def main():
ks = [3, 5, 10, 20]
mapk = 200
train, test = load_data()
train, test = train.as_matrix(), test.as_matrix()
x = train.T
res = np.zeros(9)
number_of_user = train.shape[0]
print int(number_of_user * 0.01)
pca = PCA(n_components=int(number_of_user * 0.01))
new_x = pca.fit_transform(x)
for u in xrange(train.shape[0]):
y = x[:, u]
truth = test[u]
clf = LogisticRegression(random_state=42, C=0.001, solver='lbfgs')
clf.fit(new_x, y)
#print u, classification.accuracy_score(clf.predict(x), y)
pred_buy_proba = clf.predict_proba(new_x)[:, 1].ravel()
pruned_buy_proba = pred_buy_proba - y.ravel()
pred_order = pruned_buy_proba.argsort()[::-1]
actual_bought = truth.nonzero()[0]
score = apk(actual_bought, pred_order, mapk)
tmp = [score]
for k in ks:
tmp.append(prec(actual_bought, pred_order, k))
tmp.append(recall(actual_bought, pred_order, k))
res += np.array(tmp)
if u % 50 == 0:
print res / (u + 1)
return res / (u + 1)
def main():
ks = [3, 5, 10, 20]
mapk = 200
train, test = load_data()
train, test = train.as_matrix(), test.as_matrix()
pred = train.sum(axis=0)
res = np.zeros(9)
x = train.T
for u in xrange(train.shape[0]):
y = x[:, u]
truth = test[u]
pred_buy_proba = pred
y[y > 0] = float('-inf')
pruned_buy_proba = pred_buy_proba + y.ravel()
pred_order = pruned_buy_proba.argsort()[::-1]
actual_bought = truth.nonzero()[0]
score = apk(actual_bought, pred_order, mapk)
tmp = [score]
for k in ks:
tmp.append(prec(actual_bought, pred_order, k))
tmp.append(recall(actual_bought, pred_order, k))
res += np.array(tmp)
if u % 50 == 0:
print res / (u + 1)
return res / (u + 1)
def main():
ks = [3, 5, 10, 20]
mapk = 200
train, test = load_data()
train, test = train.as_matrix(), test.as_matrix()
x = train
pca = PCA(n_components=int(train.shape[1] * 0.01))
pca.fit(train)
new_x = pca.transform(train)
res = []
for i in xrange(train.shape[1]):
y = x[:, i]
clf = LogisticRegression(random_state=42, C=0.001, solver='lbfgs')
clf.fit(new_x, y)
pred_buy_proba = clf.predict_proba(new_x)[:, 1].ravel()
res.append(pred_buy_proba)
res = np.array(res).T
pred = (res - train).argsort(axis=1)[::-1]
res = np.zeros(9)
for u in xrange(train.shape[0]):
truth = test[u]
pred_order = pred[u]
actual_bought = truth.nonzero()[0]
score = apk(actual_bought, pred_order, mapk)
tmp = [score]
for k in ks:
tmp.append(prec(actual_bought, pred_order, k))
tmp.append(recall(actual_bought, pred_order, k))
res += np.array(tmp)
if u % 50 == 0:
print res / (u + 1)
return res / (u + 1)
def main():
ks = [3, 5, 10, 20]
mapk = 200
epoch = 1
res = np.zeros(9)
train, test = load_data()
train, test = train.as_matrix(), test.as_matrix()
x = train.T
clf = SGDClassifier(random_state=42, loss='log')
for i in xrange(epoch):
for u in xrange(train.shape[0]):
y = x[:, u]
new_x = np.append(x,
np.repeat(y[:, np.newaxis], x.shape[1], axis=1),
axis=1)
clf.partial_fit(new_x, y, classes=[0, 1])
for u in xrange(train.shape[0]):
y = x[:, u]
truth = test[u]
new_x = np.append(x,
np.repeat(y[:, np.newaxis], x.shape[1], axis=1),
axis=1)
clf.predict(new_x)
pred_buy_proba = clf.predict_proba(new_x)[:, 1].ravel()
pruned_buy_proba = pred_buy_proba - y.ravel()
pred_order = pruned_buy_proba.argsort()[::-1]
actual_bought = truth.nonzero()[0]
score = apk(actual_bought, pred_order, mapk)
tmp = [score]
for k in ks:
tmp.append(prec(actual_bought, pred_order, k))
tmp.append(recall(actual_bought, pred_order, k))
res += np.array(tmp)
if u % 50 == 0:
print res / (u + 1)
return res / (u + 1)