def test_pa3(self):
testdata = zip([(1024, 77), (1024, 268), (1024, 462), (1024, 393),
(1024, 36955), (2048, 77), (2048, 36955), (2048, 788)],
[
"1024,77,4.3848,Memento (2000)",
"1024,268,2.8646,Batman (1989)",
"1024,462,3.1082,Erin Brockovich (2000)",
"1024,393,3.8722,Kill Bill: Vol. 2 (2004)",
"1024,36955,2.3524,True Lies (1994)",
"2048,77,4.8493,Memento (2000)",
"2048,36955,3.9698,True Lies (1994)",
"2048,788,3.8509,Mrs. Doubtfire (1993)",
])
data = DataIO(verbose=False)
data.load('testdata/ratings.csv',
items_file='testdata/movie-titles.csv')
model = UserModel(verbose=False, normalize=True)
model.build(data)
for ((u, i), s) in testdata:
self.assertTrue(
'%s' % s == '%d,%d,%.4f,%s' %
(u, i,
user_based_knn(model,
30, [data.new_user_idx(u)],
[data.new_item_idx(i)],
cosine,
promote_users=True,
normalize='centered'), data.title(i)))
class TestUserModel(unittest.TestCase):
def setUp(self):
self.data = DummyDataset()
self.model = UserModel(verbose = False)
self.model.build(self.data)
def test_mean(self):
expected = np.matrix([[ 3.2 ],
[ 3.94 ],
[ 4.13333333]])
self.assertTrue(stringify_matrix(self.model.mean()) == stringify_matrix(expected))
def test_users(self):
expected = sparse.csr_matrix(
[[ 0.8 , -2.2 , 0.0 , 1.4 , 0.0 ],
[ 0.56 , 0.06 , -0.44 , -0.24 , 0.06 ],
[ 0.86666667, 0.0 , -0.73333333, 0.0 , -0.13333333]])
self.assertTrue(stringify_matrix(self.model.R().todense()) == stringify_matrix(expected.todense()))
def test_pa3(self):
testdata = zip([(1024,77),(1024,268),(1024,462),(1024,393),(1024,36955),(2048,77),(2048,36955),(2048,788)],
[
"1024,77,4.3848,Memento (2000)",
"1024,268,2.8646,Batman (1989)",
"1024,462,3.1082,Erin Brockovich (2000)",
"1024,393,3.8722,Kill Bill: Vol. 2 (2004)",
"1024,36955,2.3524,True Lies (1994)",
"2048,77,4.8493,Memento (2000)",
"2048,36955,3.9698,True Lies (1994)",
"2048,788,3.8509,Mrs. Doubtfire (1993)",
])
data = DataIO(verbose = False)
data.load('testdata/ratings.csv', items_file = 'testdata/movie-titles.csv')
model = UserModel(verbose = False, normalize = True)
model.build(data)
for ((u,i),s) in testdata:
self.assertTrue('%s' % s ==
'%d,%d,%.4f,%s' % (u,i,user_based_knn(model, 30, [data.new_user_idx(u)],[data.new_item_idx(i)],
cosine, promote_users = True, normalize = 'centered'), data.title(i)))
from model import UserModel
from suggest import top_ns
ratings_file = "ratings.csv"
given_users = [3867, 860]
NN = 5
n = 3
part_1_file = "part_1.csv"
part_2_file = "part_2.csv"
# part 1
data = DataIO()
data.load(ratings_file)
model = UserModel(normalize=False)
model.build(data)
given_users = data.translate_users(given_users)
given_items = range(data.num_items())
R = user_based_knn(model, NN, given_users, given_items, pearson, promote_users=False)
recs = top_ns(R, n, keep_order=True)
file = open(part_1_file, "w")
file.write("\n".join(["%d %.3f" % (data.old_item_idx(i), s) for u in recs for (i, s) in u]))
file.close()
# part 2
R = user_based_knn(
model, NN, given_users, given_items, pearson, promote_users=False, exclude_seen=False, normalize=True
from score import user_based_knn, cosine
from dataset import DataIO
from model import UserModel
ratings_file = '../data/ratings.csv'
items_file = '../data/movie-titles.csv'
NN = 30
answer_file = 'part_1.csv'
# part 1
data = DataIO()
data.load(ratings_file, items_file=items_file)
model = UserModel(normalize=True)
model.build(data)
inputs = [(4169, 161), (4169, 36955), (4169, 453), (4169, 857), (4169, 238),
(5399, 1891), (5399, 14), (5399, 187), (5399, 602), (5399, 629),
(3613, 329), (3613, 604), (3613, 134), (3613, 1637), (3613, 278),
(1873, 786), (1873, 2502), (1873, 550), (1873, 1894), (1873, 1422),
(4914, 268), (4914, 36658), (4914, 786), (4914, 161), (4914, 854)]
file = open(answer_file, 'w')
file.write('\n'.join([
'%d,%d,%.4f,%s' %
(u, i,
user_based_knn(model,
NN, [data.new_user_idx(u)], [data.new_item_idx(i)],
cosine,
promote_users=True,
class WA4Test(unittest.TestCase):
def setUp(self):
self.data = DataIO(verbose = False)
self.data.load('testdata/ratings-ma4.csv')
self.model = UserModel(normalize = False, verbose = False)
self.model.build(self.data)
def test_pearson(self):
# test correlation
S = pearson(self.model.R(), self.model.R()).todense()
# 1. check we don't have numbers more than 1
# user string comparison to avoid float nuances
self.assertTrue('%.2f' % S.max() == '1.00');
# 2. check there are only '1' on the diagonal
self.assertTrue(sum([S[i,i] for i in range(S.shape[0])]) == S.shape[0])
# 3. check a couple of correlation coefficients
corr_test = [(1648, 5136, 0.40298),
(918, 2824, -0.31706)]
for (u1,u2,c) in corr_test:
# check what's in the full matrix
u1 = self.data.new_user_idx(u1)
u2 = self.data.new_user_idx(u2)
# check precomputed
self.assertTrue('%.5f' % S[u1,u2] == '%.5f' % c)
# compute here
self.assertTrue('%.5f' % pearson(self.model.R()[u1,:], self.model.R()[u2,:]).todense() == '%.5f' % c)
def test_5nn(self):
u = 3712
nns = [(2824,0.46291), (3867,0.400275), (5062,0.247693), (442,0.22713), (3853,0.19366)]
S = pearson(self.model.R(), self.model.R())
leave_top_n(S,6)
top_neighbours = [(self.data.old_user_idx(i),S[i,self.data.new_user_idx(u)])
for i in S[:,self.data.new_user_idx(u)].nonzero()[0]]
top_neighbours.sort(key = lambda a: a[1], reverse = True)
# skip the first element (corr = 1)
self.assertTrue(','.join(['%d,%.6f' % a for a in top_neighbours[1:]]) ==
','.join(['%d,%.6f' % a for a in nns]))
# consider moving this test to test_recsys.py
def test_unnormalized(self):
u = 3712
expected = [(641,5.000), (603,4.856), (105,4.739)]
R = user_based_knn(self.model, 5, [self.data.new_user_idx(u)], range(self.data.num_items()),
pearson, promote_users = False)
recs = top_ns([R],3, keep_order = True)
self.assertTrue(','.join(['%d,%.3f' % (self.data.old_item_idx(a),b) for (a,b) in recs[0]]) ==
','.join(['%d,%.3f' % a for a in expected]))
# consider moving this test to test_recsys.py
def test_normalized(self):
u = 3712
expected = [(641,5.900), (603,5.546), (105,5.501)]
R = user_based_knn(self.model, 5, [self.data.new_user_idx(u)], range(self.data.num_items()),
pearson, promote_users = False, normalize = 'normalize')
recs = top_ns([R],3, keep_order = True)
self.assertTrue(','.join(['%d,%.3f' % (self.data.old_item_idx(a),b) for (a,b) in recs[0]]) ==
','.join(['%d,%.3f' % a for a in expected]))
class WA4Test(unittest.TestCase):
def setUp(self):
self.data = DataIO(verbose=False)
self.data.load('testdata/ratings-ma4.csv')
self.model = UserModel(normalize=False, verbose=False)
self.model.build(self.data)
def test_pearson(self):
# test correlation
S = pearson(self.model.R(), self.model.R()).todense()
# 1. check we don't have numbers more than 1
# user string comparison to avoid float nuances
self.assertTrue('%.2f' % S.max() == '1.00')
# 2. check there are only '1' on the diagonal
self.assertTrue(
sum([S[i, i] for i in range(S.shape[0])]) == S.shape[0])
# 3. check a couple of correlation coefficients
corr_test = [(1648, 5136, 0.40298), (918, 2824, -0.31706)]
for (u1, u2, c) in corr_test:
# check what's in the full matrix
u1 = self.data.new_user_idx(u1)
u2 = self.data.new_user_idx(u2)
# check precomputed
self.assertTrue('%.5f' % S[u1, u2] == '%.5f' % c)
# compute here
self.assertTrue(
'%.5f' % pearson(self.model.R()[u1, :],
self.model.R()[u2, :]).todense() == '%.5f' %
c)
def test_5nn(self):
u = 3712
nns = [(2824, 0.46291), (3867, 0.400275), (5062, 0.247693),
(442, 0.22713), (3853, 0.19366)]
S = pearson(self.model.R(), self.model.R())
leave_top_n(S, 6)
top_neighbours = [
(self.data.old_user_idx(i), S[i, self.data.new_user_idx(u)])
for i in S[:, self.data.new_user_idx(u)].nonzero()[0]
]
top_neighbours.sort(key=lambda a: a[1], reverse=True)
# skip the first element (corr = 1)
self.assertTrue(','.join(['%d,%.6f' % a for a in top_neighbours[1:]])
== ','.join(['%d,%.6f' % a for a in nns]))
# consider moving this test to test_recsys.py
def test_unnormalized(self):
u = 3712
expected = [(641, 5.000), (603, 4.856), (105, 4.739)]
R = user_based_knn(self.model,
5, [self.data.new_user_idx(u)],
range(self.data.num_items()),
pearson,
promote_users=False)
recs = top_ns([R], 3, keep_order=True)
self.assertTrue(','.join(
['%d,%.3f' % (self.data.old_item_idx(a), b) for (
a,
b) in recs[0]]) == ','.join(['%d,%.3f' % a for a in expected]))
# consider moving this test to test_recsys.py
def test_normalized(self):
u = 3712
expected = [(641, 5.900), (603, 5.546), (105, 5.501)]
R = user_based_knn(self.model,
5, [self.data.new_user_idx(u)],
range(self.data.num_items()),
pearson,
promote_users=False,
normalize='normalize')
recs = top_ns([R], 3, keep_order=True)
self.assertTrue(','.join(
['%d,%.3f' % (self.data.old_item_idx(a), b) for (
a,
b) in recs[0]]) == ','.join(['%d,%.3f' % a for a in expected]))
