The test_classifier_chain_vs_independent_models test in sklearn/tests/test_multioutput.py currently uses fetch_mldata('yeast'). Perhaps we should rely on synthetic data rather than requiring our testing to involve an unreliable web download.