Reproducible Data Science with R
Download as PPTX, PDF10 likes16,339 views
The document discusses reproducible data science using R, emphasizing the importance of reproducibility for saving time, enhancing scientific integrity, and facilitating collaboration. It outlines key practices such as transparent data sourcing, automated analysis pipelines, and the use of tools like R scripts and packages (e.g., tidyverse, knitr) for managing data and generating documents. The final takeaways highlight the benefits of reproducibility and provide practical steps for implementing it in data science projects.
Reproducible Data Science with R
- 1. Reproducible Data Science with R David Smith R Community Lead, Microsoft @revodavid
- 2. What is reproducibility? “Two honest researchers would get the same result” – John Mount • Transparent data sourcing and availability • Fully automated analysis pipeline (code provided) • Traceability from published results back to data 2
- 3. Why Reproducibility? • Save time • Better science • More authoritative research • Reduce risk of errors • Facilitate collaboration 3
- 4. Why Reproducibility? • Save time • Better science • More authoritative research • Reduce risk of errors • Facilitate collaboration 4
- 5. Why Reproducibility? • Save time • Better science • More authoritative research • Reduce risk of errors • Facilitate collaboration 5
- 6. Why Reproducibility? • Save time • Better science • More authoritative research • Reduce risk of errors • Facilitate collaboration 6
- 7. Why Reproducibility? • Save time • Better science • More authoritative research • Reduce risk of errors • Facilitate collaboration 7
- 8. Accessing Data • Data sources – databases, sensor logs, spreadsheets, file download, APIs, … – Remember, these are sources: don’t modify them! • Snapshot data into local static files (text is good) – You will likely include some ETL steps here – Record a timestamp of when data was extracted: Sys.time() – With big data, sample during development, scale up to finalize – Document how this was all done, preferably with code (source file)! • Import text files using R functions – Recommended package: readr (avoids issues with locale, dates, etc) 8
- 9. Analysis process • Interactively explore data & develop analyses as usual • Capture the entire process in an R script – library(tidyverse) is helpful for cleaning, feature generation, etc. – Re-usable components can be shared as (private) packages • Generate artifacts using scripts – Graphics (please, no JPGs!) – Tables – Documents • Include timestamps in output: Sys.time() 9
- 10. A reproducible analysis environment • Operating system and R version – For most purposes, not the biggest cause of issues • But do document your R session info: sessionInfo() – For production, consider VM / container • A clean R environment – Organize work into independent projects (directories) – Use relative paths in scripts – Avoid use of .Rprofile – Set explicit random seeds – Do not save R workspace 10
- 11. Managing a changing R package ecosystem • One command to lock package versions to a specific date: checkpoint("2017-04-11") (For collaborators, same command downloads required package versions) • Applies just to this project – Global package upgrades won’t break reproducibility in other projects • checkpoint package avail on CRAN, included with Microsoft R 11
- 12. Presenting results • Eliminate manual processes (as far as possible) – Annotations (graphs / tables) – Cut-and-paste into documents • Notebooks – Combines code, output and narrative – Good for collaboration with other researchers • Document Generation – Best for automating reports 12
- 13. knitr / Rmarkdown • Generate HTML, Word, or PDF reports – Or books, blogs, slides, … • Combine narrative and R code in a single document • Human-readable, easy to edit • Single-click update! 13
- 14. Collaboration and sharing • Just share R project folder • Publish on Github – Happy Git and GitHub for the useR, Jenny Bryan http://happygitwithr.com/ – Version retention and tracking – Collaboration (code and comments) 14
- 15. Take-Aways Reproducibility is Beneficial • Saves time • Produces better science • More trusted research • Reduced risk of errors • Encourages collaboration Reproducibility is Simple • Document and automate processes with R scripts • Read and clean data with tidyverse • Use checkpoint to manage package versions • Generate documents with knitr • Share reproducible projects with Github 15
- 16. Reproducible Data Science with R David Smith R Community Lead, Microsoft @revodavid