def __init__(self, dataFrame, criterion, depth, split_feature=''):
self.df = dataFrame
self.depth = depth
self.split_feature = split_feature
self.impurity = round(
impurity_func(dataFrame.values[:, -1], criterion), 6)
self.child = {}
self.is_leaf = False
self.label = None
def Find_Best_Feature(self, df):
"""
This function retuns the feature that can split the data.
in which, the impurity is the least.
you should implement the part of calculating for
impurity (entropy or gini_index) for given feature.변경하다
이 함수는 불순도가 가장 작게 데이터를 분할할 수 있는 feature를
반환하는 함수입니다.
여러분이 구현할 내용은 주어진 feature의 impurity(entropy or gini_index)를
계산하는 것입니다.
[Parameter]:
df : [dataFrame (got by pandas) ][N x D] : Training_data
[Variables] :
header : [list of string ] [1 x D] the set of attribute_name, last element is for output_feature.
input_feature : [list of string] [1 x (D-1)] the set of attribute name except for last feature(output)
Y_data : [column vector of label] [N x 1] label_data, To get the (vector or matrix) not dataFrame,
you can write ' data = df.values '
self.Category_feature_idx : [list of integers]: the set of idx only for Category feature.(referred top of the code.)
split_value : in numeric_data, You just divide the data two part.
one is the less than split_value, the other is no less than split_value.
[Objects]: (the part of implement)
impurity : (float) entropy or gini_index got by splitting data given attribute.
impurity_list [list of float] [1 x D] : the list which store the all impurity in order.
[return]:
Best_feature : (string) feature_name
feature_type : (string) the type of feature ('Category' or 'Numeric')
"""
header = df.columns.values
input_feature = header[:-1]
output_feature = header[-1]
Y_data = df[output_feature].values
impurity_list = []
# for all features in DataFrame,
for idx, h in enumerate(input_feature):
# ============ Edit here ==================
col_data = df[h]
Y = df.values[:, -1]
distinct_data = np.unique(col_data)
# Category Feature Case
if idx in self.Category_feature_idx:
impurity = 0
for attr in distinct_data:
attr_idx = (col_data == attr)
y0 = Y[attr_idx]
p0 = len(y0) / len(Y)
impurity += p0 * impurity_func(y0, self.criterion)
# Numeric Feature Case
else:
split_value = Finding_split_point(df, h, self.criterion)
for val in distinct_data:
less_idx = (col_data < split_value)
y0 = Y[less_idx]
y1 = Y[~less_idx]
p0 = len(y0) / len(Y)
p1 = len(y1) / len(Y)
impurity = np.sum([
p0 * impurity_func(y0, self.criterion),
p1 * impurity_func(y1, self.criterion)
])
#=====================================================
impurity_list.append(np.round(impurity, 6))
idx = np.argmin(impurity_list)
Best_feature = input_feature[idx]
feature_type = idx in self.Category_feature_idx and 'Category' or 'Numeric'
return Best_feature, feature_type
def Find_Best_Feature(self, df):
"""
This function retuns the feature that can split the data.
in which, the impurity is the least.
you should implement the part of calculating for
impurity (entropy or gini_index) for given feature.변경하다
이 함수는 불순도가 가장 작게 데이터를 분할할 수 있는 feature를
반환하는 함수입니다.
여러분이 구현할 내용은 주어진 feature의 impurity(entropy or gini_index)를
계산하는 것입니다.
[Parameter]:
df : [dataFrame (got by pandas) ][N x D] : Training_data
[Variables] :
header : [list of string ] [1 x D] the set of attribute_name, last element is for output_feature.
input_feature : [list of string] [1 x (D-1)] the set of attribute name except for last feature(output)
Y_data : [column vector of label] [N x 1] label_data, To get the (vector or matrix) not dataFrame,
you can write ' data = df.values '
self.Category_feature_idx : [list of integers]: the set of idx only for Category feature.(referred top of the code.)
split_value : in numeric_data, You just divide the data two part.
one is the less than split_value, the other is no less than split_value.
[Objects]: (the part of implement)
impurity : (float) entropy or gini_index got by splitting data given attribute.
impurity_list [list of float] [1 x D] : the list which store the all impurity in order.
[return]:
Best_feature : (string) feature_name
feature_type : (string) the type of feature ('Category' or 'Numeric')
"""
header = df.columns.values
input_feature = header[:-1]
output_feature = header[-1]
Y_data = df[output_feature].values
impurity_list = []
# for all features in DataFrame,
for idx, h in enumerate(input_feature):
# ============ Edit here ==================
impurity = 0 # 이 위에 포문에서 피쳐(컬럼) 한줄 씩 가져옴. Humidty, temp, outlook 이런거야
# Category Feature Case
if idx in self.Category_feature_idx: # 이건 카테고리 데이터일때, 예를 들어 humidity high랑 normal 뿐이니까
values = []
total_value_len = len(Y_data) # 총 데이터의 개수
for value in df[h].values:
if value not in values:
values.append(value)
# 위에 3줄은 이제 그 특정 피쳐에서 값을 뽑아내는건데 df[h]에 이제 high, high, normal, high 이런거 들어있고
# df[h]가 판다스라 어케 뽑아내는지 몰라서 저렇게함. values에 이제 high, normal 이렇게 담기게 되는 거
for value in values:
data = Y_data[
df[h].values ==
value] # 이제 values에서 하나 하나 뽑아서 Y_data 값을 가져오는거 예를들어, df[h] = (h, h, n ,h) Y_data = (yes, no, yes, yes) 이면 value가 h 일때 data 는 (yes, no, yes)가 되는거지
impurity += (len(data) / total_value_len) * impurity_func(
data, self.criterion
) # impurity 구하는거는 이제 그 따로따로 임퓨리티 구해서 전체에서의 비율 곱하기
# 위 예제에서는 (3/4)*I(2, 1) + (1/4)*I(1,0) 이렇게 되는거
# Numeric Feature Case
else: # 컨티뉴 일때는
split_value = Finding_split_point(
df, h, self.criterion) # 이건 조교가 짜준코드 가장 적절한 스플릿 포인트를 찾아
less_idx = (df[h] < split_value) # 스플릿 포인트보다 작은거의 인덱스
y0 = Y_data[less_idx] # 스플릿 포인트보다 작은 놈들의 Y_data
y1 = Y_data[~less_idx] # 이건 반대
p0 = len(y0) / len(Y_data)
p1 = len(y1) / len(Y_data)
impurity = np.sum([
p0 * impurity_func(y0, self.criterion),
p1 * impurity_func(y1, self.criterion)
]) # 두개의 impurity합이 이제 impurity가 됨/
#=====================================================
impurity_list.append(np.round(impurity, 6))
idx = np.argmin(impurity_list)
Best_feature = input_feature[idx]
feature_type = idx in self.Category_feature_idx and 'Category' or 'Numeric'
return Best_feature, feature_type
def Find_Best_Feature(self, df):
"""
This function retuns the feature that can split the data.
in which, the impurity is the least.
you should implement the part of calculating for
impurity (entropy or gini_index) for given feature.변경하다
이 함수는 불순도가 가장 작게 데이터를 분할할 수 있는 feature를
반환하는 함수입니다.
여러분이 구현할 내용은 주어진 feature의 impurity(entropy or gini_index)를
계산하는 것입니다.
[Parameter]:
df : [dataFrame (got by pandas) ][N x D] : Training_data
[Variables] :
header : [list of string ] [1 x D] the set of attribute_name, last element is for output_feature.
input_feature : [list of string] [1 x (D-1)] the set of attribute name except for last feature(output)
Y_data : [column vector of label] [N x 1] label_data, To get the (vector or matrix) not dataFrame,
you can write ' data = df.values '
self.Category_feature_idx : [list of integers]: the set of idx only for Category feature.(referred top of the code.)
split_value : in numeric_data, You just divide the data two part.
one is the less than split_value, the other is no less than split_value.
[Objects]: (the part of implement)
impurity : (float) entropy or gini_index got by splitting data given attribute.
impurity_list [list of float] [1 x D] : the list which store the all impurity in order.
[return]:
Best_feature : (string) feature_name
feature_type : (string) the type of feature ('Category' or 'Numeric')
"""
header = df.columns.values
input_feature = header[:-1]
output_feature = header[-1]
Y_data = df[output_feature].values
impurity_list = []
# for all features in DataFrame,
for idx, h in enumerate(input_feature):
# ============ Edit here ==================
num = df[h].values.shape[0]
impurity = 0
if idx in self.Category_feature_idx:
for key, cnt in zip(
*np.unique(df[h].values, return_counts=True)):
impurity += impurity_func(Y_data[df[h].values == key],
self.criterion) * cnt / num
else:
split_value = Finding_split_point(df, h, self.criterion)
split = df[h].values < split_value
num_T, num_F = split.sum(), (~split).sum()
impurity += impurity_func(Y_data[split],
self.criterion) * num_T / num
impurity += impurity_func(Y_data[~split],
self.criterion) * num_F / num
#=====================================================
impurity_list.append(np.round(impurity, 6))
idx = np.argmin(impurity_list)
Best_feature = input_feature[idx]
feature_type = idx in self.Category_feature_idx and 'Category' or 'Numeric'
return Best_feature, feature_type