def findDecision(df, config):
algorithm = config['algorithm']
decision_classes = df["Decision"].unique()
#-----------------------------
if algorithm == 'Regression':
stdev = df['Decision'].std(ddof=0)
entropy = 0
if algorithm == "ID3" or algorithm == "C4.5":
entropy = calculateEntropy(df, config)
#print("entropy: ",entropy)
columns = df.shape[1]; instances = df.shape[0]
gains = []; gainratios = []; ginis = []; reducted_stdevs = []; chi_squared_values = []
for i in range(0, columns-1):
column_name = df.columns[i]
column_type = df[column_name].dtypes
#print(column_name,"->",column_type)
if column_type != 'object':
df = Preprocess.processContinuousFeatures(algorithm, df, column_name, entropy, config)
classes = df[column_name].value_counts()
gain = entropy * 1; splitinfo = 0; gini = 0; weighted_stdev = 0; chi_squared_value = 0
for j in range(0, len(classes)):
current_class = classes.keys().tolist()[j]
#print(column_name,"->",current_class)
subdataset = df[df[column_name] == current_class]
#print(subdataset)
subset_instances = subdataset.shape[0]
class_probability = subset_instances/instances
if algorithm == 'ID3' or algorithm == 'C4.5':
subset_entropy = calculateEntropy(subdataset, config)
#print("entropy for this sub dataset is ", subset_entropy)
gain = gain - class_probability * subset_entropy
if algorithm == 'C4.5':
splitinfo = splitinfo - class_probability*math.log(class_probability, 2)
elif algorithm == 'CART': #GINI index
decision_list = subdataset['Decision'].value_counts().tolist()
subgini = 1
for k in range(0, len(decision_list)):
subgini = subgini - math.pow((decision_list[k]/subset_instances), 2)
gini = gini + (subset_instances / instances) * subgini
elif algorithm == 'CHAID':
num_of_decisions = len(decision_classes)
expected = subset_instances / num_of_decisions
for d in decision_classes:
num_of_d = subdataset[subdataset["Decision"] == d].shape[0]
chi_square_of_d = math.sqrt(((num_of_d - expected) * (num_of_d - expected)) / expected)
chi_squared_value += chi_square_of_d
elif algorithm == 'Regression':
subset_stdev = subdataset['Decision'].std(ddof=0)
weighted_stdev = weighted_stdev + (subset_instances/instances)*subset_stdev
#iterating over classes for loop end
#-------------------------------
if algorithm == "ID3":
gains.append(gain)
elif algorithm == "C4.5":
if splitinfo == 0:
splitinfo = 100 #this can be if data set consists of 2 rows and current column consists of 1 class. still decision can be made (decisions for these 2 rows same). set splitinfo to very large value to make gain ratio very small. in this way, we won't find this column as the most dominant one.
gainratio = gain / splitinfo
gainratios.append(gainratio)
elif algorithm == "CART":
ginis.append(gini)
elif algorithm == "CHAID":
chi_squared_values.append(chi_squared_value)
elif algorithm == 'Regression':
reducted_stdev = stdev - weighted_stdev
reducted_stdevs.append(reducted_stdev)
#print(df)
if algorithm == "ID3":
winner_index = gains.index(max(gains))
metric_value = entropy
metric_name = "Entropy"
elif algorithm == "C4.5":
winner_index = gainratios.index(max(gainratios))
metric_value = entropy
metric_name = "Entropy"
elif algorithm == "CART":
winner_index = ginis.index(min(ginis))
metric_value = min(ginis)
metric_name = "Gini"
elif algorithm == "CHAID":
winner_index = chi_squared_values.index(max(chi_squared_values))
metric_value = max(chi_squared_values)
metric_name = "ChiSquared"
elif algorithm == "Regression":
winner_index = reducted_stdevs.index(max(reducted_stdevs))
metric_value = max(reducted_stdevs)
metric_name = "Std"
winner_name = df.columns[winner_index]
return winner_name, df.shape[0], metric_value, metric_name
def findGains(df, config):
algorithm = config['algorithm']
decision_classes = df["Decision"].unique()
#-----------------------------
entropy = 0
if algorithm == "ID3" or algorithm == "C4.5":
entropy = calculateEntropy(df, config)
columns = df.shape[1]
instances = df.shape[0]
gains = []
for i in range(0, columns - 1):
column_name = df.columns[i]
column_type = df[column_name].dtypes
#print(column_name,"->",column_type)
if column_type != 'object':
df = Preprocess.processContinuousFeatures(algorithm, df,
column_name, entropy,
config)
classes = df[column_name].value_counts()
splitinfo = 0
if algorithm == 'ID3' or algorithm == 'C4.5':
gain = entropy * 1
else:
gain = 0
for j in range(0, len(classes)):
current_class = classes.keys().tolist()[j]
#print(column_name,"->",current_class)
subdataset = df[df[column_name] == current_class]
#print(subdataset)
subset_instances = subdataset.shape[0]
class_probability = subset_instances / instances
if algorithm == 'ID3' or algorithm == 'C4.5':
subset_entropy = calculateEntropy(subdataset, config)
gain = gain - class_probability * subset_entropy
if algorithm == 'C4.5':
splitinfo = splitinfo - class_probability * math.log(
class_probability, 2)
elif algorithm == 'CART': #GINI index
decision_list = subdataset['Decision'].value_counts().tolist()
subgini = 1
for k in range(0, len(decision_list)):
subgini = subgini - math.pow(
(decision_list[k] / subset_instances), 2)
gain = gain + (subset_instances / instances) * subgini
elif algorithm == 'CHAID':
num_of_decisions = len(decision_classes)
expected = subset_instances / num_of_decisions
for d in decision_classes:
num_of_d = subdataset[subdataset["Decision"] == d].shape[0]
chi_square_of_d = math.sqrt(
((num_of_d - expected) *
(num_of_d - expected)) / expected)
gain += chi_square_of_d
elif algorithm == 'Regression':
subset_stdev = subdataset['Decision'].std(ddof=0)
gain = gain + (subset_instances / instances) * subset_stdev
#iterating over classes for loop end
#-------------------------------
if algorithm == 'Regression':
stdev = df['Decision'].std(ddof=0)
gain = stdev - gain
if algorithm == 'C4.5':
if splitinfo == 0:
splitinfo = 100 #this can be if data set consists of 2 rows and current column consists of 1 class. still decision can be made (decisions for these 2 rows same). set splitinfo to very large value to make gain ratio very small. in this way, we won't find this column as the most dominant one.
gain = gain / splitinfo
#----------------------------------
gains.append(gain)
#-------------------------------------------------
resp_obj = {}
resp_obj["gains"] = {}
for idx, feature in enumerate(
df.columns[0:-1]): #Decision is always the last column
#print(idx, feature)
resp_obj["gains"][feature] = gains[idx]
resp_obj["entropy"] = entropy
return resp_obj