Graph-based representations have gained attention for their ability to model structural and semantic information capturing relevant characteristics and features of source code for training deep learning models. However, existing methods face limitations: they lack flexibility in constructing cross-language graphs, produce non-interoperable outputs, and generate excessively large graphs, hindering their adoption and raising scalability and efficiency issues in graph-based neural network training.

In this work, we introduce CONCORD, a domain-specific language (DSL), to address these challenges. We aim to 1) enable customizable graph-based code representations across programming languages, 2) reduce graph size complexity through simplification heuristics, and 3) improve scalability and reproducibility in software engineering tasks.

CONCORD provides a configurable DSL to automate graph construction and implements heuristics to reduce graph size while preserving critical information. We evaluate its effectiveness on two tasks: code smell detection and vulnerability detection. For each, we compare performance and graph size against baseline representations without simplification heuristics.

On code smell detection, CONCORD preserved 95.1% of baseline performance, and exceeded it by 5% in one setting, while reducing, on average, the number of nodes and edges by 13.11% and 13.67%, respectively. For vulnerability detection, it improved performance by 3.65% over the baseline while reducing the number of nodes and edges by 3.63% and 3.62%, respectively. This demonstrates that CONCORD’s heuristics maintain or enhance performance while improving scalability.

CONCORD represents a step towards advancing graph-based code analysis by offering a flexible, language-agnostic approach to generate streamlined code representations. Its simplification heuristics balance performance and scalability, enabling efficient training of neural models without sacrificing accuracy. In addition, it reduces development overhead, promotes reproducibility through standardized representations, and broadens accessibility to graph-based methods for software engineering tasks.