Go to DBLP homepageCross-Domain Multi-Label Prediction of Metamorphic Relation Patterns Leveraging Multimodal Features
Abstract: One of the core challenges in Metamorphic Testing (MT) is the automatic identification of diverse Metamorphic Relations (MRs). Machine learning-based methods have attempted to predict predefined MRs for mathematical programs at the unit testing level. However, these methods are typically limited to single MR prediction, with unclear applicability to integration testing or other domains. As an essential approach to deriving MRs, Metamorphic Relation Patterns (MRPs) generalize common MRs across domains at an abstract level, enabling domain-specific instantiation and systematic MR identification. Existing researches rely heavily on manually correlating programs with MRPs and instantiating them through domain expertise, which is both time-consuming and labor-intensive. To address these limitations, we propose RBRL-MRP, a cross-domain multi-label MRP prediction approach based on joint Ranking support vector machine and Binary Relevance with robust Low-rank learning (RBRL). RBRL-MRP represents program at both unit and integration testing levels utilizing multimodal feature fusion, followed by dimensionality reduction for optimization. A multi-label MRP set is constructed by incorporating program characteristics and associated MRs. We then train a multi-label classification (MLC) model to predict whether a program satisfies predefined MRPs. The predicted MRPs are combined using certain rules to generate candidate MRs, which are validated against the real MR set to determine the final set. Experimental results demonstrate that RBRL-MRP achieves superior performance in MRP prediction, instantiation effectiveness, and robustness across multiple programming languages and domains. Future research could integrate dynamic analysis to enhance feature representation and improve model applicability for large-scale applications.
External IDs:dblp:journals/et/LiWCJDD25
Loading