def owen_zhang(x_train, y_train, x_test, seed=0, alpha=0, noise=0.01):
"""
Owen Zhang's leave-one-out + noise likelihood encoding
"Winning data science competitions"
http://de.slideshare.net/ShangxuanZhang/winning-data-science-competitions-presented-by-owen-zhang
"""
if len(x_train.shape) == 1:
x_train = x_train.reshape(-1, 1)
x_test = x_test.reshape(-1, 1)
ncols = x_train.shape[1]
nclass = np.unique(y_train).shape[0]
if not is_numpy(x_train):
x_train = np.array(x_train)
x_test = np.array(x_test)
xx_train = None
xx_test = None
for i in range(ncols):
le_train = LikelihoodEstimator(noise=noise, alpha=alpha, leave_one_out=True, seed=seed). \
fit(x_train[:, i], y_train)
le_test = LikelihoodEstimator(noise=0, alpha=alpha, leave_one_out=False, seed=seed). \
fit(x_train[:, i], y_train)
lh_train = le_train.x_likelihoods.copy()
lh_test = le_test.predict_proba(x_test[:, i])
if nclass <= 2:
lh_train = lh_train.T[1].reshape(-1, 1)
lh_test = lh_test.T[1].reshape(-1, 1)
xx_train = np.hstack((lh_train,)) if xx_train is None else np.hstack((xx_train, lh_train))
xx_test = np.hstack((lh_test,)) if xx_test is None else np.hstack((xx_test, lh_test))
return xx_train, xx_test
def fit(self, x):
self.ecdfs = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
self.ecdfs.update({i: ECDF(x[:, i] if is_np else x.iloc[:, i].values)})
return self
def fit(self, x):
self.ecdfs = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
self.ecdfs.update(
{i: ECDF(x[:, i] if is_np else x.iloc[:, i].values)})
return self
def fit(self, x, y):
print x.shape
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
if not is_numpy(x):
x = np.array(x)
self.nclass = np.unique(y).shape[0]
for i in range(ncols):
self.estimators.append(LikelihoodEstimator(**self.get_params()).fit(x[:, i], y))
return self
def fit(self, x, y):
print x.shape
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
if not is_numpy(x):
x = np.array(x)
self.nclass = np.unique(y).shape[0]
for i in range(ncols):
self.estimators.append(
LikelihoodEstimator(**self.get_params()).fit(x[:, i], y))
return self
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
ecdf = self.ecdfs[i]
if is_np:
x[:, i] = self.ppf(ecdf(x[:, i]))
else:
x.iloc[:, i] = self.ppf(ecdf(x.iloc[:, i]))
return x
def transform(self, x):
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
if not is_numpy(x):
x = np.array(x)
likelihoods = None
for i in range(ncols):
lh = self.estimators[i].predict(x[:, i])
if self.nclass <= 2:
lh = lh.T[1].reshape(-1, 1)
likelihoods = np.hstack((lh,)) if likelihoods is None else np.hstack((likelihoods, lh))
return likelihoods
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
ecdf = self.ecdfs[i]
if is_np:
x[:, i] = self.ppf(ecdf(x[:, i]))
else:
x.iloc[:, i] = self.ppf(ecdf(x.iloc[:, i]))
return x
def fit(self, x):
self.new_values = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
if is_np:
val = dict(Counter(x[:, i]))
else:
val = x.iloc[:, i].value_counts().to_dict()
val = dict((k, self.value if v < self.threshold else k) for k, v in val.items())
self.new_values.update({i: val})
return self
def fit(self, x):
self.new_values = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
if is_np:
val = dict(Counter(x[:, i]))
else:
val = x.iloc[:, i].value_counts().to_dict()
val = dict((k, self.value if v < self.threshold else k)
for k, v in val.items())
self.new_values.update({i: val})
return self
def fit(self, x):
self.counts = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
if is_np:
cnt = dict(Counter(x[:, i]))
else:
cnt = x.iloc[:, i].value_counts().to_dict()
if self.min_count > 0:
cnt = dict((k, self.nan_value if v < self.min_count else v) for k, v in cnt.items())
self.counts.update({i: cnt})
return self
def transform(self, x):
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
if not is_numpy(x):
x = np.array(x)
likelihoods = None
for i in range(ncols):
lh = self.estimators[i].predict(x[:, i])
if self.nclass <= 2:
lh = lh.T[1].reshape(-1, 1)
likelihoods = np.hstack(
(lh, )) if likelihoods is None else np.hstack(
(likelihoods, lh))
return likelihoods
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
cnt = self.counts[i]
if is_np:
k, v = np.array(list(zip(*sorted(cnt.items()))))
ix = np.digitize(x[:, i], k, right=True)
x[:, i] = v[ix]
else:
x.iloc[:, i].replace(cnt, inplace=True)
return x
def fit(self, x):
self.counts = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
if is_np:
cnt = dict(Counter(x[:, i]))
else:
cnt = x.iloc[:, i].value_counts().to_dict()
if self.min_count > 0:
cnt = dict((k, self.nan_value if v < self.min_count else v)
for k, v in cnt.items())
self.counts.update({i: cnt})
return self
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
for i in range(ncols):
cnt = self.counts[i]
if is_np:
k, v = np.array(list(zip(*sorted(cnt.items()))))
ix = np.digitize(x[:, i], k, right=True)
x[:, i] = v[ix]
else:
x.iloc[:, i].replace(cnt, inplace=True)
return x
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
if self.ive is not None:
x = self.ive.transform(x)
for i in range(ncols):
enc = self.encoders[i]
if is_np:
x[:, i] = enc.transform(x[:, i]) + self.first_category
else:
x.iloc[:, i] = enc.transform(x.iloc[:, i]) + self.first_category
return x
def fit(self, x):
self.encoders = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
if self.min_count > 0:
self.ive = InfrequentValueEncoder(threshold=self.min_count, value=np.finfo(float).min)
x = self.ive.fit_transform(x)
for i in range(ncols):
if is_np:
enc = LabelEncoder().fit(x[:, i])
else:
enc = LabelEncoder().fit(x.iloc[:, i])
self.encoders.update({i: enc})
return self
def fit(self, x):
self.encoders = {}
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
if self.min_count > 0:
self.ive = InfrequentValueEncoder(threshold=self.min_count,
value=np.finfo(float).min)
x = self.ive.fit_transform(x)
for i in range(ncols):
if is_np:
enc = LabelEncoder().fit(x[:, i])
else:
enc = LabelEncoder().fit(x.iloc[:, i])
self.encoders.update({i: enc})
return self
def transform(self, x):
if self.copy:
x = x.copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1)
ncols = x.shape[1]
is_np = is_numpy(x)
if self.ive is not None:
x = self.ive.transform(x)
for i in range(ncols):
enc = self.encoders[i]
if is_np:
x[:, i] = enc.transform(x[:, i]) + self.first_category
else:
x.iloc[:,
i] = enc.transform(x.iloc[:, i]) + self.first_category
return x
def owen_zhang(x_train, y_train, x_test, seed=0, alpha=0, noise=0.01):
"""
Owen Zhang's leave-one-out + noise likelihood encoding
"Winning data science competitions"
http://de.slideshare.net/ShangxuanZhang/winning-data-science-competitions-presented-by-owen-zhang
"""
if len(x_train.shape) == 1:
x_train = x_train.reshape(-1, 1)
x_test = x_test.reshape(-1, 1)
ncols = x_train.shape[1]
nclass = np.unique(y_train).shape[0]
if not is_numpy(x_train):
x_train = np.array(x_train)
x_test = np.array(x_test)
xx_train = None
xx_test = None
for i in range(ncols):
le_train = LikelihoodEstimator(noise=noise, alpha=alpha, leave_one_out=True, seed=seed). \
fit(x_train[:, i], y_train)
le_test = LikelihoodEstimator(noise=0, alpha=alpha, leave_one_out=False, seed=seed). \
fit(x_train[:, i], y_train)
lh_train = le_train.x_likelihoods.copy()
lh_test = le_test.predict_proba(x_test[:, i])
if nclass <= 2:
lh_train = lh_train.T[1].reshape(-1, 1)
lh_test = lh_test.T[1].reshape(-1, 1)
xx_train = np.hstack(
(lh_train, )) if xx_train is None else np.hstack(
(xx_train, lh_train))
xx_test = np.hstack((lh_test, )) if xx_test is None else np.hstack(
(xx_test, lh_test))
return xx_train, xx_test
