source information as an additional recovery signal to capture provenance-induced component priors, and introducing an adaptive cross-source penalty in clustering to discourage weakly justified cross-source merges while still allowing well-supported merges. Experiments on a mixed-source architecture dataset demonstrate that REFINE-Mixed consistently improves recovery precision over representative baselines.