def test_perform_greater_k(self):
metric = RecallAtK(k=999, relevant_threshold=3)
result = metric.perform(split_w_new_items)
expected = Recall(relevant_threshold=3).perform(split_w_new_items)
result = np.sort(result, axis=0)
expected = np.sort(expected, axis=0)
# If k > than number of rows, then it's a regular Precision
self.assertTrue(np.array_equal(expected, result))
def test_perform_i1_i4_missing(self):
metric = RecallAtK(k=3, relevant_threshold=3)
result_macro = metric.perform(split_missing)
expected_u1 = 0 / 2
result_macro_u1 = float(result_macro.query('from_id == "u1"')[str(metric)])
self.assertAlmostEqual(expected_u1, result_macro_u1)
expected_u2 = 2 / 3
result_macro_u2 = float(result_macro.query('from_id == "u2"')[str(metric)])
self.assertAlmostEqual(expected_u2, result_macro_u2)
expected_macro_sys = (expected_u1 + expected_u2) / 2
result_macro_sys = float(result_macro.query('from_id == "sys"')[str(metric)])
self.assertAlmostEqual(expected_macro_sys, result_macro_sys)
def test_perform_mean(self):
metric = RecallAtK(k=3)
result = metric.perform(split_w_new_items)
expected_u1 = 1/2
result_mean_u1 = float(result.query('from_id == "u1"')[str(metric)])
self.assertAlmostEqual(expected_u1, result_mean_u1)
expected_u2 = 1/1
result_mean_u2 = float(result.query('from_id == "u2"')[str(metric)])
self.assertAlmostEqual(expected_u2, result_mean_u2)
expected_mean_sys = (expected_u1 + expected_u2) / 2
result_mean_sys = float(result.query('from_id == "sys"')[str(metric)])
self.assertAlmostEqual(expected_mean_sys, result_mean_sys)
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_eval_ranking_needed_metrics_implicit_split(self):
# We set the split_list directly by the class attribute
c = MetricCalculator()
RankingNeededMetric.rank_truth_list = self.rank_split_list
system_res, each_user_res = c.eval_metrics([
Precision(),
PrecisionAtK(2),
RPrecision(),
Recall(),
RecallAtK(2),
FMeasure(),
FMeasureAtK(2),
NDCG(),
NDCGAtK(2),
MRR(),
MRRAtK(2),
Correlation('pearson'),
Correlation('kendall'),
Correlation('spearman'),
PredictionCoverage(self.catalog),
CatalogCoverage(self.catalog, top_n=2),
GiniIndex(),
DeltaGap(user_groups={
'a': 0.5,
'b': 0.5
}),
LongTailDistr(out_dir='test_plot'),
PopProfileVsRecs(user_groups={
'a': 0.5,
'b': 0.5
},
out_dir='test_plot'),
PopRecsCorrelation(out_dir='test_plot')
])
self.assertIsInstance(system_res, pd.DataFrame)
self.assertIsInstance(each_user_res, pd.DataFrame)