An official website of the United States government
Data scientists reportedly spend 60 to 80 percent of their time in their daily routines on data wrangling, i.e. cleaning data and extracting features. However, data wrangling code is often repetitive and error-prone to write. Moreover, it is easy to introduce subtle bugs when reusing and adopting existing code, which result not in crashes but reduce model quality. To support data scientists with data wrangling, we present a technique to generate interactive documentation for data wrangling code. We use (1) program synthesis techniques to automatically summarize data transformations and (2) test case selection techniques to purposefully select representative examples from the data based on execution information collected with tailored dynamic program analysis. We demonstrate that a JupyterLab extension with our technique can provide documentation for many cells in popular notebooks and find in a user study that users with our plugin are faster and more effective at finding realistic bugs in data wrangling code.
more »
« less
CORAL: COde RepresentAtion learning with weakly-supervised transformers for analyzing data analysis
Abstract Large scale analysis of source code, and in particular scientific source code, holds the promise of better understanding the data science process, identifying analytical best practices, and providing insights to the builders of scientific toolkits. However, large corpora have remained unanalyzed in depth, as descriptive labels are absent and require expert domain knowledge to generate. We propose a novel weakly supervised transformer-based architecture for computing joint representations of code from both abstract syntax trees and surrounding natural language comments. We then evaluate the model on a new classification task for labeling computational notebook cells as stages in the data analysis process from data import to wrangling, exploration, modeling, and evaluation. We show that our model, leveraging only easily-available weak supervision, achieves a 38% increase in accuracy over expert-supplied heuristics and outperforms a suite of baselines. Our model enables us to examine a set of 118,000 Jupyter Notebooks to uncover common data analysis patterns. Focusing on notebooks with relationships to academic articles, we conduct the largest study of scientific code to date and find that notebooks which devote an higher fraction of code to the typically labor-intensive process of wrangling data in expectation exhibit decreased citation counts for corresponding papers. We also show significant differences between academic and non-academic notebooks, including that academic notebooks devote substantially more code to wrangling and exploring data, and less on modeling.
more »
« less
- Award ID(s):
- 1901386
- PAR ID:
- 10363959
- Publisher / Repository:
- Springer Science + Business Media
- Date Published:
- Journal Name:
- EPJ Data Science
- Volume:
- 11
- Issue:
- 1
- ISSN:
- 2193-1127
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract We present a Python package geared toward the intuitive analysis and visualization of paleoclimate timeseries,Pyleoclim. The code is open‐source, object‐oriented, and built upon the standard scientific Python stack, allowing users to take advantage of a large collection of existing and emerging techniques. We describe the code's philosophy, structure, and base functionalities and apply it to three paleoclimate problems: (a) orbital‐scale climate variability in a deep‐sea core, illustrating spectral, wavelet, and coherency analysis in the presence of age uncertainties; (b) correlating a high‐resolution speleothem to a climate field, illustrating correlation analysis in the presence of various statistical pitfalls (including age uncertainties); (c) model‐data confrontations in the frequency domain, illustrating the characterization of scaling behavior. We show how the package may be used for transparent and reproducible analysis of paleoclimate and paleoceanographic datasets, supporting Findable, Accessible, Interoperable, and Reusable software and an open science ethos. The package is supported by an extensive documentation and a growing library of tutorials shared publicly as videos and cloud‐executable Jupyter notebooks, to encourage adoption by new users.more » « less
-
Computational notebooks combine code, visualizations, and text in a single document. Researchers, data analysts, and even journalists are rapidly adopting this new medium. We present three studies of how they are using notebooks to document and share exploratory data analyses. In the first, we analyzed over 1 million computational notebooks on GitHub, finding that one in four had no explanatory text but consisted entirely of visualizations or code. In a second study, we examined over 200 academic computational notebooks, finding that although the vast majority described methods, only a minority discussed reasoning or results. In a third study, we interviewed 15 academic data analysts, finding that most considered computational notebooks personal, exploratory, and messy. Importantly, they typically used other media to share analyses. These studies demonstrate a tension between exploration and explanation in constructing and sharing computational notebooks. We conclude with opportunities to encourage explanation in computational media without hindering exploration.more » « less
-
Many big data systems are written in languages such as C, C++, Java, and Scala to process large amounts of data efficiently, while data analysts often use Python to conduct data wrangling, statistical analysis, and machine learning. User-defined functions (UDFs) are commonly used in these systems to bridge the gap between the two ecosystems. In this paper, we propose Udon, a novel debugger to support fine-grained debugging of UDFs. Udon encapsulates the modern line-by-line debugging primitives, such as the ability to set breakpoints, perform code inspections, and make code modifications while executing a UDF on a single tuple. It includes a novel debug-aware UDF execution model to ensure the responsiveness of the operator during debugging. It utilizes advanced state-transfer techniques to satisfy breakpoint conditions that span across multiple UDFs. It incorporates various optimization techniques to reduce the runtime overhead. We conduct experiments with multiple UDF workloads on various datasets and show its high efficiency and scalability.more » « less
-
We present in this paper an automated method to assess the quality of Jupyter notebooks. The quality of notebooks is assessed in terms of reproducibility and executability. Specifically, we automatically extract a number of expert-defined features for each notebook, perform a feature selection step, and then trained supervised binary classifiers to predict whether a notebook is reproducible and executable, respectively. We also experimented with semantic code embeddings to capture the notebooks' semantics. We have evaluated these methods on a dataset of 306,539 notebooks and achieved an F1 score of 0.87 for reproducibility and 0.96 for executability (using expert-defined features) and an F1 score of 0.81 for reproducibility and 0.78 for executability (using code embeddings). Our results suggest that semantic code embeddings can be used to determine with good performance the reproducibility and executability of Jupyter notebooks, and since they can be automatically derived, they have the advantage of no need for expert involvement to define features.more » « less
