def fit(self, X, y):
"""
Fit the tree classifier
Arguments:
X {[df]} -- Dataframe containing the features
y {pd.series} -- Label vector
"""
X = assert_df(X)
y = assert_series(y)
self.attributes = [
a for a in self.domain.attributes if a.name in X.columns.values
]
self.columns = [a.name for a in self.attributes]
s_domain = Domain(self.attributes, class_vars=self.domain.class_var)
rows = pd.concat([X[self.columns], y], axis=1).values.tolist()
train = Table.from_list(domain=s_domain, rows=rows)
if isinstance(self.domain.class_var, DiscreteVariable):
self.tree = TreeClassifier().fit_storage(train)
else:
self.tree = TreeRegressor().fit_storage(train)
return self
def __init__(self, f_types, l_type, shape, **kwargs):
self.f_types = assert_series(f_types)
self.l_type = assert_l_type(l_type)
self.shape = shape
self._init_parameters(**kwargs)
self.is_fitted = False
self.names = self.f_types.index.tolist()
self.feature_importances = {name: -1 for name in self.names}
def __init__(self, f_types, l_type, **kwargs):
"""
Class which predicts label for unseen samples
Arguments:
f_types {pd.series} -- Series of feature types
l_type {str} -- Type of label
"""
self.f_types = assert_series(f_types)
self.l_type = assert_l_type(l_type)
self.params = {
"knn_neighbors": kwargs.get("knn_neighbors", 6),
"nominal_distance": kwargs.get("nominal_distance", 1),
"distance_metric": kwargs.get("distance_metric", "partial"),
}
def fit(self, X, y):
if self.is_fitted:
print("Selector is already fitted")
return self
X = assert_df(X).reset_index(drop=True)
y = assert_series(y).reset_index(drop=True)
data = Data(X, y, self.f_types, self.l_type, X.shape)
self.data = data.shuffle_rows()
self.domain = None
if self.params["eval_method"] == "tree":
self.domain = self.data.to_table().domain
if self.params["eval_method"] == "mi":
self.data = self.data.add_salt()
self._fit()
self.is_fitted = True
return self
def _get_mi_cc(X, y, f_types, l_type, k, dist):
"""
Estimate mutual information for continous label types
and at least one continous feature
Checks how many samples are inside a given radius
Arguments:
"""
nx = np.ones(X.shape[0]) * -1
ny = np.ones(X.shape[0]) * -1
D_x = get_dist_matrix(X, f_types, nominal_distance=dist)
D_x.sort()
new_y = assert_df(y)
new_types = assert_series(l_type)
D_y = get_dist_matrix(new_y, new_types, nominal_distance=dist)
D_y.sort()
for row in range(X.shape[0]):
# Get distances inside features and labels
dist_x = D_x[row, :]
dist_y = D_y[row, :]
# Update statistics if sample contains non-nan values
radius = max(dist_x[k + 1], dist_y[k + 1])
if not np.isinf(radius):
nx[row] = (dist_x <= radius).sum() - 1
ny[row] = (dist_y <= radius).sum() - 1
nx = nx[nx >= 0]
ny = ny[nx >= 0]
mi = digamma(len(nx)) + digamma(k) - (1 / k) - \
digamma(np.mean(nx)) - digamma(np.mean(ny))
return max(mi, 0)