Dependency-Guided Contrastive Tuning for Semi-supervised Code to Architecture Mapping

Sammanfattning

As software systems evolve, the link between their source files and their intended architectural modules often becomes incomplete or outdated. Recovering this link is the source-to-architecture mapping problem. The source files are all available, but assigning them to modules requires expert effort, so labeling is costly, and only a small portion of the system is realistically annotated. The challenge is therefore to map the remaining files accurately from that limited supervision. The thesis develops and evaluates a mapping method built around three complementary input features of a source-code file: its code structure, its identifier vocabulary, and its static dependency context. Each feature is transformed into a vector representation, and the three vectors are combined into a unified file embedding. A key component of the approach is dependency-guided contrastive tuning (DGCT), which adapts the code-structure encoder before classification. Using only the software dependency graph and no architectural labels, DGCT encourages files within one or two dependency hops of each other to obtain more similar representations. The resulting file embeddings are then used by a semi-supervised classifier that learnsfrom a small labeled subset and iteratively expands its training set by incorporating high-confidence predictions on unlabeled files. Our approach is evaluated on seven open-source Java systems and compared against an established Naive Bayes baseline, a lexical baseline that encodes the same three input features with a Term Frequency-Inverse Document Frequency (TF-IDF)vectorizer, and ablated versions of itself: a frozen variant whose encoders are nottuned, single-signal variants that use only one input feature, and variants in which folder and package cues are removed. Using only 5% of the available files as labeled training data, the dependency-tuned model attains the best average mapping accuracy, about 67 percent macro-averaged F1 across the systems, and performs best on four of the seven systems. It improves over its untuned counterpart on every system by roughly 6 macro-F1 points on average, surpasses the lexical baseline by about 5 macro-F1 points on average, and exceeds the established Naive Bayes baseline by around 36 macro-F1 points.