Towards a Science of Causal Interpretability in Deep Learning for Software Engineering

Abstract

This dissertation addresses achieving causal interpretability in Deep Learning for Software Engineering (DL4 SE). Although Neural Code Models (NCMs) demonstrate promising performance in automating software engineering tasks, their lack of transparency on causal relationships between inputs and outputs hinders a complete understanding of their capabilities and limitations. Software researchers and, more generally, practitioners should develop the ability to explain code predictions to increment the trust and trustworthiness on NCMs. However, traditional associational interpretability, which focuses on identifying correlations, is considered insufficient for tasks requiring interventions and understanding changes’ impact. To overcome this limitation, this dissertation introduces do_code, a novel post hoc interpretability method specifically designed for NCMs. do_code leverages causal inference to provide programming language-oriented explanations of model predictions. It comprises a formal four-step pipeline: modeling a causal problem using Structural Causal Models (SCMs), identifying the causal estimand, estimating causal effects using metrics such as Average Treatment Effect (ATE), and refuting the effect estimates. The theoretical underpinnings of do_code are extensible, and a concrete instantiation is provided to mitigate the impact of spurious correlations by grounding explanations in properties of programming languages. A comprehensive case study on deep code generation across various interpretability scenarios and on the uses of different deep learning architectures demonstrates the practical benefits of do_code. The findings of the study reveal insights into the sensitivity of NCMs to changes in code syntax and NCMs’ ability to learn certain concepts of programming languages with minimized confounding bias. Next, the dissertation explores the role of associational interpretability as a foundation, examining the causal nature of software information and using information theory with tools such as COMET (a Hierarchical Bayesian Software Retrieval Model) and TraceXplainer to understand software traceability. Furthermore, the work emphasizes the importance of identifying code confounders for a more rigorous evaluation of DL4 SE models. Finally, the dissertation offers guidelines for applying causal interpretability to Neural Code Models, contributing a formal framework and practical considerations towards building more reliable and trustworthy AI in Software Engineering.