Python has good bindings to venerable libraries in C and fortran that nobody would want to implement from scratch. Python libs like numpy and scipy were very convenient ways to call the highly optimized code in say BLAS. Then scikit-learn came and made machine learning easy for the masses. Before, libs in R would provide different methods but they were anything but unified (things like CARET notwithstanding).<p>In the last 2-3 years the main driver was deep learning. Since the whole point was to run on a GPU and CUDA remains a pain of low-level C, python made it very palatable.<p>R has lost the race by not jumping on the deep learning train. Scala shows promise because of Spark, its killer app.
Reasonable language, well-known libraries.<p>There are other approaches in other languages that are gaining ground, but little offer the "easy in" that numpy/scipy/pandas provides.