Demystifying a Dark Art: Understanding Real-World Machine Learning Model Development

Abstract

It is well-known that the process of developing machine learning (ML) workflows is a dark-art; even experts struggle to find an optimal workflow leading to a high accuracy model. Users currently rely on empirical trial-and-error to obtain their own set of battle-tested guidelines to inform their modeling decisions. In this study, we aim to demystify this dark art by understanding how people iterate on ML workflows in practice. We analyze over 475k user-generated workflows on OpenML, an open-source platform for tracking and sharing ML workflows. We find that users often adopt a manual, automated, or mixed approach when iterating on their workflows. We observe that manual approaches result in fewer wasted iterations compared to automated approaches. Yet, automated approaches often involve more preprocessing and hyperparameter options explored, resulting in higher performance overall-suggesting potential benefits for a human-in-the-loop ML system that appropriately recommends a clever combination of the two strategies.

"We Have No Idea How Models will Behave in Production until Production": How Engineers Operationalize Machine Learning

“…Hill et al [29] interview ML application developers and report challenges related to building first versions of ML models, especially around the early stages of exploration and experimentation (e.g., feature engineering, model training). They describe the process of building models as "magic"-similarly echoed by Lee et al [50] when analyzing ML projects from Github-with unique practices of debugging data in addition to code. Serban et al [88] conduct a survey of practitioners and list 29 software engineering practices for ML, such as "Use Continuous Integration" and "Peer Review Training Scripts."…”

“…Prior work has extensively documented the Agile tendencies of MLEs, describing how they iterate quickly (i.e. with velocity) to explore a large ML or data science search space [3,30,50,74,110]. Amershi et al [3] describe how experimentation can be sped up when labels are annotated faster (i.e., rapid data preparation).…”

Hindsight Logging for Model Training

“…There is a wide range of research and tooling being developed to support these many tasks. ML lifecycle management is especially challenging because it involves many cycles of trial-and-error [30], and its dependencies are hard to scope [53]. When something goes wrong, ML engineers may need to rollback their model to an earlier version [41,64], inspect old versions of the training data [24,27,37], or audit the code that was used for training [40,51].…”

Leam: An Interactive System for In-situ Visual Text Analysis

“…We document these challenges in Section 1. These challenges are a culmination of (1) our conversations with practitioners while working in an industry research lab, part of a larger company with more than three hundred e-commerce subsidiaries, (2) prior research, particularly those reporting from interview studies on data analysis workflows [20,38], and (3) our experience in developing and evaluating interactive data systems. As such, the list of challenges here is intended to be a useful guide informing research and development on text analytics systems, not a comprehensive enumeration, and inevitably reflects our personal taste.…”