def test_explain_linear_feature_filter(newsgroups_train, vec):
clf = LogisticRegression(random_state=42)
docs, y, target_names = newsgroups_train
X = vec.fit_transform(docs)
clf.fit(X, y)
if isinstance(vec, HashingVectorizer):
vec = InvertableHashingVectorizer(vec)
vec.fit(docs)
res = explain_weights(clf, vec=vec, feature_re='^ath')
text_expl, _ = expls = format_as_all(res, clf)
for expl in expls:
assert 'atheists' in expl
assert 'atheism' in expl
assert 'space' not in expl
assert 'BIAS' not in expl
res = explain_weights(
clf,
vec=vec,
feature_filter=lambda name: name.startswith('ath') or name == '<BIAS>')
text_expl, _ = expls = format_as_all(res, clf)
for expl in expls:
assert 'atheists' in expl
assert 'atheism' in expl
assert 'space' not in expl
assert 'BIAS' in expl
assert '<BIAS>' in text_expl
def test_explain_hashing_vectorizer(newsgroups_train_binary):
# test that we can pass InvertableHashingVectorizer explicitly
vec = HashingVectorizer(n_features=1000)
ivec = InvertableHashingVectorizer(vec)
clf = LogisticRegression(random_state=42)
docs, y, target_names = newsgroups_train_binary
ivec.fit([docs[0]])
X = vec.fit_transform(docs)
clf.fit(X, y)
get_res = lambda **kwargs: explain_prediction(
clf, docs[0], vec=ivec, target_names=target_names, top=20, **kwargs)
res = get_res()
check_explain_linear_binary(res, clf)
assert res == get_res()
res_vectorized = explain_prediction(clf,
vec.transform([docs[0]])[0],
vec=ivec,
target_names=target_names,
top=20,
vectorized=True)
pprint(res_vectorized)
assert res_vectorized == _without_weighted_spans(res)
assert res == get_res(feature_names=ivec.get_feature_names(
always_signed=False))
def test_explain_linear_hashed(newsgroups_train, clf):
docs, y, target_names = newsgroups_train
vec = HashingVectorizer(n_features=10000)
ivec = InvertableHashingVectorizer(vec)
X = vec.fit_transform(docs)
clf.fit(X, y)
# use half of the docs to find common terms, to make it more realistic
ivec.fit(docs[::2])
check_newsgroups_explanation_linear(clf, ivec, target_names)
def test_explain_linear_hashed_pos_neg(newsgroups_train, pass_feature_weights):
docs, y, target_names = newsgroups_train
# make it binary
y = y.copy()
y[y != 0] = 1
target_names = [target_names[0], 'other']
vec = HashingVectorizer(norm=None)
ivec = InvertableHashingVectorizer(vec)
clf = LogisticRegression(random_state=42)
clf.fit(vec.fit_transform(docs), y)
ivec.fit(docs)
if pass_feature_weights:
res = explain_weights(
clf,
top=(10, 10),
target_names=target_names,
feature_names=ivec.get_feature_names(always_signed=False),
coef_scale=ivec.column_signs_)
else:
res = explain_weights(clf,
ivec,
top=(10, 10),
target_names=target_names)
# HashingVectorizer with norm=None is "the same" as CountVectorizer,
# so we can compare it and check that explanation is almost the same.
count_vec = CountVectorizer()
count_clf = LogisticRegression(random_state=42)
count_clf.fit(count_vec.fit_transform(docs), y)
count_res = explain_weights(count_clf,
vec=count_vec,
top=(10, 10),
target_names=target_names)
for key in ['pos', 'neg']:
values, count_values = [
sorted(get_names_coefs(getattr(r.targets[0].feature_weights, key)))
for r in [res, count_res]
]
assert len(values) == len(count_values)
for (name, coef), (count_name,
count_coef) in zip(values, count_values):
assert name == count_name
assert abs(coef - count_coef) < 0.05