R vs. Python for Data Science

I disagree with the "learning curve"; if you've learned other programming languages Python has a pretty simple and familiar core, and Pandas (while the API is an inconsistent mess) is well documented. Base R is quirky compared with modern programming languages, and the API is pretty inconsistent.

I also strongly disagree with the Tidyverse bashing. I'd say it has the shortest learning curve (especially for someone familiar with SQL), and is one of the main reasons I still use R today outside of deep learning - I find it much more friendly to work with than any alternative.

You don't miss anything by skipping this.

>By contrast, just now I tried to find nearest-neighbor code for Python and at least with my cursory search, came up empty-handed; there was just one implementation that described itself as simple and straightforward, nothing fast.

> The following searches in PyPI turned up nothing: log-linear model; Poisson regression; instrumental variables; spatial data; familywise error rate; etc.

This is not how you search for things. Usually I search on google for "poisson regression scipy" if I'm looking for poisson regression.

I'm not sure I follow this, you can just set the interpreter.

I've started using rMarkdown more heavily, with reticulate & python for most data munging and r for plotting. Partly because I already know how to solve the problems I have in python more quickly than in R. The only thing I have against it at the moment is the debugging story isn't very nice by default, though I've not looked into how to improve this.

edit - if you've not looked into rmarkdown, I heartily recommend it. It is to me what the final output of notebooks should be. I can easily interleave code and descriptions, hide what I want, run it from scratch entirely as a default, and produce a range of outputs including interactive static webpages. Once web packaging is finally sorted, it'll be near perfect.

* R has Better statistical correctness based on "some dude"?

* R has better OO programming because you can print functions to the command line?

In my workplace R and Python are both well represented and I always hear from the R users that there is no "real support" for classes in R as there is in Python and that they miss it. I can't judge for myself though.

I'd be interested in seeing it included in the comparison, though (I'm afraid it would lose on multiple points if only due to lack of funding, but it is very usable).

I currently have a fairly big project written in Octave, but will likely rewrite it in python for maintainability (and would rewrite it again in something else if it grew too much for Python).

there's also Sage, which is an interesting contender, but I do not know enough about it to know how it compares. Arguably, though, Sage and Octave are more geared towards numerical computing than data science, and I think that's where they shine. So, depending on your data and the processing you need, those could be more adequate.

- Learning curve

- Machine Learning

- Parallel computation

- C/C++ interface

- Object orientation

All of these are wins for Python, some of them like the learning curve are wins by a huge margin. I should probably do a point by point rebuttal later but so many of his points are incorrect and/or poorly justified.

https://github.com/search?q=poisson+regression

I've not used Rcpp, but Pybind11 is pretty mature, and works well, so I'm not sure why he's saying it's under development; by the same measure there was an update to Rcpp last week, so that is too. He mentions Cython which allows you to compile Python code, but my main use case for that is exactly what he says - wrapping up C/C++ libraries, which is very easy in it.

"Python is currently undergoing a transition from version 2.7 to 3.x. This will cause some disruption, but nothing too elaborate."

This is pretty out of date; in Jetbrain's survey, 84% of devs had transitioned.

R

- The tidyverse ecosystem had given a huge boost to R. It had brought intuitiveness and consistency to R which was much required especially if you are a programmer coming from other languages. Also there are other ecosystems like Bioconductor which are also very mature.

- Rstudio and especially Rmarkdown notebooks are much better for reproducible analysis than Jupyter.

- It is very difficult though to develop standalone tools with R. For example it doesn't have a good argument parser.

Python

- The language is much more intuitive and more ideal for developing standalone tools.

- The ecosystem is many cases very fragmented though with a lot of libraries doing similar things.

- It lacks a good plotting system. Matplotlib is very powerful but has a very steep learning curve. In comparison ggplot2 in R is very intuitive.

Someone noted below that python is better for deploys: true. R is better for interactive use, and has a better package universe, though quality control for packages is vastly lower than something like scikit learn. R also completely dominates in classical stats, which is generally bread and butter compared to having the latest goofy neural thing. Assuming you actually do data science.

There are lots of regression options. Scipy, Scikit learn, Pymc3, PyStan.

Metaprogramming in Python is easier than in R, and arguably more predictable and consistent.

In any case, I've worked extensively in both environments and I don't think the author has considered every aspect. Below or my two cents.

- Elegance: slightly disagree. R might look more concise, but the language comes with many strange aspects (quoting, non standard evaluation) that can put a wig between novice and experienced team members. Python is more verbose, perhaps, but cleaner overall

- Learning curve: disagree. Even when working in R, modern practice would ask you to learn the tidyverse or data.table first instead of sticking with base R. Good tutorials are available for both

- Libraries: depends, the notion of "libraries" is too broad anyway, better to split it up according to the subcategories below. Both come with lots of packages, so I'd agree with it being a tie

- Statistics: agree with R. R is still the statisticians language, and many implementations of some more obscure techniques are only available in R. This being said, most ML shops today would be more interested in e.g. a good GBM implementation or deep learning rather than some robust statistics package. In R: think regression, ANOVA, significance tests, time series and niche subfields like bioengineering. In Python: think RF, GBM, t-SNE, deep learning

- Parallel computation: I'd say both are lacking, and you'd need to look more towards tooling such as Spark anyway. I'd also say out of memory computing becomes your first concern more often. Dask and Pandas on Ray are very nice on Python

- Foreign interface: kind of disagree. I think Python has matured better here

- Object oriented programming: disagree. The problem with R is in fact that is has about 4 (or more) OOP ways

- Interop: agree that you should avoid it, at the moment, it will only make deployment more cumbersome

Some other concerns I'd consider.

- Pipeline approach to ML ("model dev / model run"): better in Python. E.g. the clear approach of scikit learn to consider both preprocessing as the model itself as part of the fit-transform-predict pipeline with clear methods is way better than R. I've seen many novice R users fall into the trap of preprocessing a data set before splitting in train/test, for example. This has been one of the biggest drivers to push me towards Python coming from R. Most established libraries in Python commit to a shared, best-practice way of thinking whereas every package in R seems to come with its own ideas in terms of pipeline and usage

- Deployment: also a win for Python. Better package management / reproducibility, though it is possible in R as well

- Data exploration: I find this easier in R. Packages like dplyr help a lot here. Pandas' API is somewhat cumbersome

- Charts / visualizations: ggplot2 in R is still a champion, though good dashboarding tools exist for Python as well. Still, I find this easier to use in R

- Spatial analysis: both come with very solid libraries, though I find whipping up a quick visualization easier in R

- Deep learning: clear win for Python. Tensorflow, PyTorch and even Keras are not fun to use in R

- Reports authoring: possible in both, though R's markdown functionality combined with RStudio is fantastic. Nevertheless, Jupyter notebooks can be made to act as a reporting tool for both languages

> Given R's magic metaprogramming features (code that produces code), computer scientists ought to be drooling over R.

The benefit of this... is very debatable.