If you look carefully at your stacktrace this looks like a scipy problem, you should be able to reproduce it just using from scipy.sparse.linalg import lsqr as sparse_lsqr.

Try uninstalling + reinstalling scipy.

Yes, as I mentioned reinstalling scipy with numpy 1.10.1 solves this. The issue is, why is this so and (because this seems to be an issue for scipy) should there be a hint to use sklearn with numpy 1.10.1 (or maybe i just haven't found it)?

Sorry, I am going to try to be clearer:

• the problem you are encountering seems, with a probability very close to one, to be a scipy install problem on your machine.
• scipy and for that matter scikit-learn works fine with numpy 1.10.2 to the best of my knowledge.

The advice I was trying to give you was to try uninstalling + reinstalling scipy while keeping numpy 1.10.2, for example if you use conda (if you use pip replace conda by pip in the instructions):

conda uninstall scipy
conda install scipy

Still the same problem with numpy 1.10.2 .
A different user on a different machine and os (i use arch) had the same problem:

http://askubuntu.com/questions/712485/installing-scikit-learn-with-python3-superlu-cpython-34m-so-undefined-symbol

Still the same problem with numpy 1.10.2 .
A different user on a different machine and os (i use arch) had the same problem:

Fair enough, it may be a genuine problem, hard to tell. I didn't manage to reproduce this in a fresh conda environment using pip install scipy. In any case this seems like a scipy specific problem and the scikit-learn bug tracker may not be the best place to discuss it.

For the record I am using conda and I don't seem to be affected by this issue.

I get the same issue with numpy 1.10.4 + scipy 0.16.1

Downgrading numpy to 1.10.1 & reinstalling scipy solves this:

$ sudo pip install numpy==1.10.1
$ sudo pip install --upgrade --force-reinstall scipy
$ uname -a
Linux 4.3.3-2-ARCH #1 SMP PREEMPT Wed Dec 23 20:09:18 CET 2015 x86_64 GNU/Linux

To me, that looks very much like a binary incompatibility problem.

In other terms, I believe that you are installing various packages from
different sources that have not been compiled to work together. This is
why Sachin Garg has solved this problem by forcing recompilations.

So are you compiling numpy and scipy from source? That is not really recommended (also per our documentation). I'd recommend using your package manager or conda instead.

As the error is in scipy and not scikit-learn, I'm closing.

As the error is in scipy and not scikit-learn, I'm closing.

I agree on this one.

FWIW I tried to reproduce the problem on an ArchLinux distro within VirtualBox since both @AndreRoehrig and @sgargbugreporter seemed to have the problem on Arch. I wasn't able to reproduce the problem.

btw try from scipy.sparse.linalg import lsqr as sparse_lsqr to see that it is a scipy issue.

I had to also add --no-cache-dir when re-installing scipy

I experienced the same issue and I made a report to scipy as it's definitely not the right place here.