Luigi Quaranta

Fabio Calefato, Filippo Lanubile, Luigi Quaranta

Background. The rapid and growing popularity of machine learning (ML) applications has led to an increasing interest in MLOps, that is, the practice of continuous integration and deployment (CI/CD) of ML-enabled systems. Aims. Since changes may affect not only the code but also the ML model parameters and the data themselves, the automation of traditional CI/CD needs to be extended to manage model retraining in production. Method. In this paper, we present an initial investigation of the MLOps practices implemented in a set of ML-enabled systems retrieved from GitHub, focusing on GitHub Actions and CML, two solutions to automate the development workflow. Results. Our preliminary results suggest that the adoption of MLOps workflows in open-source GitHub projects is currently rather limited. Conclusions. Issues are also identified, which can guide future research work.

Luigi Quaranta, Fabio Calefato, Filippo Lanubile

Jupyter Notebook is the tool of choice of many data scientists in the early stages of ML workflows. The notebook format, however, has been criticized for inducing bad programming practices; indeed, researchers have already shown that open-source repositories are inundated by poor-quality notebooks. Low-quality output from the prototypical stages of ML workflows constitutes a clear bottleneck towards the productization of ML models. To foster the creation of better notebooks, we developed Pynblint, a static analyzer for Jupyter notebooks written in Python. The tool checks the compliance of notebooks (and surrounding repositories) with a set of empirically validated best practices and provides targeted recommendations when violations are detected.

Filippo Lanubile, Silverio Martínez-Fernández, Luigi Quaranta

Building and maintaining production-grade ML-enabled components is a complex endeavor that goes beyond the current approach of academic education, focused on the optimization of ML model performance in the lab. In this paper, we present a project-based learning approach to teaching MLOps, focused on the demonstration and experience with emerging practices and tools to automatize the construction of ML-enabled components. We examine the design of a course based on this approach, including laboratory sessions that cover the end-to-end ML component life cycle, from model building to production deployment. Moreover, we report on preliminary results from the first edition of the course. During the present year, an updated version of the same course is being delivered in two independent universities; the related learning outcomes will be evaluated to analyze the effectiveness of project-based learning for this specific subject.

Fabio Calefato, Luigi Quaranta, Filippo Lanubile et al.

Background. Due to the widespread adoption of Artificial Intelligence (AI) and Machine Learning (ML) for building software applications, companies are struggling to recruit employees with a deep understanding of such technologies. In this scenario, AutoML is soaring as a promising solution to fill the AI/ML skills gap since it promises to automate the building of end-to-end AI/ML pipelines that would normally be engineered by specialized team members. Aims. Despite the growing interest and high expectations, there is a dearth of information about the extent to which AutoML is currently adopted by teams developing AI/ML-enabled systems and how it is perceived by practitioners and researchers. Method. To fill these gaps, in this paper, we present a mixed-method study comprising a benchmark of 12 end-to-end AutoML tools on two SE datasets and a user survey with follow-up interviews to further our understanding of AutoML adoption and perception. Results. We found that AutoML solutions can generate models that outperform those trained and optimized by researchers to perform classification tasks in the SE domain. Also, our findings show that the currently available AutoML solutions do not live up to their names as they do not equally support automation across the stages of the ML development workflow and for all the team members. Conclusions. We derive insights to inform the SE research community on how AutoML can facilitate their activities and tool builders on how to design the next generation of AutoML technologies.

Luigi Quaranta

The massive trend of integrating data-driven AI capabilities into traditional software systems is rising new intriguing challenges. One of such challenges is achieving a smooth transition from the explorative phase of Machine Learning projects - in which data scientists build prototypical models in the lab - to their production phase - in which software engineers translate prototypes into production-ready AI components. To narrow down the gap between these two phases, tools and practices adopted by data scientists might be improved by incorporating consolidated software engineering solutions. In particular, computational notebooks have a prominent role in determining the quality of data science prototypes. In my research project, I address this challenge by studying the best practices for collaboration with computational notebooks and proposing proof-of-concept tools to foster guidelines compliance.

Fabio Calefato, Filippo Lanubile, Nicole Novielli et al.

The Emotion Mining Toolkit (EMTk) is a suite of modules and datasets offering a comprehensive solution for mining sentiment and emotions from technical text contributed by developers on communication channels. The toolkit is written in Java, Python, and R, and is released under the MIT open source license. In this paper, we describe its architecture and the benchmark against the previous, standalone versions of our sentiment analysis tools. Results show large improvements in terms of speed.

Luigi Quaranta, Fabio Calefato, Filippo Lanubile

Despite the widespread adoption of computational notebooks, little is known about best practices for their usage in collaborative contexts. In this paper, we fill this gap by eliciting a catalog of best practices for collaborative data science with computational notebooks. With this aim, we first look for best practices through a multivocal literature review. Then, we conduct interviews with professional data scientists to assess their awareness of these best practices. Finally, we assess the adoption of best practices through the analysis of 1,380 Jupyter notebooks retrieved from the Kaggle platform. Findings reveal that experts are mostly aware of the best practices and tend to adopt them in their daily work. Nonetheless, they do not consistently follow all the recommendations as, depending on specific contexts, some are deemed unfeasible or counterproductive due to the lack of proper tool support. As such, we envision the design of notebook solutions that allow data scientists not to have to prioritize exploration and rapid prototyping over writing code of quality.

Filippo Lanubile, Fabio Calefato, Luigi Quaranta et al.

The transition from AI/ML models to production-ready AI-based systems is a challenge for both data scientists and software engineers. In this paper, we report the results of a workshop conducted in a consulting company to understand how this transition is perceived by practitioners. Starting from the need for making AI experiments reproducible, the main themes that emerged are related to the use of the Jupyter Notebook as the primary prototyping tool, and the lack of support for software engineering best practices as well as data science specific functionalities.

Davide Fucci, Daniela Girardi, Nicole Novielli et al.

Code comprehension has been recently investigated from physiological and cognitive perspectives through the use of medical imaging. Floyd et al (i.e., the original study) used fMRI to classify the type of comprehension tasks performed by developers and relate such results to their expertise. We replicate the original study using lightweight biometrics sensors which participants (28 undergrads in computer science) wore when performing comprehension tasks on source code and natural language prose. We developed machine learning models to automatically identify what kind of tasks developers are working on leveraging their brain-, heart-, and skin-related signals. The best improvement over the original study performance is achieved using solely the heart signal obtained through a single device (BAC 87% vs. 79.1%). Differently from the original study, we were not able to observe a correlation between the participants' expertise and the classifier performance (tau = 0.16, p = 0.31). Our findings show that lightweight biometric sensors can be used to accurately recognize comprehension tasks opening interesting scenarios for research and practice.