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)))
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 score import user_based_knn, pearson
from dataset import DataIO
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
# make python find our new modules
import sys
sys.path.append("../../../recsys")
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),
class DatasetTest(unittest.TestCase):
def setUp(self):
self.ratings_file = 'testdata/ratings.csv'
self.item_tags_file = 'testdata/movie-tags.csv'
self.ds = DataIO(False)
def test_ratings(self):
self.ds.load(self.ratings_file)
self.__ratings_norm_test()
self.__ratings_test()
self.__printer_test()
def test_item_tags(self):
self.ds.load(self.ratings_file, self.item_tags_file)
self.__ratings_norm_test()
self.__ratings_test()
self.__tags_test()
self.__tags_norm_test()
def __printer_test(self):
expected_users = '1: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n51: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n100: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)'
expected_items = '11: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n603: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n36955: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)'
recs = [
zip(range(10),
range(10)[::-1]),
] * 3
ids = [0, 50, 99]
self.assertTrue(
self.ds.print_recs(recs, given_items=ids) == expected_items)
self.assertTrue(
self.ds.print_recs(recs, given_users=ids) == expected_users)
def __ratings_test(self):
# lines count
self.assertTrue(
len(self.ds.ratings) == self.__wccount(self.ratings_file))
# values
head_ratings = [(1, 809, 4.0), (1, 601, 5.0), (1, 238, 5.0),
(1, 664, 4.5), (1, 3049, 3.0)]
self.assertTrue(self.ds.ratings[0:5] == [(self.ds.new_user_idx(u),
self.ds.new_item_idx(i), r)
for (u, i,
r) in head_ratings])
tail_ratings = [(5573, 114, 2.5), (5573, 22, 4.5), (5573, 11, 3.0),
(5573, 557, 4.0), (5573, 98, 3.5)]
self.assertTrue(self.ds.ratings[-5:] == [(self.ds.new_user_idx(u),
self.ds.new_item_idx(i), r)
for (u, i,
r) in tail_ratings])
def __ratings_norm_test(self):
(user_col, item_col) = zip(*self.ds.ratings)[:2]
self.assertTrue(len(set(user_col)) == self.ds.num_users())
self.assertTrue(len(set(item_col)) == self.ds.num_items())
self.assertTrue(
range(self.ds.num_users()) == [
self.ds.new_user_idx(self.ds.old_user_idx(i))
for i in range(self.ds.num_users())
])
self.assertTrue(
range(self.ds.num_items()) == [
self.ds.new_item_idx(self.ds.old_item_idx(i))
for i in range(self.ds.num_items())
])
def __tags_test(self):
# read tags file and check that all (item,tag) combinations appear in the dataset
# get item-tag combinations from the original file
file = open(self.item_tags_file, 'rbU')
csv_reader = csv.reader(file, delimiter=',')
item_tag_set_orig = set([(self.ds.new_item_idx(int(i)),
self.ds.tag_idx(t))
for (i, t) in csv_reader])
file.close()
# item-tag combinations in the dataset
item_tag_set = set(zip(*zip(*self.ds.item_tags)[:2]))
self.assertTrue(
len(item_tag_set_orig.symmetric_difference(item_tag_set)) == 0)
# tag values
tag_values = [(114, 'afternoon section', 1), (114, 'capitalism', 4),
(114, "YOUNG WOMEN'S FAVORATE", 1),
(10020, '18th century', 2), (581, 'wolves', 1)]
self.assertTrue(
all([
self.ds.item_tags.index(
(self.ds.new_item_idx(i), self.ds.tag_idx(t), c))
for (i, t, c) in tag_values
]))
# tag count
tag_count_expected = dict([(114, 1), (680, 1), (581, 1)])
# take list of unique (item,tag) pairs, replace tag with 1s and group-sum by the first argument
item_tagcount = dict(
self.__sum_group_by_first(
zip(
zip(*self.ds.item_tags)[0], [
1,
] * len(self.ds.item_tags))))
self.assertTrue([
item_tagcount[self.ds.new_item_idx(i)] == tag_count_expected[i]
for i in [114, 680, 581]
])
def __tags_norm_test(self):
# collect all users and items
(item_col, tag_col) = zip(*self.ds.item_tags)[:2]
# check that there are as many new indexes as different users and items
self.assertTrue(len(set(item_col)) == self.ds.num_items())
# actually, this may not hold, but let's keep for now
self.assertTrue(len(set(tag_col)) == self.ds.num_tags())
# for all tags, check that new(old(new) = new
self.assertTrue(
range(self.ds.num_tags()) == [
self.ds.tag_idx(self.ds.tags(i))
for i in range(self.ds.num_tags())
])
# takes a list of pairs
# group by the first element and do summ aggregate of the second
# credits http://stackoverflow.com/questions/11058001/python-group-by-and-sum-a-list-of-tuples
def __sum_group_by_first(self, list_of_pairs):
return [(x, sum([z[1] for z in y])) for (
x, y) in groupby(sorted(list_of_pairs, key=operator.itemgetter(0)),
key=operator.itemgetter(0))]
#credits https://gist.github.com/zed/0ac760859e614cd03652
def __wccount(self, filename):
out = subprocess.Popen(['wc', '-l', filename],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT).communicate()[0]
return int(out.strip().partition(b' ')[0])
def setUp(self):
self.ratings_file = 'testdata/ratings.csv'
self.item_tags_file = 'testdata/movie-tags.csv'
self.ds = DataIO(False)
class DatasetTest(unittest.TestCase):
def setUp(self):
self.ratings_file = 'testdata/ratings.csv'
self.item_tags_file = 'testdata/movie-tags.csv'
self.ds = DataIO(False)
def test_ratings(self):
self.ds.load(self.ratings_file)
self.__ratings_norm_test()
self.__ratings_test()
self.__printer_test()
def test_item_tags(self):
self.ds.load(self.ratings_file, self.item_tags_file)
self.__ratings_norm_test()
self.__ratings_test()
self.__tags_test()
self.__tags_norm_test()
def __printer_test(self):
expected_users = '1: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n51: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n100: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)'
expected_items = '11: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n603: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)\n36955: (11,9.00), (12,8.00), (13,7.00), (14,6.00), (22,5.00), (24,4.00), (38,3.00), (63,2.00), (77,1.00), (85,0.00)'
recs = [zip(range(10), range(10)[::-1]),]*3
ids = [0,50,99]
self.assertTrue(self.ds.print_recs(recs, given_items = ids) == expected_items)
self.assertTrue(self.ds.print_recs(recs, given_users = ids) == expected_users)
def __ratings_test(self):
# lines count
self.assertTrue(len(self.ds.ratings) == self.__wccount(self.ratings_file))
# values
head_ratings = [(1,809,4.0),(1,601,5.0),(1,238,5.0),(1,664,4.5),(1,3049,3.0)]
self.assertTrue(self.ds.ratings[0:5] == [(self.ds.new_user_idx(u),self.ds.new_item_idx(i),r) for (u,i,r) in head_ratings])
tail_ratings = [(5573,114,2.5),(5573,22,4.5),(5573,11,3.0),(5573,557,4.0),(5573,98,3.5)]
self.assertTrue(self.ds.ratings[-5:] == [(self.ds.new_user_idx(u),self.ds.new_item_idx(i),r) for (u,i,r) in tail_ratings])
def __ratings_norm_test(self):
(user_col, item_col) = zip(*self.ds.ratings)[:2]
self.assertTrue(len(set(user_col)) == self.ds.num_users())
self.assertTrue(len(set(item_col)) == self.ds.num_items())
self.assertTrue(range(self.ds.num_users()) ==
[self.ds.new_user_idx(self.ds.old_user_idx(i)) for i in range(self.ds.num_users())])
self.assertTrue(range(self.ds.num_items()) ==
[self.ds.new_item_idx(self.ds.old_item_idx(i)) for i in range(self.ds.num_items())])
def __tags_test(self):
# read tags file and check that all (item,tag) combinations appear in the dataset
# get item-tag combinations from the original file
file = open(self.item_tags_file, 'rbU')
csv_reader = csv.reader(file, delimiter=',')
item_tag_set_orig = set([(self.ds.new_item_idx(int(i)), self.ds.tag_idx(t)) for (i,t) in csv_reader])
file.close()
# item-tag combinations in the dataset
item_tag_set = set(zip(*zip(*self.ds.item_tags)[:2]))
self.assertTrue(len(item_tag_set_orig.symmetric_difference(item_tag_set)) == 0)
# tag values
tag_values = [(114,'afternoon section',1),
(114,'capitalism',4),
(114,"YOUNG WOMEN'S FAVORATE",1),
(10020,'18th century',2),
(581,'wolves',1)]
self.assertTrue(all([self.ds.item_tags.index((self.ds.new_item_idx(i), self.ds.tag_idx(t), c )) for (i,t,c) in tag_values]))
# tag count
tag_count_expected = dict([(114,1),
(680,1),
(581,1)])
# take list of unique (item,tag) pairs, replace tag with 1s and group-sum by the first argument
item_tagcount = dict(self.__sum_group_by_first( zip(zip(*self.ds.item_tags)[0], [1,]*len(self.ds.item_tags)) ))
self.assertTrue([item_tagcount[self.ds.new_item_idx(i)] == tag_count_expected[i] for i in [114,680,581]])
def __tags_norm_test(self):
# collect all users and items
(item_col, tag_col) = zip(*self.ds.item_tags)[:2]
# check that there are as many new indexes as different users and items
self.assertTrue(len(set(item_col)) == self.ds.num_items())
# actually, this may not hold, but let's keep for now
self.assertTrue(len(set(tag_col)) == self.ds.num_tags())
# for all tags, check that new(old(new) = new
self.assertTrue(range(self.ds.num_tags()) ==
[self.ds.tag_idx(self.ds.tags(i)) for i in range(self.ds.num_tags())])
# takes a list of pairs
# group by the first element and do summ aggregate of the second
# credits http://stackoverflow.com/questions/11058001/python-group-by-and-sum-a-list-of-tuples
def __sum_group_by_first(self, list_of_pairs):
return [(x,sum([z[1] for z in y])) for (x,y)
in groupby(sorted(list_of_pairs, key = operator.itemgetter(0)),
key = operator.itemgetter(0))]
#credits https://gist.github.com/zed/0ac760859e614cd03652
def __wccount(self, filename):
out = subprocess.Popen(['wc', '-l', filename],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT
).communicate()[0]
return int(out.strip().partition(b' ')[0])
def setUp(self):
self.ratings_file = 'testdata/ratings.csv'
self.item_tags_file = 'testdata/movie-tags.csv'
self.ds = DataIO(False)
# make python find our new modules
import sys
sys.path.append("../../../recsys")
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,
from score import user_based_knn, pearson
from dataset import DataIO
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)
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)
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]))
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)
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]))