def main():
np.random.seed(0)
data = f.readData()
train, validation, test = f.splitData(data.shape[0])
C = [0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1]
# C = 0.061 F-1 = 0.525
trainRegularizationStrengthForl2(data, train, validation, C)
# C = 0.175 F-1 = 0.526170798898
trainRegularizationStrengthForl1(data, train, validation, C)
def main():
np.random.seed(0)
data = f.readData()
train, validation, test = f.splitData(data.shape[0])
# trainNeuralNetworks(data, train, validation)
alphas = np.arange(
0.0001, 0.0015,
0.0001) #[0.00001, 0.00005, 0.0001, 0.0005, 0.001, 0.005, 0.01]
N = [75, 100]
F_1 = np.zeros([len(alphas), len(N)])
for i in range(len(alphas)):
for j in range(len(N)):
F_1[i, j] = trainNeuralNetworks(data,
train,
validation,
N=N[j],
alpha=alphas[i])
def main():
np.random.seed(0)
data = f.readData()
train, validation, test = f.splitData(data.shape[0])
# C = [0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1]
# n_estimators = [5000, 10000, 50000, 100000, 500000]
# n_estimators = np.arange(10, 200, 10)
# n_estimators = np.repeat([100],100)
# number_of_trees = 100, average F-1 on 100 forests = 0.377228139802
# trainRandomForest(data, train, validation, n_estimators, max_features = None)
n_estimators = [100, 200, 500, 1000]
# number_of_boosting_stages = 100, average F-1 on 100 boostings = 0.377228139802
trainGradientBoosting(data,
train,
validation,
n_estimators,
max_features='auto')
def main():
np.random.seed(0)
data = f.readData()
train, validation, test = f.splitData(data.shape[0])
# C = [0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1]
C = [0.000001, 0.000005, 0.00001, 0.00005, 0.0001, 0.0005]
# C = [10, 50, 100, 500, 1000]
# C = np.arange(1, 10, 1)
# C = 137, F-1 = 0.541310541311
trainSVMWithGaussianKernel(data, train, validation, C)
trainSVMWithLinearKernel(data, train, validation, C)
# SVM with Linear Kernel
# l1, squared hinge, C = 50, F-1 = 0.525447042641
# l2, hinge, C = 0.001 , F-1 = 0.512968299712
# l2, squared hinge, C = 1, F-1 = 0.524725274725
trainSVMWithLinearKernel2(data, train, validation, C)
if fName == 'abalone.data':
abalonePath = os.path.join(path, fName)
abaloneNames = [
'sex', 'length', 'diameter', 'height', 'wholeHt', 'shuckWt',
'visceraWt', 'shellWt', 'rings'
]
data, features, classNum = prepData(abalonePath, abaloneNames, slice(-1),
'rings')
classes = classNum.astype(str)
classes[classNum <= 8] = '1-8'
classes[classNum >= 11] = '11+'
else:
sys.exit("No such data set.")
(xvData, xvLabel), xvFolds, pruningSet = splitData(data, classes)
fullErr, prnErr = crossValidate(xvData, xvLabel, pruningSet, xvFolds,
printTree)
print("Full Tree - Mean Error: %f" % fullErr.mean())
print("Full Tree - St Dev Error: %f" % fullErr.std())
print("Pruned Tree - Mean Error: %f" % prnErr.mean())
print("Pruned Tree - St Dev Error: %f" % prnErr.std())
tr = TrainDTree(xvData, xvLabel) # train DTree using full cross-val sample
tr.combineChildNodes() # combine subtrees with homogeneous classes
if printTree: # print full tree
print("\n===full tree===")
print(tr)
PruneDTree(tr, pruningSet[0], pruningSet[1]) # prune with pruning set
def main():
np.random.seed(0)
data = f.readData()
train, validation, test = f.splitData(data.shape[0])
X_train, y_train = v.makeMatrix(data, train)
X_test, y_test = v.makeMatrix(data, test)
print("Logistic Regression")
clf = LogisticRegression(C=0.061,
class_weight='balanced',
max_iter=10000,
solver='sag',
n_jobs=-1)
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print("F-1 measure for Logistic Regression with l2 and C = %s is %s" %
(0.061, f1))
clf = LogisticRegression(penalty='l1',
C=0.175,
class_weight='balanced',
max_iter=5000,
n_jobs=-1)
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print("F-1 measure for Logistic Regression with l1 and C = %s is %s" %
(0.175, f1))
print("SVM")
clf = SVC(C=137, class_weight='balanced')
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print("F-1 measure for SVM with RBF and C = %s is %s" % (137, f1))
# l1, squared hinge, C = 50, F-1 = 0.525447042641
# l2, hinge, C = 0.001 , F-1 = 0.512968299712
# l2, squared hinge, C = 1, F-1 = 0.524725274725
C = 50
loss = "squared_hinge"
penalty = 'l1'
clf = LinearSVC(C=C,
loss=loss,
penalty=penalty,
class_weight='balanced',
dual=False)
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print(
"F-1 measure for SVM with Linear Kernal, %s loss, %s penalty, and C = %s is %s"
% (loss, penalty, C, f1))
C = 0.001
loss = "hinge"
penalty = 'l2'
clf = LinearSVC(C=C,
loss=loss,
penalty=penalty,
class_weight='balanced',
dual=True)
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print(
"F-1 measure for SVM with Linear Kernal, %s loss, %s penalty, and C = %s is %s"
% (loss, penalty, C, f1))
C = 1
loss = "squared_hinge"
penalty = 'l2'
clf = LinearSVC(C=C,
loss=loss,
penalty=penalty,
class_weight='balanced',
dual=False)
f1 = v.validate(data, X_train, y_train, X_test, y_test, clf)
print(
"F-1 measure for SVM with Linear Kernal, %s loss, %s penalty, and C = %s is %s"
% (loss, penalty, C, f1))
TODO: Kann ein feature welches immer 0 ist hohe gewichte im lin fit kriegen? -> Ja! sehr einfach sogar
Ändert die Ridge regression das? -> Ja! w[i] ist dann 0
Dazu entweder einfach ein feature = 0 setzen und lernen oder ein gelerntes modell nehmen und feature X runter und gewicht zu X sehr hoch setzen - sollte die
performance wenig belasten wenn das feature X ca. 0 ist.
"""
### data
data = F.loadPreparedData()
### norm each feature to 1 (makes the weights comparable - not necessary with shap, but nice to have)
data = data / data.max(axis=0)
## append sqrt(fare) to better the distribution
data["sqrt(fare)"] = np.sqrt(data["fare"])
## append 0 data colum
data["testZero"] = rnd.random(len(data)) * 0.00001
## test train split
x_train, y_train, x_test, y_test = F.splitData(data)
### correlation
corr = data.corr()["survived"]
F.prettyPrint("Correlation",
corr.sort_values(key=lambda x: abs(x), ascending=False))
DO_LIN_FIT = 0
DO_NN_FIT = 1
DO_NNLIN_FIT = 0
if DO_LIN_FIT:
print("-----------------------------------------")
print(" ######## LIN MODEL ########")
print("-----------------------------------------")
### Linear Model
linModel, w = F.linRegression(x_train, y_train, x_test, y_test)