def test_novelty(self):
recommender = baseline.MostPopularRecommender(training_set, items)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertAlmostEqual((math.log2(0.375) + math.log2(0.5) + math.log2(0.125)) / -3,
evaluator.novelty(recommender))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertAlmostEqual((math.log2(0.375) + math.log2(0.5)) / -2, evaluator.novelty(recommender))
def test_serendipity(self):
recommender = baseline.MostPopularRecommender(training_set, items)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertEqual(0.0, evaluator.serendipity(recommender))
self.assertAlmostEqual(1 / 6, evaluator.serendipity(recommender, primitive_k=2))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertEqual(0.0, evaluator.serendipity(recommender))
self.assertAlmostEqual(3 / 6, evaluator.serendipity(recommender, primitive_k=1))
def test_diversity(self):
recommender = baseline.MostPopularRecommender(training_set, items)
similarity = sequeval.CosineSimilarity(training_set, items)
# Compute the possible diversities
d1 = 1 - similarity.similarity(2, 1)
d2 = 1 - similarity.similarity(2, 3)
d3 = 1 - similarity.similarity(1, 3)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertAlmostEqual((d1 + d2 + d3) / 3, evaluator.diversity(recommender, similarity))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertAlmostEqual(d1, evaluator.diversity(recommender, similarity))
def run():
_user_ratings = int(request.args.get('user-ratings'))
_item_ratings = int(request.args.get('item-ratings'))
_splitter = request.args.get('splitter')
_ratio = float(request.args.get('ratio')) / 100
_k = int(request.args.get('length'))
loader = sequeval.UIRTLoader(user_ratings=_user_ratings,
item_ratings=_item_ratings)
ratings = loader.load('datasets/yes_reduced.csv')
builder = sequeval.Builder('1000 s')
sequences, items = builder.build(ratings)
profiler = sequeval.Profiler(sequences)
response = {
'profiler': {
'users': profiler.users(),
'items': profiler.items(),
'ratings': profiler.ratings(),
'sequences': profiler.sequences(),
'sparsity': parse(profiler.sparsity()),
'length': parse(profiler.sequence_length())
}
}
if _splitter == 'random':
splitter = sequeval.RandomSplitter(_ratio)
else:
splitter = sequeval.TimestampSplitter(_ratio)
training_set, test_set = splitter.split(sequences)
response['splitter'] = {
'training': len(training_set),
'test': len(test_set)
}
evaluator = sequeval.Evaluator(training_set, test_set, items, _k)
cosine = sequeval.CosineSimilarity(training_set, items)
response['evaluator'] = []
most_popular = baseline.MostPopularRecommender(training_set, items)
response['evaluator'].append(evaluation(evaluator, most_popular, cosine))
random = baseline.RandomRecommender(training_set, items)
response['evaluator'].append(evaluation(evaluator, random, cosine))
unigram = baseline.UnigramRecommender(training_set, items)
response['evaluator'].append(evaluation(evaluator, unigram, cosine))
bigram = baseline.BigramRecommender(training_set, items)
response['evaluator'].append(evaluation(evaluator, bigram, cosine))
return json.dumps(response)
def test_perplexity(self):
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
recommender = baseline.MostPopularRecommender(training_set, items)
self.assertAlmostEqual(math.inf, evaluator.perplexity(recommender))
recommender = baseline.RandomRecommender(training_set, items)
self.assertAlmostEqual(3.0, evaluator.perplexity(recommender))
def test_confidence(self):
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
recommender = baseline.MostPopularRecommender(training_set, items)
self.assertEqual(1.0, evaluator.confidence(recommender))
recommender = baseline.RandomRecommender(training_set, items)
self.assertAlmostEqual(1 / 3, evaluator.confidence(recommender))
def test_ndpm(self):
recommender = baseline.MostPopularRecommender(training_set, items)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertAlmostEqual(5 / 12, evaluator.ndpm(recommender))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertEqual(3 / 4, evaluator.ndpm(recommender))
def test_precision(self):
recommender = baseline.MostPopularRecommender(training_set, items)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertAlmostEqual(5 / 6, evaluator.precision(recommender))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertEqual(1.0, evaluator.precision(recommender))
def test_coverage(self):
recommender = baseline.MostPopularRecommender(training_set, items)
evaluator = sequeval.Evaluator(training_set, test_set, items, 3)
self.assertEqual(3 / 3, evaluator.coverage(recommender))
evaluator = sequeval.Evaluator(training_set, test_set, items, 2)
self.assertEqual(2 / 3, evaluator.coverage(recommender))
if args.splitter == 'random':
print("\n# Random splitter")
splitter = sequeval.RandomSplitter(args.ratio)
elif args.splitter == 'timestamp':
print("\n# Timestamp splitter")
splitter = sequeval.TimestampSplitter(args.ratio)
else:
raise RuntimeError('Unknown splitter ' + args.splitter)
training_set, test_set = splitter.split(sequences)
print("Training set:", len(training_set))
print("Test set:", len(test_set))
print("\n# Evaluator")
print("%10s\t%10s\t%10s\t%10s\t%10s\t%10s\t%10s\t%10s\t%10s" %
("Algorithm", "Coverage", "Precision", "nDPM", "Diversity",
"Novelty", "Serendipity", "Confidence", "Perplexity"))
evaluator = sequeval.Evaluator(training_set, test_set, items, args.length)
cosine = sequeval.CosineSimilarity(training_set, items)
most_popular = baseline.MostPopularRecommender(training_set, items)
evaluation(evaluator, most_popular, cosine)
random = baseline.RandomRecommender(training_set, items)
evaluation(evaluator, random, cosine)
unigram = baseline.UnigramRecommender(training_set, items)
evaluation(evaluator, unigram, cosine)
bigram = baseline.BigramRecommender(training_set, items)
evaluation(evaluator, bigram, cosine)