Skip to content

Latest commit

 

History

History
127 lines (76 loc) · 4.6 KB

Python_vs_R.md

File metadata and controls

127 lines (76 loc) · 4.6 KB
Raw

python versus R

Data preprocessing

  • R requires far fewer packages, and we never have to deal with instantiating objects etc.
  • In python it is easier to split into training/testing sets, and to impute values.
  • However, everything else is easier in R.
  • OneHotEncoder, although cumbersome to use, often makes life a lot easier down the road.

Verdict: R, unless we need OneHotEncoder.

General data visualisation

  • R has many convenient built in plot functions for specific algorithms, e.g the SVM and decision tree.
  • Plots by default seem to be a bit prettier in R. But ggplot is pretty nasty to use.
  • However, the coloured scatterplot for classification is noticeably slower in R than python.
  • Seaborn is beautiful.

Verdict: python: performance combined with pretty Seaborn visuals

Linear regression

  • All of the functions are built in for R. Very intuitive syntax
  • Backward elimination to improve the model requires about 10 lines of code in python, but it is a single line builtin in R.

Verdict: Strong R

Polynomial regression

  • The method presented in the Udemy tutorial for fitting a polynomial regression model in R is terrible. However, there is in fact a built in function called poly() which makes things a lot easier.

Verdict: R, but only if you use poly()

SVR

  • The python library requires us to scale our values, while the R library e1071 takes care of it.
  • Not many differences otherwise

Verdict: R

Decision tree regression

  • Easy to implement in both languages, but it is slightly more intuitive in python.

Verdict: Python

Logistic regression

  • Much more intuitive in python.

Verdict: Python

SVM classification

  • About the same ease of use in both.
  • But R has a default plot function for the SVM classifier which, in one line, does what normally requires 15 in both R and python.

Verdict: R, because of the plot()

Decision tree classification

  • About the same difficulty in both
  • As with SVM, R has a very convenient way of visualising the tree.

Verdict: R because of the tree visualisation

Clustering

  • kmeans is a builtin function in R, and it is marginally easier to use than the python version
  • Likewise, hierarchical clustering is slightly easier in R.
  • However, the plots in python are so much prettier.

Verdict: python if you need visualisation

Association rule learning

  • The arules package in R is very easy to use, and the output is very readable.
  • This is contrasted to the particular formatting required by apyori, and the awful output.

Verdict: strong R

Reinforcement learning

  • I coded the UCB and Thompson sampling algorithms from scratch: any time I have to code something from scratch, I believe python is a better choice (but this may be biased).
  • In any case, the R code was adapted straight from the python code.

Verdict: python

Natural language processing

  • Interacting with the corpus is a pain in R.
  • However, you can clean the text and create the sparse matrix in one simple function (though Hadelin chooses not to do this).

Verdict: strong R

Neural Networks

  • Neural networks seem to be much slower in python, because h2o in R connects to an external server which does the computation for us. The difference is massive.
  • However, R still doesn't have strong support for convolutional neural networks.
  • Not shown in the course, but python can connect to h2o too.
  • Python has better support for general neural networks.

*Verdict: python, with an honourable mention to R h2o

Model selection

  • Cross validation is overall a bit nicer in python.
  • The sklearn grid search is far more intuitive, and seems to allow tuning of more different parameters, however there is a lot of hard coding required.
  • The caret library in R on the other hand is almost mystical.

Verdict: python (slightly)

XGBoost

  • The installation in python is nightmarish, but it is just install.packages('xgboost') in R!
  • That being said, I believe nowadays XGBoost is much easier to install in python
  • But after it is installed, it is just as easy to run in both R and python. Thus it's a matter of which syntax you prefer.

Verdict: slight R, but mostly because it is hard to install for python

Ambivalence

For a number of algorithms, I found that the language really doesn't make a difference, so in these cases just choose python or R based on your particular application – i.e if it's a practical project like a backend web app, go with python, but if it's a statistical research project, consider R.

  • Random forest regression and classification
  • K-nearest neighbours
  • Naïve Bayes classification