import numpy as np

class NormalizePositive(object):

def fit(self, X, y=None):

# count features that are greater than zero in axis 0:

binary = (X > 0)

count0 = binary.sum(0)

# to avoid division by zero, set zero counts to one:

count0 += (count0 == 0)

self.mean = X.sum(0)/count0

# Compute variance by average squared difference to the mean, but only

# consider differences where binary is True (i.e., where there was a

# true rating):

diff = (X - self.mean) * binary

diff **= 2

self.std = np.sqrt(0.1 + diff.sum(0)/count0)

return self

def fit_transform(self, X):

return self.fit(X).transform(X)

def transform(self, X):

binary = (X > 0)

X = X - self.mean

X /= self.std

X *= binary

return X

def inverse_transform(self, X, copy=True):

if copy:

X = X.copy()

X *= self.std

X += self.mean

return X

def predict(train):

norm = NormalizePositive()

train = norm.fit_transform(train)

return norm.inverse_transform(train * 0.)

def main(transpose_inputs=False):

from load_ml100k import get_train_test

from sklearn import metrics

train,test = get_train_test(random_state=12)

if transpose_inputs:

train = train.T

test = test.T

predicted = predict(train)

r2 = metrics.r2_score(test[test > 0], predicted[test > 0])

print('R2 score ({} normalization): {:.1%}'.format(

('movie' if transpose_inputs else 'user'),

r2))

if __name__ == '__main__':

main()

main(transpose_inputs=True)