SVM test_spars gives random results with Python 3.5

[alpo]

in test_spars list(xes.indices)) AssertionError: Lists differ: [0, 1, 2, 3] != [1, 3, 0, 2]

It can be related to different dictionary keys order in Python 2.7 (scrambled) and Python 3.5 (randomized). In Python 3.6 (ordered) the result can differ. It is hidden in internals of scipy.sparse.csr_martix perhaps. Investigation required.