taar/tests/test_similarityrecommender.py

339 строки
11 KiB
Python

# This Source Code Form is subject to the terms of the Mozilla Public
# License, v. 2.0. If a copy of the MPL was not distributed with this file,
# You can obtain one at http://mozilla.org/MPL/2.0/.
import json
import six
import pytest
import numpy as np
import scipy.stats
from taar.cache import JSONCache, Clock
from taar.recommenders.similarity_recommender import \
CATEGORICAL_FEATURES, CONTINUOUS_FEATURES, DONOR_LIST_KEY, LR_CURVES_SIMILARITY_TO_PROBABILITY, \
SimilarityRecommender
from .similarity_data import CONTINUOUS_FEATURE_FIXTURE_DATA
from .similarity_data import CATEGORICAL_FEATURE_FIXTURE_DATA
def generate_fake_lr_curves(num_elements, ceiling=10.0):
"""
Generate a mock likelihood ratio (LR) curve that can be used for
testing.
"""
lr_index = list(np.linspace(0, ceiling, num_elements))
# This sets up a normal distribution with a mean of 0.5 and std
# deviation of 0.5
numerator_density = [scipy.stats.norm.pdf(float(i), 0.5, 0.5) for i in lr_index]
# This sets up a normal distribution with a mean of 5.0 and std
# deviation of 1.5. So this is right shifted and has a wide std
# deviation compared to the first curve.
# This results in a small overlap. Using ranges between 0.5 to
# 5.0 will yield large values to small values.
denominator_density = [scipy.stats.norm.pdf(float(i), 5.0, 1.5) for i in lr_index]
return list(zip(lr_index, zip(numerator_density, denominator_density)))
def generate_a_fake_taar_client():
return {
"client_id": "test-client-001",
"activeAddons": [],
"geo_city": "brasilia-br",
"subsession_length": 4911,
"locale": "br-PT",
"os": "mac",
"bookmark_count": 7,
"tab_open_count": 4,
"total_uri": 222,
"unique_tlds": 21
}
class MockNoDataUtils:
def get_s3_json_content(self, *args, **kwargs):
return None
class MockCategoricalData:
cat_data = json.loads(json.dumps(CATEGORICAL_FEATURE_FIXTURE_DATA))
lrs_data = json.loads(json.dumps(generate_fake_lr_curves(1000)))
def get_s3_json_content(self, bucket, key):
if key == DONOR_LIST_KEY:
return self.cat_data
if key == LR_CURVES_SIMILARITY_TO_PROBABILITY:
return self.lrs_data
class MockContinuousData:
cts_data = json.loads(json.dumps(CONTINUOUS_FEATURE_FIXTURE_DATA))
lrs_data = json.loads(json.dumps(generate_fake_lr_curves(1000)))
def get_s3_json_content(self, bucket, key):
if key == DONOR_LIST_KEY:
return self.cts_data
if key == LR_CURVES_SIMILARITY_TO_PROBABILITY:
return self.lrs_data
@pytest.fixture
def cat_test_ctx(test_ctx):
ctx = test_ctx
ctx['utils'] = MockCategoricalData()
ctx['clock'] = Clock()
ctx['cache'] = JSONCache(ctx)
return ctx.child()
@pytest.fixture
def cts_test_ctx(test_ctx):
ctx = test_ctx
ctx['utils'] = MockContinuousData()
ctx['clock'] = Clock()
ctx['cache'] = JSONCache(ctx)
return ctx.child()
def test_soft_fail(test_ctx):
# Create a new instance of a SimilarityRecommender.
ctx = test_ctx
ctx['utils'] = MockNoDataUtils()
ctx['clock'] = Clock()
ctx['cache'] = JSONCache(ctx)
r = SimilarityRecommender(ctx)
# Don't recommend if the source files cannot be found.
assert not r.can_recommend({})
def test_can_recommend(cts_test_ctx):
# Create a new instance of a SimilarityRecommender.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
# Test that we can't recommend if we have not enough client info.
assert not r.can_recommend({})
# Test that we can recommend for a normal client.
assert r.can_recommend(generate_a_fake_taar_client())
# Check that we can not recommend if any required client field is missing.
required_fields = CATEGORICAL_FEATURES + CONTINUOUS_FEATURES
for required_field in required_fields:
profile_without_x = generate_a_fake_taar_client()
# Make an empty value in a required field in the client info dict.
profile_without_x[required_field] = None
assert not r.can_recommend(profile_without_x)
# Completely remove (in place) the entire required field from the dict.
del profile_without_x[required_field]
assert not r.can_recommend(profile_without_x)
def test_recommendations(cts_test_ctx):
# Create a new instance of a SimilarityRecommender.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
# TODO: clobber the SimilarityRecommender::lr_curves
recommendation_list = r.recommend(generate_a_fake_taar_client(), 1)
assert isinstance(recommendation_list, list)
assert len(recommendation_list) == 1
recommendation, weight = recommendation_list[0]
# Make sure that the reported addons are the expected ones from the most similar donor.
assert "{test-guid-1}" == recommendation
assert type(weight) == np.float64
def test_recommender_str(cts_test_ctx):
# Tests that the string representation of the recommender is correct.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
assert str(r) == "SimilarityRecommender"
def test_get_lr(cts_test_ctx):
# Tests that the likelihood ratio values are not empty for extreme values and are realistic.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
assert r.get_lr(0.0001) is not None
assert r.get_lr(10.0) is not None
assert r.get_lr(0.001) > r.get_lr(5.0)
def test_compute_clients_dist(cts_test_ctx):
# Test the distance function computation.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
test_clients = [
{
"client_id": "test-client-002",
"activeAddons": [],
"geo_city": "sfo-us",
"subsession_length": 1,
"locale": "en-US",
"os": "windows",
"bookmark_count": 1,
"tab_open_count": 1,
"total_uri": 1,
"unique_tlds": 1
},
{
"client_id": "test-client-003",
"activeAddons": [],
"geo_city": "brasilia-br",
"subsession_length": 1,
"locale": "br-PT",
"os": "windows",
"bookmark_count": 10,
"tab_open_count": 1,
"total_uri": 1,
"unique_tlds": 1
},
{
"client_id": "test-client-004",
"activeAddons": [],
"geo_city": "brasilia-br",
"subsession_length": 100,
"locale": "br-PT",
"os": "windows",
"bookmark_count": 10,
"tab_open_count": 10,
"total_uri": 100,
"unique_tlds": 10
}
]
per_client_test = []
# Compute a different set of distances for each set of clients.
for tc in test_clients:
test_distances = r.compute_clients_dist(tc)
assert len(test_distances) == len(CONTINUOUS_FEATURE_FIXTURE_DATA)
per_client_test.append(test_distances[2][0])
# Ensure the different clients also had different distances to a specific donor.
assert per_client_test[0] >= per_client_test[1] >= per_client_test[2]
def test_distance_functions(cts_test_ctx):
# Tests the similarity functions via expected output when passing modified client data.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
# Generate a fake client.
test_client = generate_a_fake_taar_client()
recs = r.recommend(test_client, 10)
assert len(recs) > 0
# Make it a generally poor match for the donors.
test_client.update({'total_uri': 10, 'bookmark_count': 2, 'subsession_length': 10})
all_client_values_zero = test_client
# Make all categorical variables non-matching with any donor.
all_client_values_zero.update({key: 'zero' for key in test_client.keys() if key in CATEGORICAL_FEATURES})
recs = r.recommend(all_client_values_zero, 10)
assert len(recs) == 0
# Make all continuous variables equal to zero.
all_client_values_zero.update({key: 0 for key in test_client.keys() if key in CONTINUOUS_FEATURES})
recs = r.recommend(all_client_values_zero, 10)
assert len(recs) == 0
# Make all categorical variables non-matching with any donor.
all_client_values_high = test_client
all_client_values_high.update({key: 'one billion' for key in test_client.keys() if key in CATEGORICAL_FEATURES})
recs = r.recommend(all_client_values_high, 10)
assert len(recs) == 0
# Make all continuous variables equal to a very high numerical value.
all_client_values_high.update({key: 1e60 for key in test_client.keys() if key in CONTINUOUS_FEATURES})
recs = r.recommend(all_client_values_high, 10)
assert len(recs) == 0
# Test for 0.0 values if j_c is not normalized and j_d is fine.
j_c = 0.0
j_d = 0.42
assert abs(j_c * j_d) == 0.0
assert abs((j_c + 0.01) * j_d) != 0.0
def test_weights_continuous(cts_test_ctx):
# Create a new instance of a SimilarityRecommender.
ctx = cts_test_ctx
r = SimilarityRecommender(ctx)
# In the ensemble method recommendations should be a sorted list of tuples
# containing [(guid, weight), (guid, weight)... (guid, weight)].
recommendation_list = r.recommend(generate_a_fake_taar_client(), 2)
with open('/tmp/similarity_recommender.json', 'w') as fout:
fout.write(json.dumps(recommendation_list))
# Make sure the structure of the recommendations is correct and
# that we recommended the the right addons.
assert len(recommendation_list) == 2
for recommendation, weight in recommendation_list:
assert isinstance(recommendation, six.string_types)
assert isinstance(weight, float)
# Test that sorting is appropriate.
rec0 = recommendation_list[0]
rec1 = recommendation_list[1]
rec0_weight = rec0[1]
rec1_weight = rec1[1]
# Duplicate presence of test-guid-1 should mean rec0_weight is double
# rec1_weight, and both should be greater than 1.0
assert rec0_weight > rec1_weight > 1.0
def test_weights_categorical(cat_test_ctx, cts_test_ctx):
'''
This should get :
["{test-guid-1}", "{test-guid-2}", "{test-guid-3}", "{test-guid-4}"],
["{test-guid-9}", "{test-guid-10}", "{test-guid-11}", "{test-guid-12}"]
from the first two entries in the sample data where the geo_city
data
'''
# Create a new instance of a SimilarityRecommender.
ctx = cat_test_ctx
ctx2 = cts_test_ctx
wrapped = ctx2.wrap(ctx)
r = SimilarityRecommender(wrapped)
# In the ensemble method recommendations should be a sorted list of tuples
# containing [(guid, weight), (guid, weight)... (guid, weight)].
recommendation_list = r.recommend(generate_a_fake_taar_client(), 2)
assert len(recommendation_list) == 2
# Make sure the structure of the recommendations is correct and that we recommended the the right addons.
for recommendation, weight in recommendation_list:
assert isinstance(recommendation, six.string_types)
assert isinstance(weight, float)
# Test that sorting is appropriate.
rec0 = recommendation_list[0]
rec1 = recommendation_list[1]
rec0_weight = rec0[1]
rec1_weight = rec1[1]
assert rec0_weight > rec1_weight > 0