Norm Matloff, Prof. of Computer Science, UC Davis; my bio
Hello! This Web page is aimed at shedding some light on the perennial R-vs.-Python debates in the Data Science community. As a professional computer scientist and statistician, I hope to shed some useful light on the topic. I have potential bias: I've written four R-related books, I've given a keynote talk at useR!; I currently serve as Editor-in-Chief of the R Journal; etc. But I hope this analysis will be considered fair and helpful.
Clear win for Python.
This is subjective, of course, but having written (and taught) in many different programming languages, I really appreciate Python's greatly reduced use of parentheses and braces:
if x > y:
z = 5
w = 8vs.
if (x > y)
{
z = 5
w = 8
}Python is sleek!
Huge win for R.
This is of particular interest to me, as an educator. I've taught a number of subjects -- math, stat, CS and even English As a Second Language -- and have given intense thought to the learning process for many, many years.
To even get started in Data Science with Python, one must learn a lot of material not in base Python, e.g., NumPy, Pandas and matplotlib. These libraries require a fair amount of computer systems sophistication.
By contrast, matrix types and basic graphics are built-in to base R. The novice can be doing simple data analyses within minutes.
Python libraries can be tricky to configure, even for the systems-savvy, while most R packages run right out of the box.
Slight edge to R.
CRAN has over 14,000 packages. PyPI has over 183,000, but it seems thin on Data Science.
For example, I once needed code to do fast calculation of nearest-neighbors of a given data point. (NOT code using that to do classification.) I was able to immediately find not one but two packages to do this. By contrast, just now I tried to find nearest-neighbor code for Python and at least with my cursory search in PyPi, came up empty-handed; there was just one implementation that described itself as simple and straightforward, nothing fast.
The following (again, cursory) searches in PyPI turned up nothing: EM algorithm; log-linear model; Poisson regression; instrumental variables; spatial data; familywise error rate; etc.
This is not to say no Python libraries exist for these things; I am simply saying that they are not easily found in PyPI, whereas it is easy to find them in CRAN.
The fact that R has a canonical package structure is a big advantage.
Slight edge to Python here.
The R-vs.-Python debate is largely a statistics-vs.-CS debate, and since most research in neural networks has come from CS, available software for NNs is mostly in Python. RStudio has done some excellent work in developing a Keras implementation, but so far R is limited in this realm.
On the other hand, random forest research has been mainly pursued by the stat community, and in this realm I'd submit that R has the superior software. R also has excellent packages for gradient boosting.
I give the edge to Python here because for many people, machine learning means NNs.
Big win for R.
In my book, the Art of R Programming, I made the statement, "R is written by statisticians, for statisticians," which I'm pleased to see pop up here and there on occasion. It's important!
To be blunt, I find the machine learning people, who mostly advocate Python, often have a poor understanding of, and in some cases even a disdain for, the statistical issues in ML. I was shocked recently, for instance, to see one of the most prominent ML people, state in his otherwise outstanding book that standardizing the data to mean-0, variance-1 means one is assuming the data are Gaussian — absolutely false and misleading.
Let's call it a tie.
Neither the base version of R nor Python have good support for multicore computation. Threads in Python are nice for I/O, but parallel computation using them is impossible, due to the infamous Global Interpreter Lock. Python's multiprocessing package is not a good workaround, nor is R's 'parallel' package. External libraries supporting cluster computation are OK in both languages.
Currently Python has better interfaces to GPUs.
Slight win for R.
Though there are tools like swig etc. for interfacing Python to C/C++, as far is I know there is nothing remotely as powerful as R's Rcpp for this at present. The Pybind11 package is being developed.
In addition, R's new ALTREP idea has great potential for enhancing performance and usability.
On the other hand, the Cython and PyPy variants of Python can in some cases obviate the need for explicit C/C++ interface in the first place; indeed some would say Cython IS a C/C++ interface.
Slight win for R.
For instance, though functions are objects in both languages, R takes that more seriously than does Python. Whenever I work in Python, I'm annoyed by the fact that I cannot print a function to the terminal, which I do a lot in R.
Python has just one OOP paradigm. In R, you have your choice of several (S3, S4, R6 etc.), though some may debate whether this is a good thing.
Given R's magic metaprogramming features (code that produces code), computer scientists ought to be drooling over R. But most CS people are unaware of it.
Horrible loss for R.
(During the Twitter discussion of this essay, this section nearly tripled in size, as more and more issues came up that I thought should be addressed. Soon I will move this to a separate document, leaving just a summary here.)
Python is currently undergoing a transition from version 2.7 to 3.x. This will cause some disruption, but nothing too elaborate.
By contrast, R is rapidly devolving into two mutually unintelligible dialects, ordinary R and the Tidyverse. I, as a seasoned R programmer, cannot read Tidy code, as it calls numerous Tidyverse functions that I don't know. Conversely, as one person in the Twitter discussion of this document noted (approvingly), "One can code in the Tidyverse while knowing very little R."
Note that, in using the term "Tidyverse," I am not including pre-existing projects by Hadley Wickham, e.g. ggplot2 (which I use a lot) and dplyr. Instead, I am referring to things such as tibbles and pipes. I am especially concerned about the plethora of extra functions Tidyverse advocates burden students with learning, on top of rudimentary base R. These functions are unnecssary, as most can be done simply in base R.
Though advocates claim it works well in teaching beginning programmers, there has been no study of this claim, and I think the perceived success is psychological, a Bandwagon Effect. On the contrary, I believe it is quite the opposite, i.e. using the Tidyverse makes things more difficult for learners without prior programming background.
As noted, the students are being asked to learn a much larger volume of material, which is clearly bad pedagogy. See also "The Tidyverse Curse", in which the author says inter alia that he uses "only" 60 Tidyverse functions -- 60! Similarly, it is bad pedagogy to force students to learn tibbles, a more complex technology, instead of data frames, a simpler one.
Again, I am certainly not saying one should only use base R; on the contrary, see my point above that CRAN is a major advantage of R. But the Tidyverse should be considered advanced R, not for beginners, just as is the case for most complex CRAN packages.
The Tidyverse is a vigorous promotional effort by a commercial entity that has come to dominate the R world, RStudio. I know and admire the people at RStudio, but a commercial entity should not have such undue influence on an open-source project.
For instance, consider the lightning-fast data.table package. Rather than welcoming it as a hugely valuable contribution to R, RStudio has treated it as a competitor, downplaying it and promoting their own product, dplyr. This is simply not healthy for an open-source language.
In addition, all this is causing tension between some leaders in the two camps. All this does not augur well for the future of the language.
Win for Python.
Not a big issue in Data Science, but it does come up in some contexts.
Classical computer science data structures, e.g. binary trees, are easy to implement in Python. It is not part of base R, but can be done in various ways, e.g. the datastructures package, which wraps the widely-used Boost C++ library.
Big win for R.
To begin with, R's basic help() function is much more informative than Python's. It's nicely supplemented by example(). And most important, the custom of writing vignettes in R packages makes R a hands-down winner in this aspect.
RStudio is to be commended for developing the reticulate package, to serve as a bridge between Python and R. It's an outstanding effort, and works well for pure computation. But as far as I can tell, it does not solve the knotty problems that arise in Python, e.g. virtual environments and the like.
At present, I do not recommend writing mixed Python/R code.
I have a quick tutorial on R for non-programmers, an evolving project. I also have a book, the Art of R Programming, NSP, 2011.
I have a tutorial on Python, for those with a strong programming background.
This document has benefited from various reader comments, notably from Dirk Eddelbuettel, as well as Paul Hewson, Bob Muenchen and Inaki Ucar.