Reliable and Explainable Recommender Systems for Efficient Software Development

点击翻译按钮获取中文摘要

Modern software development is complex, and the number of choices developers face, such as ways to implement a feature, is often overwhelming. Developers thus often spend an enormous amount of time determining the optimal choice. This problem worsens with the flood of information provided by an increasing number of software development support tools/platforms. To filter information and improve the efficiency of software development, software engineering (SE) recommender systems, which provide suggestions for information items (code, experts, etc.) that are most likely of interest to developers, have emerged. However, despite the increasing experimental performance of existing SE recommender systems, recent surveys reveal that developers are still hesitant to adopt data-driven recommender systems due to their unstable performance in practice and inability to explain the provided recommendations. The proposed research program will create reliable and explainable SE recommender systems to enable developers to trust and fully utilize the coming generation of artificial intelligence (AI) empowered software development tools.   A reliable SE recommender should perform consistently given an evolving recommendation context. However, it is not scalable for developers to implement dedicated versions of recommenders that are suitable for each specific context of use. As such, being context-aware and adaptive are essential to achieving reliable SE recommender systems. Existing context-aware SE recommenders are far from ideal as they ignore the abstraction of the context and are unable to adapt accordingly. To fill this gap, we will design a context interpretation component for each target recommendation task and an adapter in the recommendation model that can handle a broad scope of changes leveraging implicit and explicit feedback from users. To ensure reliable SE recommenders, we will also develop new methodologies to improve software data quality.   Most existing SE recommender systems are treated as black boxes because of their unclear working mechanisms, resulting in mistrust of the systems and the need to use a time-consuming trial-and-error process to deploy a high-performance recommender system. To solve this challenge and build explainable recommenders, we will identify expected explanation forms for SE recommendation tasks by analyzing developers' online behaviours and surveying practitioners and will design machine learning models that can provide the expected explanations automatically.   Under this program, 3 PhD, 3 MSc and 2 undergraduate students will be trained in managing large software datasets and building intelligence tools to facilitate efficient software development. The program will benefit the rapidly-growing information technology industry by providing effective, reliable and explainable automation solutions for software development, thereby enhancing Canada's leadership in building an AI-empowered software development environment.