Open
Description
Description
Performance is much worse when using partial_fit method on multilabel y than using fit on the same data. I suspect that the issue is partial_fit supports multi-class but not multi-label. Why is this the case when fit supports multi-label?
Steps/Code to Reproduce
X_train.shape, y_train.shape # --> ((3963, 4572), (3963, 39))
# where y is binary [0,1] for each of the 39 columns
mlp = MLPClassifier(hidden_layer_sizes=(500, ), activation='relu', solver='adam', alpha=0.0001, batch_size='auto', learning_rate='constant', learning_rate_init=0.001, power_t=0.5, max_iter=500, shuffle=True, random_state=123, tol=0.0001, early_stopping=False, validation_fraction=0.1, beta_1=0.9, beta_2=0.999, epsilon=1e-08, n_iter_no_change=10)
mlp.partial_fit(X_train, y_train, classes=list(range(y_train.shape[1])))
y_pred = mlp.predict(X_test)
precision_score(y_test, y_pred, average='weighted')Expected Results
precision score is 0.635 when using just fit.
Actual Results
precision score is 0.216 when using partial_fit method.
Versions
System
python: 3.6.5 (default, Apr 25 2018, 14:23:58) [GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.1)]
executable: /Users/clhq/.local/share/virtualenvs/LTR-looks-tag-recommender-QOgMS24J/bin/python
machine: Darwin-17.7.0-x86_64-i386-64bit
BLAS
macros: NO_ATLAS_INFO=3, HAVE_CBLAS=None
lib_dirs:
cblas_libs: cblas
Python deps
pip: 18.1
setuptools: 40.4.3
sklearn: 0.20.0
numpy: 1.15.3
scipy: 1.1.0
Cython: None
pandas: 0.23.4