Search

Features in product lines and highly configurable systems can interact in ways that are contrary to developers' intent. Current methods to identify such unanticipated feature interactions are costly and inadequate. To address this problem we propose a new approach to learn feature interactions, both in those product lines where constraints on feature combinations are specified and in feature-rich configurable systems where such specifications often are not available. The contribution of the paper is to use program analysis to extract feature-relevant learning models from the source code in order to detect unwanted feature interactions. Where specifications of feature constraints are unavailable, our approach infers the constraints using feature-related data-flow dependency information. Evaluation in experiments on three software product line benchmarks and a highly configurable system shows that this approach is fast and effective. The contribution is to support developers by automatically detecting feature combinations in a new product or version that can interact in unwanted or unrecognized ways. This enables better understanding of latent interactions and identifies software components that should be tested together because their features interact in some configurations.

An ongoing challenge for the requirements engineering of software product lines is to predict whether a new combination of features (units of functionality) will create an unwanted or even hazardous feature interaction. We thus seek to improve and automate the prediction of unwanted feature interactions early in development. In this paper, we show how the detection of unwanted feature interactions in a software product line can be effectively represented as a link prediction problem. Link prediction uses machine learning algorithms and similarity scores among a graph's nodes to identify likely new edges. We here model the software product line features as nodes and the unwanted interactions among the features as edges. We investigate six link-based similarity metrics, some using local and some using global knowledge of the graph, for use in this context. We evaluate our approach on a software product line benchmark in the literature, building six machine-learning models from the graph-based similarity data. Results show that the best ML algorithms achieved an accuracy of 0.75 to 1 for classifying feature interactions as unwanted or wanted in this small study and that global similarity metrics performed better than local similarity metrics. The work shows how link-prediction models can help find missing edges, which represent unwanted feature interactions that are undocumented or unrecognized, earlier in development.