def test_fit(self):
rec = ExplicitItemFinder()
data = load_movielens_mini()
true_sc = np.where(data.score <= 3, 0, 1)
data.binarize_score()
rec.fit(data, event_index=(0, 1))
# check whether score info is correctly set
assert_allclose(rec.score_domain, [0, 1, 1])
assert_allclose(rec.score, true_sc)
assert_equal(rec.n_score_levels, 2)
# error checker
data.score_domain = (0, 5, 1)
with assert_raises(ValueError):
rec.fit(data)
data.score_domain = (0, 1, 1)
data.n_score_levels = 3
with assert_raises(ValueError):
rec.fit(data)
data.n_score_levels = 2
data.score = np.zeros(30)
with assert_raises(ValueError):
rec.fit(data)
data.score = np.ones(30)
with assert_raises(ValueError):
rec.fit(data)
data.score = np.repeat(np.arange(3), 10)
with assert_raises(ValueError):
rec.fit(data)
def test_class(self):
# setup
data = load_movielens_mini()
# __init__()
rec = Recommender(random_state=1234)
self.assertEqual(rec.n_otypes, 0)
self.assertIsNone(rec.n_objects)
self.assertIsNone(rec.eid)
self.assertIsNone(rec.iid)
self.assertEqual(rec.random_state, 1234)
self.assertIsNone(rec._rng)
self.assertDictEqual(rec.fit_results_, {})
# _set_object_info()
rec._set_object_info(data)
self.assertEqual(rec.n_otypes, 2)
assert_array_equal(rec.n_objects, [8, 10])
assert_array_equal(rec.eid, data.eid)
assert_array_equal(rec.iid, data.iid)
# to_eid()
self.assertEqual(rec.to_eid(0, 0), 1)
self.assertEqual(rec.to_eid(1, 0), 1)
# to_iid()
self.assertEqual(rec.to_iid(0, 1), 0)
self.assertEqual(rec.to_iid(1, 1), 0)
def test_fit(self):
rec = ExplicitItemFinder()
data = load_movielens_mini()
true_sc = np.where(data.score <= 3, 0, 1)
data.binarize_score()
rec.fit(data, event_index=(0, 1))
# check whether score info is correctly set
assert_allclose(rec.score_domain, [0, 1, 1])
assert_allclose(rec.score, true_sc)
assert_equal(rec.n_score_levels, 2)
# error checker
data.score_domain = (0, 5, 1)
with assert_raises(ValueError):
rec.fit(data)
data.score_domain = (0, 1, 1)
data.n_score_levels = 3
with assert_raises(ValueError):
rec.fit(data)
data.n_score_levels = 2
data.score = np.zeros(30)
with assert_raises(ValueError):
rec.fit(data)
data.score = np.ones(30)
with assert_raises(ValueError):
rec.fit(data)
data.score = np.repeat(np.arange(3), 10)
with assert_raises(ValueError):
rec.fit(data)
def test_class(self):
# setup
data = load_movielens_mini()
# __init__()
rec = Recommender(random_state=1234)
self.assertEqual(rec.n_otypes, 0)
self.assertIsNone(rec.n_objects)
self.assertIsNone(rec.eid)
self.assertIsNone(rec.iid)
self.assertEqual(rec.random_state, 1234)
self.assertIsNone(rec._rng)
self.assertDictEqual(rec.fit_results_, {})
# _set_object_info()
rec._set_object_info(data)
self.assertEqual(rec.n_otypes, 2)
assert_array_equal(rec.n_objects, [8, 10])
assert_array_equal(rec.eid, data.eid)
assert_array_equal(rec.iid, data.iid)
# to_eid()
self.assertEqual(rec.to_eid(0, 0), 1)
self.assertEqual(rec.to_eid(1, 0), 1)
# to_iid()
self.assertEqual(rec.to_iid(0, 1), 0)
self.assertEqual(rec.to_iid(1, 1), 0)
def test_binarize_score(self):
data = load_movielens_mini()
data.binarize_score(2)
assert_array_equal(data.score_domain, [0, 1, 1])
assert_array_equal(
data.score,
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1])
assert_equal(data.n_score_levels, 2)
data = load_movielens_mini()
data.binarize_score()
assert_array_equal(
data.score,
[0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 1,
0, 1, 0, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1])
def test_binarize_score(self):
data = load_movielens_mini()
data.binarize_score(2)
assert_array_equal(data.score_domain, [0, 1, 1])
assert_array_equal(data.score, [
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1
])
assert_equal(data.n_score_levels, 2)
data = load_movielens_mini()
data.binarize_score()
assert_array_equal(data.score, [
0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1,
1, 0, 1, 1, 1, 1, 1, 1
])
def test_filter_event(self):
from kamrecsys.data import EventWithScoreData
# load movie_lens
data = load_movielens_mini()
# filter events
filter_cond = np.arange(data.n_events) % 3 == 0
filtered_data = super(
EventWithScoreData, data).filter_event(filter_cond)
assert_array_equal(
filtered_data.event[:, 0], [1, 5, 3, 4, 0, 0, 0, 2, 2, 0])
assert_array_equal(
filtered_data.event[:, 1], [1, 3, 6, 5, 7, 6, 4, 0, 7, 2])
assert_array_equal(
filtered_data.to_eid(0, filtered_data.event[:, 0]),
data.to_eid(0, data.event[filter_cond, 0]))
assert_array_equal(
filtered_data.to_eid(1, filtered_data.event[:, 1]),
data.to_eid(1, data.event[filter_cond, 1]))
assert_array_equal(
filtered_data.event_feature['timestamp'],
[875636053, 877889130, 891351328, 879362287, 878543541,
875072484, 889751712, 883599478, 883599205, 878542960])
assert_array_equal(filtered_data.eid[0], [1, 5, 6, 7, 8, 10])
assert_array_equal(filtered_data.eid[1], [1, 2, 3, 4, 5, 7, 8, 9])
assert_equal(
filtered_data.iid[0],
{1: 0, 5: 1, 6: 2, 7: 3, 8: 4, 10: 5})
assert_equal(
filtered_data.iid[1],
{1: 0, 2: 1, 3: 2, 4: 3, 5: 4, 7: 5, 8: 6, 9: 7})
assert_equal(
filtered_data.feature[0]['zip'],
[u'85711', u'15213', u'98101', u'91344', u'05201', u'90703'])
assert_equal(
filtered_data.feature[1]['name'],
[u'Toy Story (1995)', u'GoldenEye (1995)', u'Four Rooms (1995)',
u'Get Shorty (1995)', u'Copycat (1995)', u'Twelve Monkeys (1995)',
u'Babe (1995)', u'Dead Man Walking (1995)'])
# dummy event data
data = EventData()
data.set_event(np.tile(np.arange(5), (2, 2)).T)
filtered_data = data.filter_event(
[True, False, True, True, False, False, True, True, False, False])
assert_equal(filtered_data.n_events, 5)
assert_array_equal(
filtered_data.event, [[0, 0], [2, 2], [3, 3], [1, 1], [2, 2]])
def test_filter_event(self):
data = load_movielens_mini()
filtered_data = data.filter_event(data.score > 3)
assert_allclose(filtered_data.score, [
4., 4., 4., 5., 4., 5., 5., 5., 4., 5., 4., 5., 5., 4., 5., 4., 4.,
4., 4., 4., 4.
])
assert_allclose(filtered_data.to_eid(0, filtered_data.event[:, 0]), [
10, 5, 10, 1, 7, 7, 9, 7, 10, 1, 2, 1, 7, 6, 7, 6, 1, 9, 1, 6, 10
])
def test_filter_event(self):
data = load_movielens_mini()
filtered_data = data.filter_event(data.score > 3)
assert_allclose(
filtered_data.score,
[4., 4., 4., 5., 4., 5., 5., 5., 4., 5.,
4., 5., 5., 4., 5., 4., 4., 4., 4., 4., 4.])
assert_allclose(
filtered_data.to_eid(0, filtered_data.event[:, 0]),
[10, 5, 10, 1, 7, 7, 9, 7, 10, 1,
2, 1, 7, 6, 7, 6, 1, 9, 1, 6, 10])
def test_class(self):
import numpy as np
from kamrecsys.datasets import load_movielens_mini
from kamrecsys.mf.pmf import EventScorePredictor
data = load_movielens_mini()
recommender = EventScorePredictor(C=0.1, k=2, tol=1e-03,
random_state=1234)
self.assertDictEqual(
vars(recommender),
{'C': 0.1, 'n_otypes': 0, 'bu_': None, 'bi_': None, 'k': 2,
'p_': None, 'q_': None, '_coef': None, 'f_loss_': np.inf,
'iid': None, 'i_loss_': np.inf, 'eid': None, 'tol': 1e-03,
'n_objects': None, '_dt': None, 'mu_': None, 'opt_outputs_': None,
'maxiter': 200, 'random_state': 1234, '_rng': None})
recommender.fit(data, disp=False)
self.assertAlmostEqual(recommender.i_loss_,
0.74652578358324106, delta=1e-5)
self.assertAlmostEqual(recommender.f_loss_,
0.025638738121075231, delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 7)),
3.9873641434545979, delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 9)),
4.9892118821609106, delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 11)),
3.6480799850368273, delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 7)),
3.6336318795279228, delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 9)),
4.2482001235634943, delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 11)),
3.7236984083417841, delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 7)),
3.4141968145802597, delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 9)),
3.9818882049478654, delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 11)),
3.4710520150321895, delta=1e-5)
x = np.array([
[1, 7], [1, 9], [1, 11],
[3, 7], [3, 9], [3, 11],
[5, 7], [5, 9], [5, 11]])
assert_allclose(
recommender.predict(x),
[3.98736414, 4.98921188, 3.64807999, 3.63363188, 4.24820012,
3.72369841, 3.41419681, 3.9818882, 3.47105202],
rtol=1e-5)
def test_class(self):
data = load_movielens_mini()
data.binarize_score()
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(rec.fit_results_['initial_loss'],
0.7665586528005817,
rtol=1e-5)
assert_allclose(rec.fit_results_['final_loss'],
0.07714505638740253,
rtol=1e-5)
# raw_predict
assert_allclose(rec.raw_predict(np.array([[0, 6]])),
0.999880367504,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[8, 8]])),
0.976596331684,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[2, 10]])),
0.886485084449,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[0.999880367504, 0.976596331684, 0.886485084449],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 0.999880367504, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 0.985475882857, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 0.851425801703, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 0.890682628776, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 0.976596331684, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 0.932930852359, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 0.998741832467, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 0.962962765249, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 0.886485084449, rtol=1e-5)
# multiple prediction
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(rec.predict(x), [
0.999880367504, 0.985475882857, 0.851425801703, 0.890682628776,
0.976596331684, 0.932930852359, 0.998741832467, 0.962962765249,
0.886485084449
],
rtol=1e-5)
def test_class(self):
rec = ExplicitItemFinder()
data = load_movielens_mini()
true_sc = np.where(data.score <= 3, 0, 1)
data.binarize_score()
rec.fit(data, event_index=(0, 1))
# get_score()
assert_allclose(rec.get_score(), true_sc)
# remove_data
rec.remove_data()
self.assertIsNone(rec.score)
assert_allclose(rec.score_domain, [0, 1, 1])
self.assertEqual(rec.n_score_levels, 2)
def test_class(self):
rec = ExplicitItemFinder()
data = load_movielens_mini()
true_sc = np.where(data.score <= 3, 0, 1)
data.binarize_score()
rec.fit(data, event_index=(0, 1))
# get_score()
assert_allclose(rec.get_score(), true_sc)
# remove_data
rec.remove_data()
self.assertIsNone(rec.score)
assert_allclose(rec.score_domain, [0, 1, 1])
self.assertEqual(rec.n_score_levels, 2)
def test_class(self):
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(rec.fit_results_['initial_loss'],
0.7291206184050988,
rtol=1e-5)
assert_allclose(rec.fit_results_['final_loss'],
0.013777730279425669,
rtol=1e-5)
# raw_predict
assert_allclose(rec.raw_predict(np.array([[0, 6]])),
3.997101590073,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[8, 8]])),
4.24151516373,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[2, 10]])),
3.498237262002,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[3.997101590073, 4.24151516373, 3.498237262002],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 3.997101590073, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 4.977093755711, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 3.62779086784, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 3.683330861026, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 4.24151516373, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 3.70802937382, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 3.521554946725, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 4.000964107588, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 3.498237262002, rtol=1e-5)
# multiple prediction
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(rec.predict(x), [
3.997101590073, 4.977093755711, 3.62779086784, 3.683330861026,
4.24151516373, 3.70802937382, 3.521554946725, 4.000964107588,
3.498237262002
],
rtol=1e-5)
def test_class(self):
# setup
data = load_movielens_mini()
rec = EventRecommender()
rec._set_object_info(data)
# _set_event_info
rec._set_event_info(data)
self.assertEqual(rec.s_event, 2)
assert_array_equal(rec.event_otypes, [0, 1])
self.assertEqual(rec.n_events, 30)
assert_array_equal(
rec.event[0, :],
[rec.to_iid(0, 5), rec.to_iid(1, 2)])
assert_array_equal(
rec.event[-1, :],
[rec.to_iid(0, 10), rec.to_iid(1, 9)])
ts = sorted(rec.event_feature['timestamp'])
assert_array_equal(ts[:2], [874965758, 875071561])
assert_array_equal(ts[-2:], [891352220, 891352864])
self.assertIsNone(rec.event_index)
# fit
rec.fit(data, event_index=(1, 0))
assert_array_equal(rec.event_index, [1, 0])
# get_event
ev, n_objects = rec.get_event()
assert_array_equal(ev[0, :], [rec.to_iid(1, 2), rec.to_iid(0, 5)])
assert_array_equal(ev[-1, :], [rec.to_iid(1, 9), rec.to_iid(0, 10)])
assert_array_equal(n_objects, [10, 8])
# predict
self.assertEqual(rec.predict([0, 0]).ndim, 0)
self.assertEqual(rec.predict([[0, 0]]).ndim, 0)
self.assertEqual(rec.predict([[0, 0], [0, 1]]).ndim, 1)
assert_array_equal(rec.predict([[0, 0], [0, 1]]).shape, (2, ))
# remove_data
rec.remove_data()
self.assertEqual(rec.n_events, 30)
self.assertIsNone(rec.event)
self.assertIsNone(rec.event_feature)
assert_array_equal(rec.event_index, (1, 0))
def test_class(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(rec.fit_results_['initial_loss'],
1.3111240714555008,
rtol=1e-5)
assert_allclose(rec.fit_results_['final_loss'],
0.21148605953543897,
rtol=1e-5)
# raw_predict
assert_allclose(rec.raw_predict(np.array([[0, 6]])),
0.98444253128,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[8, 8]])),
0.225589860753,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[2, 10]])),
0.090758250123,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[0.98444253128, 0.225589860753, 0.090758250123],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 0.98444253128, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 0.924884069088, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 0.952482724921, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 0.618744630907, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 0.225589860753, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 0.2295914768, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 0.00080319875342, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 0.0396585424477, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 0.0907582501228, rtol=1e-5)
# multiple prediction
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(rec.predict(x), [
0.98444253128, 0.924884069088, 0.952482724921, 0.618744630907,
0.225589860753, 0.2295914768, 0.00080319875342, 0.0396585424477,
0.0907582501228
],
rtol=1e-5)
def test_class(self):
data = load_movielens_mini()
data.binarize_score()
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(
rec.fit_results_['initial_loss'], 23.7963271809711, rtol=1e-5)
assert_allclose(
rec.fit_results_['final_loss'], 3.48393252113573, rtol=1e-5)
# raw_predict
assert_allclose(
rec.raw_predict(np.array([[0, 6]])), 0.989201290366165, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[8, 8]])), 0.959375735295748, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[2, 10]])), 0.824034860051334, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[0.989201290366, 0.959375735296, 0.824034860051],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 0.989201290366165, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 0.974756390738565, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 0.816803762429987, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 0.854574774389645, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 0.959375735295748, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 0.900999255222352, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 0.888811789653219, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 0.9634835048284, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 0.824034860051334, rtol=1e-5)
# multiple prediction
x = np.array([
[1, 7], [1, 9], [1, 11],
[3, 7], [3, 9], [3, 11],
[5, 7], [5, 9], [5, 11]])
assert_allclose(
rec.predict(x),
[0.989201290366, 0.974756390739, 0.81680376243,
0.85457477439, 0.959375735296, 0.900999255222,
0.888811789653, 0.963483504828, 0.824034860051],
rtol=1e-5)
def test_class(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(
rec.fit_results_['initial_loss'], 103.98807319757, rtol=1e-5)
assert_allclose(
rec.fit_results_['final_loss'], 13.8730554159163, rtol=1e-5)
# raw_predict
assert_allclose(
rec.raw_predict(np.array([[0, 6]])), 0.997028296278715, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[8, 8]])), 0.131044842032382, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[2, 10]])), 0.058370496087627, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[0.9970282963, 0.131044842, 0.0583704961],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 0.997028296278715, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 0.944998324514573, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 0.964923027550838, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 0.624248406054208, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 0.131044842032382, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 0.174497501438718, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 4.07951069077663e-05, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 0.00230569601442113, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 0.058370496087627, rtol=1e-5)
# multiple prediction
x = np.array([
[1, 7], [1, 9], [1, 11],
[3, 7], [3, 9], [3, 11],
[5, 7], [5, 9], [5, 11]])
assert_allclose(
rec.predict(x),
[9.9702829628e-01, 9.4499832451e-01, 9.6492302755e-01,
6.2424840605e-01, 1.3104484203e-01, 1.7449750144e-01,
4.0795106908e-05, 2.3056960144e-03, 5.8370496088e-02],
rtol=1e-5)
def test_class(self):
# setup
data = load_movielens_mini()
rec = EventRecommender()
rec._set_object_info(data)
# _set_event_info
rec._set_event_info(data)
self.assertEqual(rec.s_event, 2)
assert_array_equal(rec.event_otypes, [0, 1])
self.assertEqual(rec.n_events, 30)
assert_array_equal(
rec.event[0, :], [rec.to_iid(0, 5), rec.to_iid(1, 2)])
assert_array_equal(
rec.event[-1, :], [rec.to_iid(0, 10), rec.to_iid(1, 9)])
ts = sorted(rec.event_feature['timestamp'])
assert_array_equal(ts[:2], [874965758, 875071561])
assert_array_equal(ts[-2:], [891352220, 891352864])
self.assertIsNone(rec.event_index)
# fit
rec.fit(data, event_index=(1, 0))
assert_array_equal(rec.event_index, [1, 0])
# get_event
ev, n_objects = rec.get_event()
assert_array_equal(ev[0, :], [rec.to_iid(1, 2), rec.to_iid(0, 5)])
assert_array_equal(ev[-1, :], [rec.to_iid(1, 9), rec.to_iid(0, 10)])
assert_array_equal(n_objects, [10, 8])
# predict
self.assertEqual(rec.predict([0, 0]).ndim, 0)
self.assertEqual(rec.predict([[0, 0]]).ndim, 0)
self.assertEqual(rec.predict([[0, 0], [0, 1]]).ndim, 1)
assert_array_equal(rec.predict([[0, 0], [0, 1]]).shape, (2,))
# remove_data
rec.remove_data()
self.assertEqual(rec.n_events, 30)
self.assertIsNone(rec.event)
self.assertIsNone(rec.event_feature)
assert_array_equal(rec.event_index, (1, 0))
def test_class(self):
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec.fit(data)
assert_allclose(
rec.fit_results_['initial_loss'], 11.6412258667768, rtol=1e-5)
assert_allclose(
rec.fit_results_['final_loss'], 0.683795023344032, rtol=1e-5)
# raw_predict
assert_allclose(
rec.raw_predict(np.array([[0, 6]])), 3.98981537208499, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[8, 8]])), 4.25466828602487, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[2, 10]])), 3.45221824936513, rtol=1e-5)
assert_allclose(
rec.raw_predict(np.array([[0, 6], [8, 8], [2, 10]])),
[3.9898153721, 4.254668286, 3.4522182494],
rtol=1e-5)
# single prediction
assert_allclose(rec.predict((1, 7)), 3.98981537208499, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 4.96457075887952, rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 3.64937487633011, rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 3.56199613750802, rtol=1e-5)
assert_allclose(rec.predict((3, 9)), 4.25466828602487, rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 3.74991933999025, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 3.27152744041802, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 3.99212459353293, rtol=1e-5)
assert_allclose(rec.predict((5, 11)), 3.45221824936513, rtol=1e-5)
# multiple prediction
x = np.array([
[1, 7], [1, 9], [1, 11],
[3, 7], [3, 9], [3, 11],
[5, 7], [5, 9], [5, 11]])
assert_allclose(
rec.predict(x),
[3.9898153721, 4.9645707589, 3.6493748763,
3.5619961375, 4.254668286, 3.74991934,
3.2715274404, 3.9921245935, 3.4522182494],
rtol=1e-5)
def test_class(self):
data = load_movielens_mini()
rec = ScorePredictor()
# fit()
rec.fit(data, event_index=(0, 1))
assert_allclose(rec.score_domain, [1., 5., 1.])
assert_allclose(rec.score, true_sc)
assert_equal(rec.n_score_levels, 5)
# get_score()
assert_allclose(rec.get_score(), true_sc)
# remove_data
rec.remove_data()
self.assertIsNone(rec.score)
assert_allclose(rec.score_domain, [1., 5., 1.])
assert_equal(rec.n_score_levels, 5)
def test_class(self):
data = load_movielens_mini()
rec = ScorePredictor()
# fit()
rec.fit(data, event_index=(0, 1))
assert_allclose(rec.score_domain, [1., 5., 1.])
assert_allclose(rec.score, true_sc)
assert_equal(rec.n_score_levels, 5)
# get_score()
assert_allclose(rec.get_score(), true_sc)
# remove_data
rec.remove_data()
self.assertIsNone(rec.score)
assert_allclose(rec.score_domain, [1., 5., 1.])
assert_equal(rec.n_score_levels, 5)
def test__get_event_array(self):
rec = ImplicitItemFinder()
data = load_movielens_mini()
data.filter_event(
np.logical_and(data.event[:, 0] < 5, data.event[:, 1] < 5))
rec.fit(data)
ev, n_objects = rec.get_event_array(sparse_type='array')
assert_array_equal(ev[:5, :5],
[[1, 1, 1, 1, 1], [1, 0, 0, 0, 0], [1, 1, 0, 0, 0],
[1, 0, 0, 0, 0], [0, 0, 0, 1, 0]])
ev2, n_objects = rec.get_event_array('csr')
assert_array_equal(ev, ev2.todense())
ev2, n_objects = rec.get_event_array('csc')
assert_array_equal(ev, ev2.todense())
ev2, n_objects = rec.get_event_array('lil')
assert_array_equal(ev, ev2.todense())
def test__get_event_array(self):
rec = ImplicitItemFinder()
data = load_movielens_mini()
data.filter_event(
np.logical_and(data.event[:, 0] < 5, data.event[:, 1] < 5))
rec.fit(data)
ev, n_objects = rec.get_event_array(sparse_type='array')
assert_array_equal(
ev[:5, :5],
[[1, 1, 1, 1, 1], [1, 0, 0, 0, 0], [1, 1, 0, 0, 0],
[1, 0, 0, 0, 0], [0, 0, 0, 1, 0]])
ev2, n_objects = rec.get_event_array('csr')
assert_array_equal(ev, ev2.todense())
ev2, n_objects = rec.get_event_array('csc')
assert_array_equal(ev, ev2.todense())
ev2, n_objects = rec.get_event_array('lil')
assert_array_equal(ev, ev2.todense())
def test_loss(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
n_objects = data.n_objects
ev = sparse.coo_matrix((np.ones(data.n_events, dtype=int),
(data.event[:, 0], data.event[:, 1])),
shape=n_objects)
ev = ev.tocsr()
rec._init_coef(ev, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(rec.loss(rec._coef, ev, n_objects),
1.3111240714555008,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(rec.loss(rec._coef, ev, n_objects),
0.69314718055994518,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(rec.loss(rec._coef, ev, n_objects),
3.1817153484891296,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(rec.loss(rec._coef, ev, n_objects),
2.516800033025768,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(rec.loss(rec._coef, ev, n_objects),
7.466231131678393,
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
data.binarize_score()
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -2.69422502892223, rtol=1e-5)
assert_allclose(
grad[1:5],
[1.4860354254, 0.5813833517, -0.2378488265, -0.4549466355],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-1.4192072182, -0.4856324281, -1.793371146, 1.1698467502],
rtol=1e-5)
assert_allclose(
grad[19:23],
[2.218012237, -2.8022576986, 0.7577517736, -0.189979212],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.252934635, 2.569059726, 1.5578077993, -0.4423181343],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -6.0, rtol=1e-5)
assert_allclose(
grad[1:5],
[0., 0., 0., -1.],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-1., -0.5, -2., 0.5],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0., 0., 0., 0.],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0., 0., 0., 0.],
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 8.799214472271457, rtol=1e-5)
assert_allclose(
grad[1:5],
[5.0330714908, 1.0866142982, 1.0866142982, 1.0732285963],
rtol=1e-5)
assert_allclose(
grad[15:19],
[2.0598428945, 1.0799214472, 0.0732285963, 2.0799214472],
rtol=1e-5)
assert_allclose(
grad[19:23],
[5.0330714908, 5.0330714908, 1.0866142982, 1.0866142982],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.0732285963, 0.0732285963, 2.0799214472, 2.0799214472],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 8.323751110862577, rtol=1e-5)
assert_allclose(
grad[1:5],
[4.7101163359, 0.9545317778, 0.952133346, 0.9486417943],
rtol=1e-5)
assert_allclose(
grad[15:19],
[2.0612948847, 1.117042198, 0.1220469429, 2.1435828557],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.9920232672, 4.8101163359, 0.2389063556, 1.0445317778],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.0089765286, 0.0420469429, 0.9967914278, 2.0535828557],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 8.999701753193309, rtol=1e-5)
assert_allclose(
grad[1:5],
[5.0798783079, 1.0699742535, 1.0599646235, 1.0499642981],
rtol=1e-5)
assert_allclose(
grad[15:19],
[2.1899607439, 1.2049810091, 0.2199797513, 2.2349854785],
rtol=1e-5)
assert_allclose(
grad[19:23],
[11.0797319135, 2.4999606492, 2.3379421381, 0.9999954835],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.2059342372, 0.0799797513, 6.2829547023, 2.0899854785],
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
data.binarize_score()
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -0.0898075009641, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.0498308105, 0.0196757414, -0.0076319979, -0.0152389619],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-0.0472451789, -0.0160642908, -0.060056816, 0.0394887188],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0732353188, -0.0916441815, 0.0231358636, -0.0058694524],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.0099831228, 0.084085786, 0.0506473455, -0.0145630614],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -0.2, rtol=1e-5)
assert_allclose(grad[1:5], [0., 0., 0., -0.0333333333], rtol=1e-5)
assert_allclose(
grad[15:19],
[-0.0333333333, -0.0166666667, -0.0666666667, 0.0166666667],
rtol=1e-5)
assert_allclose(grad[19:23], [0., 0., 0., 0.], rtol=1e-5)
assert_allclose(grad[-4:], [0., 0., 0., 0.], rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 0.293307149076, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.1662875682, 0.0347389951, 0.0347389951, 0.0342928051],
rtol=1e-5)
assert_allclose(
grad[15:19],
[0.0671799483, 0.0345159001, 0.0009594717, 0.0678492334],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.1662875682, 0.1662875682, 0.0347389951, 0.0347389951],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.0009594717, 0.0009594717, 0.0678492334, 0.0678492334],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 0.277458370362, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.1570038779, 0.0316695778, 0.0314414819, 0.0311769487],
rtol=1e-4)
assert_allclose(
grad[15:19],
[0.0663394591, 0.0347162214, 0.0014015648, 0.068637947],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0323267015, 0.1588557297, 0.0072228044, 0.0333362444],
rtol=1e-5)
assert_allclose(grad[-4:], [
-5.9551391499e-04, -7.9916718879e-05, 3.2930084632e-02,
6.6971280375e-02
],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 0.29999005844, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.1681440917, 0.0346287714, 0.0344432652, 0.0342580692],
rtol=1e-5)
assert_allclose(
grad[15:19],
[0.0701838766, 0.0371289966, 0.0040733991, 0.0710180345],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.3648799527, 0.0818505402, 0.0733684416, 0.0318517013],
rtol=1e-5)
assert_allclose(
grad[-4:], [0.0038126227, 0.0014808065, 0.206424416, 0.0683328493],
rtol=1e-5)
def test_get_score_levels(self):
data = load_movielens_mini()
score_levels = data.get_score_levels()
assert_allclose(score_levels, [1., 2., 3., 4., 5])
def test_loss(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
n_objects = data.n_objects
ev = sparse.coo_matrix(
(np.ones(data.n_events, dtype=int),
(data.event[:, 0], data.event[:, 1])), shape=n_objects)
ev = ev.tocsr()
rec._init_coef(ev, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(
rec.loss(rec._coef, ev, n_objects), 103.98807319757,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(
rec.loss(rec._coef, ev, n_objects), 55.4517744447956,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, n_objects), 253.23722787913,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, n_objects), 200.293169308728,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(
rec.loss(rec._coef, ev, n_objects), 593.059407200938,
rtol=1e-5)
def test_loss(self):
# setup
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 11.6412258667768,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 235.5,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 38.2,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 18.7025,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 947.15920616,
rtol=1e-5)
def test_loss(self):
# setup
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
0.7291206184050988,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
15.7,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
2.4166666666666665,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
1.14175,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
62.720038744000014,
rtol=1e-5)
def test_loss(self):
# setup
data = load_movielens_mini()
data.binarize_score(3)
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 23.7963271809711,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 20.7944154167983,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 47.9014604546737,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 36.467883082757,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(
rec.loss(rec._coef, ev, sc, n_objects), 113.246948248756,
rtol=1e-5)
def test_get_score_levels(self):
data = load_movielens_mini()
score_levels = data.get_score_levels()
assert_allclose(score_levels, [1., 2., 3., 4., 5])
def test_class(self):
import numpy as np
data = load_movielens_mini()
rec = MultinomialPLSA(k=2, random_state=1234, tol=1e-8)
# import logging
# logging.getLogger('kamrecsys').addHandler(logging.StreamHandler())
rec.fit(data)
assert_equal(rec.fit_results_['n_users'], 8)
assert_equal(rec.fit_results_['n_items'], 10)
assert_equal(rec.fit_results_['n_events'], 30)
assert_(rec.fit_results_['success'])
assert_equal(rec.fit_results_['status'], 0)
assert_allclose(rec.fit_results_['initial_loss'],
5.41836900049,
rtol=1e-5)
assert_allclose(rec.fit_results_['final_loss'],
5.17361298499,
rtol=1e-5)
assert_equal(rec.fit_results_['n_iterations'], 38)
assert_allclose(rec.score_levels, [1, 2, 3, 4, 5], rtol=1e-5)
# output size
assert_array_equal(
rec.raw_predict(data.event[0].reshape(1, 2)).shape, (1, ))
assert_array_equal(rec.raw_predict(data.event).shape, (30, ))
rec.use_expectation = False
assert_array_equal(
rec.raw_predict(data.event[0].reshape(1, 2)).shape, (1, ))
assert_array_equal(rec.raw_predict(data.event).shape, (30, ))
rec.use_expectation = True
# known user and item
assert_allclose(rec.raw_predict(np.array([[0, 6]])),
3.64580117249,
rtol=1e-5)
assert_allclose(rec.raw_predict(np.array([[2, 8]])),
3.62184516985,
rtol=1e-5)
assert_allclose(rec.predict((1, 7)), 3.64580117249, rtol=1e-5)
assert_allclose(rec.predict((1, 9)), 3.6587422493, rtol=1e-5)
assert_allclose(rec.predict((5, 7)), 3.60707987724, rtol=1e-5)
assert_allclose(rec.predict((5, 9)), 3.62184516985, rtol=1e-5)
# known user and unknown item
assert_allclose(rec.raw_predict(np.array([[2, 10]])),
3.62387542269,
rtol=1e-5)
assert_allclose(rec.predict((1, 11)), 3.66032199689, rtol=1e-5)
assert_allclose(rec.predict((5, 12)), 3.62387542269, rtol=1e-5)
# unknown user and known item
assert_allclose(rec.raw_predict(np.array([[8, 6]])),
3.60821491793,
rtol=1e-5)
assert_allclose(rec.predict((3, 7)), 3.60821491793, rtol=1e-5)
assert_allclose(rec.predict((11, 9)), 3.62304301551, rtol=1e-5)
# unknown user and item
assert_allclose(rec.raw_predict(np.array([[8, 10]])),
3.62507437787,
rtol=1e-5)
assert_allclose(rec.predict((3, 11)), 3.62507437787, rtol=1e-5)
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(rec.predict(x), [
3.64580117249, 3.6587422493, 3.66032199689, 3.60821491793,
3.62304301551, 3.62507437787, 3.60707987724, 3.62184516985,
3.62387542269
],
rtol=1e-5)
rec.use_expectation = False
assert_allclose(rec.predict(x), [4., 5., 5., 4., 4., 4., 4., 4., 4.],
rtol=1e-5)
def test_class(self):
import numpy as np
from kamrecsys.datasets import load_movielens_mini
from kamrecsys.mf.pmf import EventScorePredictor
data = load_movielens_mini()
recommender = EventScorePredictor(C=0.1,
k=2,
tol=1e-03,
random_state=1234)
self.assertDictEqual(
vars(recommender), {
'C': 0.1,
'n_otypes': 0,
'bu_': None,
'bi_': None,
'k': 2,
'p_': None,
'q_': None,
'_coef': None,
'f_loss_': np.inf,
'iid': None,
'i_loss_': np.inf,
'eid': None,
'tol': 1e-03,
'n_objects': None,
'_dt': None,
'mu_': None,
'opt_outputs_': None,
'maxiter': 200,
'random_state': 1234,
'_rng': None
})
recommender.fit(data, disp=False)
self.assertAlmostEqual(recommender.i_loss_,
0.74652578358324106,
delta=1e-5)
self.assertAlmostEqual(recommender.f_loss_,
0.025638738121075231,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 7)),
3.9873641434545979,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 9)),
4.9892118821609106,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((1, 11)),
3.6480799850368273,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 7)),
3.6336318795279228,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 9)),
4.2482001235634943,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((3, 11)),
3.7236984083417841,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 7)),
3.4141968145802597,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 9)),
3.9818882049478654,
delta=1e-5)
self.assertAlmostEqual(recommender.predict((5, 11)),
3.4710520150321895,
delta=1e-5)
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(recommender.predict(x), [
3.98736414, 4.98921188, 3.64807999, 3.63363188, 4.24820012,
3.72369841, 3.41419681, 3.9818882, 3.47105202
],
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
n_objects = data.n_objects
ev = sparse.coo_matrix((np.ones(data.n_events, dtype=int),
(data.event[:, 0], data.event[:, 1])),
shape=n_objects)
ev = ev.tocsr()
rec._init_coef(ev, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 0.369344884225, rtol=1e-5)
assert_allclose(
grad[1:5],
[-0.0163795867, 0.0766087539, 0.0494371859, 0.0523347865],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-0.0079542759, 0.0300129557, 0.0221887768, 0.0444588993],
rtol=1e-5)
assert_allclose(
grad[19:23],
[-0.00388247, -0.0106748979, 0.047492578, -0.0228685915],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.0284519191, 0.0318876738, 0.0234999572, -0.0061334269],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 0.125, atol=1e-5)
assert_allclose(grad[1:5], [-0.0625, 0.0375, 0.0375, 0.0125],
rtol=1e-5)
assert_allclose(grad[15:19], [-0.025, 0.0125, 0., 0.0125], rtol=1e-5)
assert_allclose(grad[19:23], [0., 0., 0., 0.], rtol=1e-5)
assert_allclose(grad[-4:], [0., 0., 0., 0.], rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 0.618307149076, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.0010152455, 0.1010152455, 0.1010152455, 0.0760152455],
rtol=1e-5)
assert_allclose(
grad[15:19],
[0.0261825668, 0.0636825668, 0.0511825668, 0.0636825668],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0010152455, 0.0010152455, 0.1010152455, 0.1010152455],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.0511825668, 0.0511825668, 0.0636825668, 0.0636825668],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 0.603873544619, rtol=1e-5)
assert_allclose(
grad[1:5],
[-0.0036235458, 0.0968966988, 0.0973867802, 0.0728492687],
rtol=1e-4)
assert_allclose(
grad[15:19],
[0.0262076293, 0.0640570655, 0.051891369, 0.0647118918],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0002012168, -0.0017716939, 0.0202682287, 0.0985633655],
rtol=1e-5)
assert_allclose(
grad[-4:], [0.0246493882, 0.0504098876, 0.030967057, 0.0630452251],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 0.624990535043, rtol=1e-5)
assert_allclose(
grad[1:5], [0.0014799603, 0.1012948868, 0.101109805, 0.0759247155],
rtol=1e-5)
assert_allclose(
grad[15:19],
[0.0285178333, 0.066295708, 0.0540735689, 0.0668514181],
rtol=1e-5)
assert_allclose(grad[19:23],
[0.0055522045, 0.00185136, 0.2222005993, 0.0468513936],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.1678131716, 0.0514809763, 0.2142578479, 0.0641662329],
rtol=1e-5)
def test_class(self):
import numpy as np
from kamrecsys.datasets import load_movielens_mini
from kamrecsys.tm.plsa_multi import EventScorePredictor
data = load_movielens_mini()
rcmdr = EventScorePredictor(tol=1e-8, k=2, random_state=1234)
self.assertDictEqual(
vars(rcmdr),
{
"k": 2,
"tol": 1e-8,
"maxiter": 100,
"alpha": 1.0,
"random_state": 1234,
"_rng": None,
"iid": None,
"eid": None,
"n_objects": None,
"n_otypes": 0,
"n_score_levels_": 0,
"i_loss_": np.inf,
"f_loss_": np.inf,
"n_iter_": 0,
"pZ_": None,
"pYgZ_": None,
"pRgZ_": None,
"pXgZ_": None,
"n_events_": 0,
"n_users_": 0,
"n_items_": 0,
"_q": None,
"score_levels_": None,
"use_expectation": True,
},
)
# import logging
# logging.getLogger('kamrecsys').addHandler(logging.StreamHandler())
rcmdr.fit(data)
self.assertAlmostEqual(rcmdr.i_loss_, 5.41836900049, delta=1e-5)
self.assertAlmostEqual(rcmdr.f_loss_, 5.17361298499, delta=1e-5)
assert_allclose(rcmdr.score_levels_, [1, 2, 3, 4, 5], rtol=1e-5)
# known user and item
self.assertAlmostEqual(rcmdr.predict((1, 7)), 3.64580117249, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((1, 9)), 3.6587422493, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 7)), 3.60707987724, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 9)), 3.62184516985, delta=1e-5)
# known user and unknown item
self.assertAlmostEqual(rcmdr.predict((1, 11)), 3.66032199689, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 12)), 3.62387542269, delta=1e-5)
# unknown user and known item
self.assertAlmostEqual(rcmdr.predict((3, 7)), 3.60821491793, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((11, 9)), 3.62304301551, delta=1e-5)
# unknown user and item
self.assertAlmostEqual(rcmdr.predict((3, 11)), 3.62507437787, delta=1e-5)
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7], [5, 9], [5, 11]])
assert_allclose(
rcmdr.predict(x),
[
3.64580117249,
3.6587422493,
3.66032199689,
3.60821491793,
3.62304301551,
3.62507437787,
3.60707987724,
3.62184516985,
3.62387542269,
],
rtol=1e-5,
)
rcmdr.use_expectation = False
assert_allclose(rcmdr.predict(x), [4.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0], rtol=1e-5)
def test_loss(self):
# setup
data = load_movielens_mini()
data.binarize_score(3)
rec = LogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
0.7665586528005817,
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
0.69314718055994495,
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
1.5567153484891236,
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
1.1832627694252338,
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
assert_allclose(rec.loss(rec._coef, ev, sc, n_objects),
3.6444649416252157,
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
rec = ImplicitLogisticPMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
n_objects = data.n_objects
ev = sparse.coo_matrix(
(np.ones(data.n_events, dtype=int),
(data.event[:, 0], data.event[:, 1])), shape=n_objects)
ev = ev.tocsr()
rec._init_coef(ev, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 29.54759073801552, rtol=1e-5)
assert_allclose(
grad[1:5],
[-1.3585150835, 6.1190706839, 3.9453452394, 4.1675236607],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-0.6724531834, 2.3829809039, 1.7510280667, 3.5386563872],
rtol=1e-5)
assert_allclose(
grad[19:23],
[-0.3332963261, -0.7966485582, 3.730424055, -1.8144337114],
rtol=1e-5)
assert_allclose(
grad[-4:],
[2.2257145558, 2.5006539103, 1.8384101853, -0.4847956693],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 10.0, atol=1e-5)
assert_allclose(
grad[1:5],
[-5., 3., 3., 1.],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-2., 1., 0., 1.],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0., 0., 0., 0.],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0., 0., 0., 0.],
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 49.46457192605721, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.0330714908, 8.0330714908, 8.0330714908, 6.0330714908],
rtol=1e-5)
assert_allclose(
grad[15:19],
[2.0464571926, 5.0464571926, 4.0464571926, 5.0464571926],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0330714908, 0.0330714908, 8.0330714908, 8.0330714908],
rtol=1e-5)
assert_allclose(
grad[-4:],
[4.0464571926, 4.0464571926, 5.0464571926, 5.0464571926],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 48.309883569540034, rtol=1e-5)
assert_allclose(
grad[1:5],
[-0.2898836641, 7.7469210915, 7.7813127826, 5.813497054],
rtol=1e-4)
assert_allclose(
grad[15:19],
[2.0195733081, 5.0427133869, 4.0646428559, 5.085469863],
rtol=1e-5)
assert_allclose(
grad[19:23],
[-0.0079767328, -0.1898836641, 1.5973842183, 7.8369210915],
rtol=1e-5)
assert_allclose(
grad[-4:],
[1.962321428, 3.9846428559, 2.4677349315, 4.995469863],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, n_objects)
assert_allclose(grad[0], 49.999242803470324, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.0798783079, 8.0698872383, 8.0598955089, 6.0499031689],
rtol=1e-5)
assert_allclose(
grad[15:19],
[2.1899451815, 5.204952933, 4.2199595873, 5.2349652997],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.2997319135, 0.0999606492, 17.6277516462, 3.6999633395],
rtol=1e-5)
assert_allclose(
grad[-4:],
[13.3258685412, 4.0799595873, 17.0428870887, 5.0899652997],
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -0.6389059157188512, rtol=1e-5)
assert_allclose(
grad[1:5],
[1.0788090999, 0.464982617, -0.7963940914, -0.0692397823],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-2.2818579721, -0.7857900494, -1.1561738739, 0.7205513543],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.1565786903, -2.9713197014, 0.4173437636, -0.0944053917],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.2978489043, 1.6373689599, 0.6585058144, -0.2955021786],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -115, rtol=1e-5)
assert_allclose(
grad[1:5],
[-36., -6., -7., -14.],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-22., -10., -18., -9.],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0., 0., 0., 0.],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0., 0., 0., 0.],
rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 35.0, rtol=1e-5)
assert_allclose(
grad[1:5],
[14.1, 4.1, 3.1, 6.1],
rtol=1e-5)
assert_allclose(
grad[15:19],
[8.1, 5.1, 2.1, 6.1],
rtol=1e-5)
assert_allclose(
grad[19:23],
[14.1, 14.1, 4.1, 4.1],
rtol=1e-5)
assert_allclose(
grad[-4:],
[2.1, 2.1, 6.1, 6.1],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 0.6000000000000028, rtol=1e-5)
assert_allclose(
grad[1:5],
[-0.5, 1.31, -0.28, 1.73],
rtol=1e-4)
assert_allclose(
grad[15:19],
[2.86, 2.27, -0.82, 3.69],
rtol=1e-5)
assert_allclose(
grad[19:23],
[-0.05, -0.4, 0.31, 1.4],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.48, -0.9, 1.77, 3.6],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 234.84160000000003, rtol=1e-5)
assert_allclose(
grad[1:5],
[78.28, 17.0564, 14.9512, 33.0668],
rtol=1e-5)
assert_allclose(
grad[15:19],
[49.89, 26.1354, 31.0472, 27.967],
rtol=1e-5)
assert_allclose(
grad[19:23],
[168.997, 37.31, 36.622072, 9.28216],
rtol=1e-5)
assert_allclose(
grad[-4:],
[101.688816, 30.9072, 86.899864, 27.822],
rtol=1e-5)
def test_filter_event(self):
from kamrecsys.data import EventWithScoreData
# load movie_lens
data = load_movielens_mini()
# filter events
filter_cond = np.arange(data.n_events) % 3 == 0
filtered_data = super(EventWithScoreData,
data).filter_event(filter_cond)
assert_array_equal(filtered_data.event[:, 0],
[1, 5, 3, 4, 0, 0, 0, 2, 2, 0])
assert_array_equal(filtered_data.event[:, 1],
[1, 3, 6, 5, 7, 6, 4, 0, 7, 2])
assert_array_equal(filtered_data.to_eid(0, filtered_data.event[:, 0]),
data.to_eid(0, data.event[filter_cond, 0]))
assert_array_equal(filtered_data.to_eid(1, filtered_data.event[:, 1]),
data.to_eid(1, data.event[filter_cond, 1]))
assert_array_equal(filtered_data.event_feature['timestamp'], [
875636053, 877889130, 891351328, 879362287, 878543541, 875072484,
889751712, 883599478, 883599205, 878542960
])
assert_array_equal(filtered_data.eid[0], [1, 5, 6, 7, 8, 10])
assert_array_equal(filtered_data.eid[1], [1, 2, 3, 4, 5, 7, 8, 9])
assert_equal(filtered_data.iid[0], {
1: 0,
5: 1,
6: 2,
7: 3,
8: 4,
10: 5
})
assert_equal(filtered_data.iid[1], {
1: 0,
2: 1,
3: 2,
4: 3,
5: 4,
7: 5,
8: 6,
9: 7
})
assert_equal(
filtered_data.feature[0]['zip'],
[u'85711', u'15213', u'98101', u'91344', u'05201', u'90703'])
assert_equal(filtered_data.feature[1]['name'], [
u'Toy Story (1995)', u'GoldenEye (1995)', u'Four Rooms (1995)',
u'Get Shorty (1995)', u'Copycat (1995)', u'Twelve Monkeys (1995)',
u'Babe (1995)', u'Dead Man Walking (1995)'
])
# dummy event data
data = EventData()
data.set_event(np.tile(np.arange(5), (2, 2)).T)
filtered_data = data.filter_event(
[True, False, True, True, False, False, True, True, False, False])
assert_equal(filtered_data.n_events, 5)
assert_array_equal(filtered_data.event,
[[0, 0], [2, 2], [3, 3], [1, 1], [2, 2]])
def test_class(self):
import numpy as np
from kamrecsys.datasets import load_movielens_mini
from kamrecsys.tm.plsa_multi import EventScorePredictor
data = load_movielens_mini()
rcmdr = EventScorePredictor(tol=1e-8, k=2, random_state=1234)
self.assertDictEqual(
vars(rcmdr), {
'k': 2,
'tol': 1e-8,
'maxiter': 100,
'alpha': 1.0,
'random_state': 1234,
'_rng': None,
'iid': None,
'eid': None,
'n_objects': None,
'n_otypes': 0,
'n_score_levels_': 0,
'i_loss_': np.inf,
'f_loss_': np.inf,
'n_iter_': 0,
'pZ_': None,
'pYgZ_': None,
'pRgZ_': None,
'pXgZ_': None,
'n_events_': 0,
'n_users_': 0,
'n_items_': 0,
'_q': None,
'score_levels_': None,
'use_expectation': True
})
# import logging
# logging.getLogger('kamrecsys').addHandler(logging.StreamHandler())
rcmdr.fit(data)
self.assertAlmostEqual(rcmdr.i_loss_, 5.41836900049, delta=1e-5)
self.assertAlmostEqual(rcmdr.f_loss_, 5.17361298499, delta=1e-5)
assert_allclose(rcmdr.score_levels_, [1, 2, 3, 4, 5], rtol=1e-5)
# known user and item
self.assertAlmostEqual(rcmdr.predict((1, 7)),
3.64580117249,
delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((1, 9)), 3.6587422493, delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 7)),
3.60707987724,
delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 9)),
3.62184516985,
delta=1e-5)
# known user and unknown item
self.assertAlmostEqual(rcmdr.predict((1, 11)),
3.66032199689,
delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((5, 12)),
3.62387542269,
delta=1e-5)
# unknown user and known item
self.assertAlmostEqual(rcmdr.predict((3, 7)),
3.60821491793,
delta=1e-5)
self.assertAlmostEqual(rcmdr.predict((11, 9)),
3.62304301551,
delta=1e-5)
# unknown user and item
self.assertAlmostEqual(rcmdr.predict((3, 11)),
3.62507437787,
delta=1e-5)
x = np.array([[1, 7], [1, 9], [1, 11], [3, 7], [3, 9], [3, 11], [5, 7],
[5, 9], [5, 11]])
assert_allclose(rcmdr.predict(x), [
3.64580117249, 3.6587422493, 3.66032199689, 3.60821491793,
3.62304301551, 3.62507437787, 3.60707987724, 3.62184516985,
3.62387542269
],
rtol=1e-5)
rcmdr.use_expectation = False
assert_allclose(rcmdr.predict(x), [4., 5., 5., 4., 4., 4., 4., 4., 4.],
rtol=1e-5)
def test_grad_loss(self):
# setup
data = load_movielens_mini()
rec = PMF(C=0.1, k=2, random_state=1234, tol=1e-03)
rec._rng = check_random_state(rec.random_state)
ev = data.event
sc = data.score
n_objects = data.n_objects
rec._init_coef(ev, sc, n_objects)
# set array's view
mu = rec._coef.view(rec._dt)['mu']
bu = rec._coef.view(rec._dt)['bu']
bi = rec._coef.view(rec._dt)['bi']
p = rec._coef.view(rec._dt)['p']
q = rec._coef.view(rec._dt)['q']
# initial parameters
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -0.0212968638573, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.0363059823, 0.0167339885, -0.0260526426, -0.0018141656],
rtol=1e-5)
assert_allclose(
grad[15:19],
[-0.0757162534, -0.02525473, -0.0393169069, 0.0252035637],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.0047896573, -0.0979586199, 0.012605792, -0.0028619178],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.0108829851, 0.0536257719, 0.0211630636, -0.0097388071],
rtol=1e-5)
# all zero
mu[0] = 0.0
bu[0][:] = 0.0
bi[0][:] = 0.0
p[0][:, :] = 0.0
q[0][:, :] = 0.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], -3.83333333333, rtol=1e-5)
assert_allclose(grad[1:5], [-1.2, -0.2, -0.2333333333, -0.4666666667],
rtol=1e-5)
assert_allclose(grad[15:19],
[-0.7333333333, -0.3333333333, -0.6, -0.3],
rtol=1e-5)
assert_allclose(grad[19:23], [0., 0., 0., 0.], rtol=1e-5)
assert_allclose(grad[-4:], [0., 0., 0., 0.], rtol=1e-5)
# all one
mu[0] = 1.0
bu[0][:] = 1.0
bi[0][:] = 1.0
p[0][:, :] = 1.0
q[0][:, :] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 1.16666666667, rtol=1e-5)
assert_allclose(
grad[1:5],
[0.4685185185, 0.1351851852, 0.1018518519, 0.2018518519],
rtol=1e-5)
assert_allclose(
grad[15:19],
[0.2685185185, 0.1685185185, 0.0685185185, 0.2018518519],
rtol=1e-5)
assert_allclose(
grad[19:23],
[0.4685185185, 0.4685185185, 0.1351851852, 0.1351851852],
rtol=1e-5)
assert_allclose(
grad[-4:],
[0.0685185185, 0.0685185185, 0.2018518519, 0.2018518519],
rtol=1e-5)
mu[0] = 1.0
bu[0][:] = np.arange(0.0, 0.8, 0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) + 1.0
p[0][:, 0] = 0.5
p[0][:, 1] = 1.0
q[0][:, 0] = 0.2
q[0][:, 1] = 1.0
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 0.02, rtol=1e-5)
assert_allclose(
grad[1:5],
[-0.0166666667, 0.0435185185, -0.0096296296, 0.0572222222],
rtol=1e-4)
assert_allclose(grad[15:19],
[0.092962963, 0.0731481481, -0.03, 0.1201851852],
rtol=1e-5)
assert_allclose(
grad[19:23],
[-0.0024074074, -0.0148148148, 0.0095925926, 0.0451851852],
rtol=1e-5)
assert_allclose(
grad[-4:],
[-0.0162962963, -0.0314814815, 0.0587037037, 0.1185185185],
rtol=1e-5)
mu[0] = 2.0
bu[0][:] = np.arange(0.8, 0.0, -0.1)
bi[0][:] = np.arange(0.0, 1.0, 0.1) * 1.5 + 1.0
p[0][:, 0] = np.arange(0.0, 0.8, 0.1) * 0.8 + 3
p[0][:, 1] = 1.0
q[0][:, 0] = np.arange(1.0, 0.0, -0.1) * 0.3 + 2
q[0][:, 1] = np.arange(0.0, 1.0, 0.1)
grad = rec.grad_loss(rec._coef, ev, sc, n_objects)
assert_allclose(grad[0], 7.82805333333, rtol=1e-5)
assert_allclose(
grad[1:5],
[2.6081481481, 0.5675096296, 0.4974844444, 1.1014859259],
rtol=1e-5)
assert_allclose(
grad[15:19],
[1.6601851852, 0.868142963, 1.0316474074, 0.9287518519],
rtol=1e-5)
assert_allclose(
grad[19:23],
[5.6287888889, 1.2421851852, 1.2161727704, 0.3079238519],
rtol=1e-5)
assert_allclose(
grad[-4:],
[3.3865753481, 1.0290548148, 2.8936547259, 0.9260666667],
rtol=1e-5)