def train(self):
train, test = stratified_split(self.__data)
X_train = train.iloc[:, 0].values
y_train = train.iloc[:, 1].values
X_test = test.iloc[:, 0].values
y_test = test.iloc[:, 1].values
self.clf_ = self.__cross_valid(X_train, y_train)
logging.info('best clf:\n%s' % print_pipe_model(self.clf_))
# predict
#cross_val_predict(clf, X_train, y_train, cv=5)
y_train_pred = self.clf_.predict(X_train)
y_test_pred = self.clf_.predict(X_test)
logging.info('data shape: %s' % X_test.shape)
#confusion matrix
self.confusion_matrix_ = confusion_matrix(y_test, y_test_pred)
logging.info('confusion matrix:\n%s' % pformat(self.confusion_matrix_))
# wrong prediction
self.false_neg_ = test[(y_test == 1) & (y_test_pred == 0)]
self.false_pos_ = test[(y_test == 0) & (y_test_pred == 1)]
# evaluate
self.train_score_ = evaluate(y_train, y_train_pred)
self.test_score_ = evaluate(y_test, y_test_pred)
# most importance features
need_std = get_pipe_model(self.clf_, 'std') is None
self.influence_ = self._get_influence_ft(self.clf_, X_train) if need_std else \
self._get_influence_ft(self.clf_)
logging.info('Influence:\n%s\n...\n%s' %
(self.influence_.head(), self.influence_.tail()))
def _get_influence_ft(self, model, X=None):
tfidf = get_pipe_model(model, 'tfidf')
features = tfidf.get_feature_names()
influences = self._get_influences(model)
if len(features) != len(influences):
return pd.DataFrame(
{
'feature': features,
'influence': np.zeros(len(features))
},
columns=['feature', 'influence'])
if X is not None:
influences = self._calc_std_influences(influences,
tfidf.transform(X))
df = pd.DataFrame({
'feature': features,
'influence': influences
},
columns=['feature', 'influence'])
df.sort_values(by='influence', ascending=False, inplace=True)
return df
def get_decision_features(clf, x):
tfidf = get_pipe_model(clf, 'tfidf')
logre = get_pipe_model(clf, 'logre')
# weights of features
x_trans = get_pipe_transform(clf, x).toarray().reshape(-1)
coef = logre.coef_.reshape(-1)
weights = [ (i, v*w) for i, v, w in zip(range(len(x_trans)), x_trans, coef) if v and w ]
weights = sorted(weights, key=itemgetter(1), reverse=True)
# trim non-contributing weights
y = logre.intercept_[0]
for i, (_, w) in enumerate(weights):
if y > 0: # the smaple has became positive
break
y += w
weights = weights[:i]
# contributing features and weight ratio
features = tfidf.get_feature_names()
w_sum = sum([w for _, w in weights])
decision_fts = [ (w/w_sum, features[i] ) for i, w in weights ]
return decision_fts
def _get_influences(self, model):
return get_pipe_model(model, -1).feature_importances_
def _get_influences(self, model):
return get_pipe_model(model, -1).coef_[0]
