def test__set_dataframe(self):
empty_frame = pd.DataFrame()
kf = KFoldPartitioning(n_splits=2)
with self.assertRaises(PartitionError):
kf.set_dataframe(empty_frame)
def test_iter(self):
kf = KFoldPartitioning()
kf.set_dataframe(original_frame)
for train, test in kf:
original_list = [
list(row) for row in original_frame.itertuples(index=False)
]
train_list = [list(row) for row in train.itertuples(index=False)]
test_list = [list(row) for row in test.itertuples(index=False)]
# Check that train and test are a partition
train_not_in_test = [
row for row in train_list if row not in test_list
]
self.assertCountEqual(
train_list, train_not_in_test) # Count so regardless of order
test_not_in_train = [
row for row in test_list if row not in train_list
]
self.assertCountEqual(
test_list, test_not_in_train) # Count so regardless of order
# Check that the union of the two give the original data
union_list = train_list + test_list
self.assertCountEqual(original_list,
union_list) # Count so regardless of order
def test_fit_graph_w_testrating_methodology(self):
graph = NXFullGraph(ratings)
rs = GraphBasedRS(NXPageRank(), graph)
em = EvalModel(rs, KFoldPartitioning(), metric_list=[Precision()])
sys_result, users_result = em.fit()
self.assertIsInstance(sys_result, pd.DataFrame)
self.assertIsInstance(users_result, pd.DataFrame)
def test_split__single_kfold(self):
user_ratings = pd.DataFrame.from_dict(
{'from_id': ["001", "001", "001", "001"],
'to_id': ["iphone", "ps4", "ps5", "xbox"],
'rating': [0.8, 0.7, -0.4, 1.0]})
n_split = 2
pm = PartitionModule(KFoldPartitioning(n_split))
user_splits = pm._split_single(user_ratings)
# No further tests since the partitioning technique is tested singularly
self.assertEqual(len(user_splits), n_split)
def test_split_all_kfold(self):
all_ratings = pd.DataFrame(
{'from_id': ["001", "001", "001", "001", "002", "002", "002", "003", "003"],
'to_id': ["iphone", "ps4", "ps5", "xbox", "realme", "airpods", "ps4", "beats", "dvd"],
'rating': [0.8, 0.7, -0.4, 1.0, 0.8, 0.7, -0.4, 1.0, 0.65]})
n_split = 2
pm = PartitionModule(KFoldPartitioning(n_split))
split_list = pm.split_all(all_ratings, set(all_ratings.from_id))
# No further tests since the partitioning technique is tested singularly
self.assertEqual(len(split_list), n_split)
def test_fit_cb_w_testrating_methodology(self):
rs = ContentBasedRS(
CentroidVector(
{"Plot": "tfidf"},
CosineSimilarity(),
), ratings, items_dir)
em = EvalModel(rs, KFoldPartitioning(), metric_list=[Precision()])
sys_result, users_result = em.fit()
self.assertIsInstance(sys_result, pd.DataFrame)
self.assertIsInstance(users_result, pd.DataFrame)
def test_all(self):
ratings_filename = os.path.join(contents_path, '..', 'datasets',
'examples', 'new_ratings.csv')
ratings_frame = RatingsImporter(
CSVFile(ratings_filename)).import_ratings()
rs = ContentBasedRS(
LinearPredictor(
{"Plot": ['tfidf', 'embedding']},
SkLinearRegression(),
), ratings_frame, items_dir)
catalog = set([
os.path.splitext(f)[0] for f in os.listdir(items_dir)
if os.path.isfile(os.path.join(items_dir, f)) and f.endswith('xz')
])
em = EvalModel(rs,
KFoldPartitioning(),
metric_list=[
Precision(sys_average='micro'),
PrecisionAtK(1, sys_average='micro'),
RPrecision(),
Recall(),
RecallAtK(3, ),
FMeasure(1, sys_average='macro'),
FMeasureAtK(2, beta=1, sys_average='micro'),
NDCG(),
NDCGAtK(3),
MRR(),
MRRAtK(5, ),
Correlation('pearson', top_n=5),
Correlation('kendall', top_n=3),
Correlation('spearman', top_n=4),
MAE(),
MSE(),
RMSE(),
CatalogCoverage(catalog),
CatalogCoverage(catalog, k=2),
CatalogCoverage(catalog, top_n=3),
GiniIndex(),
GiniIndex(top_n=3),
DeltaGap({
'primo': 0.5,
'secondo': 0.5
})
],
methodology=TestItemsMethodology())
result = em.fit()
def test_graph(self):
catalog = set(ratings.to_id)
users_dir = os.path.join(dir_test_files, 'complex_contents',
'users_codified/')
graph = NXFullGraph(
ratings,
user_contents_dir=users_dir,
item_contents_dir=items_dir,
item_exo_representation="dbpedia",
user_exo_representation='local',
item_exo_properties=['starring'],
user_exo_properties=['1'
] # It's the column in the users .DAT which
# identifies the gender
)
graph_rs = GraphBasedRS(NXPageRank(), graph)
em = EvalModel(graph_rs,
KFoldPartitioning(),
metric_list=[
Precision(relevant_threshold=3),
Recall(),
FMeasure(beta=1),
FMeasure(beta=2, sys_average='micro'),
MRR(),
Correlation('pearson'),
GiniIndex(),
DeltaGap({
'popular': 0.5,
'niche': 0.5
}),
PredictionCoverage(catalog),
PopProfileVsRecs(user_groups={
'popular': 0.5,
'niche': 0.5
},
out_dir='plots/'),
LongTailDistr('plots/', format='svg'),
PopRecsCorrelation('plots/')
],
verbose_predictions=True,
methodology=TestItemsMethodology())
em.fit()
def test_fit_graph_w_allitems_methodology(self):
graph = NXFullGraph(ratings)
rs = GraphBasedRS(NXPageRank(), graph)
items = set([
os.path.splitext(f)[0] for f in os.listdir(items_dir)
if os.path.isfile(os.path.join(items_dir, f)) and f.endswith('xz')
])
em = EvalModel(rs,
KFoldPartitioning(),
metric_list=[Precision()],
methodology=AllItemsMethodology(items))
sys_result, users_result = em.fit()
self.assertIsInstance(sys_result, pd.DataFrame)
self.assertIsInstance(users_result, pd.DataFrame)
def test_fit_cb_w_allitems_methodology(self):
rs = ContentBasedRS(
CentroidVector(
{"Plot": "tfidf"},
CosineSimilarity(),
), ratings, items_dir)
items = set([
os.path.splitext(f)[0] for f in os.listdir(items_dir)
if os.path.isfile(os.path.join(items_dir, f)) and f.endswith('xz')
])
em = EvalModel(rs,
KFoldPartitioning(),
metric_list=[Precision()],
methodology=AllItemsMethodology(items))
sys_result, users_result = em.fit()
self.assertIsInstance(sys_result, pd.DataFrame)
self.assertIsInstance(users_result, pd.DataFrame)
def test_all_skipping_user_exception(self):
all_ratings = pd.DataFrame(
{'from_id': ["001", "001", "001", "001", "002", "002", "002", "003", "004", "004"],
'to_id': ["iphone", "ps4", "ps5", "xbox", "realme", "airpods", "ps4", "beats", "ps4", "ps5"],
'rating': [0.8, 0.7, -0.4, 1.0, 0.8, 0.7, -0.4, 1.0, 0.3, 0.6]})
n_split = 2
pm = PartitionModule(KFoldPartitioning(n_split))
split_list = pm.split_all(all_ratings, set(all_ratings.from_id))
# No further tests since the partitioning technique is tested singularly
self.assertEqual(len(split_list), n_split)
# Check that there are all users except 003 which is skipped since it has only 1 rating
for split in split_list:
self.assertIn('001', split.train['from_id'].values)
self.assertIn('001', split.test['from_id'].values)
self.assertIn('002', split.train['from_id'].values)
self.assertIn('002', split.test['from_id'].values)
self.assertNotIn('003', split.train['from_id'].values)
self.assertNotIn('003', split.test['from_id'].values)
self.assertIn('004', split.train['from_id'].values)
self.assertIn('004', split.test['from_id'].values)