def test_bias_item_predict():
algo = Bias(users=False)
algo.fit(simple_df)
p = algo.predict_for_user(10, [1, 2, 3])
assert len(p) == 3
assert p.values == approx((algo.item_offsets_ + algo.mean_).values)
def test_sweep_norecs(tmp_path):
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path, recommend=None)
ratings = ml_test.ratings
folds = xf.partition_users(ratings, 5, xf.SampleN(5))
sweep.add_datasets(folds, DataSet='ml-small')
sweep.add_algorithms([Bias(damping=0), Bias(damping=5), Bias(damping=10)],
attrs=['damping'])
sweep.add_algorithms(Popular())
try:
sweep.run()
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert (work / 'runs.csv').exists()
assert (work / 'runs.parquet').exists()
assert (work / 'predictions.parquet').exists()
assert not (work / 'recommendations.parquet').exists()
runs = pd.read_parquet(work / 'runs.parquet')
# 4 algorithms by 5 partitions
assert len(runs) == 20
assert all(np.sort(runs.AlgoClass.unique()) == ['Bias', 'Popular'])
bias_runs = runs[runs.AlgoClass == 'Bias']
assert all(bias_runs.damping.notna())
pop_runs = runs[runs.AlgoClass == 'Popular']
assert all(pop_runs.damping.isna())
preds = pd.read_parquet(work / 'predictions.parquet')
assert all(preds.RunId.isin(bias_runs.RunId))
def test_sweep_filenames(tmp_path):
tmp_path = norm_path(tmp_path)
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path)
ratings = ml_pandas.renamed.ratings
folds = []
for part, (train, test) in enumerate(xf.partition_users(ratings, 2, xf.SampleN(5))):
trfn = work / 'p{}-train.csv'.format(part)
tefn = work / 'p{}-test.csv'.format(part)
train.to_csv(trfn)
test.to_csv(tefn)
folds.append((trfn, tefn))
sweep.add_datasets(folds, DataSet='ml-small')
sweep.add_algorithms([Bias(damping=0), Bias(damping=5), Bias(damping=10)],
attrs=['damping'])
sweep.add_algorithms(Popular())
try:
sweep.run()
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert (work / 'runs.csv').exists()
assert (work / 'runs.parquet').exists()
assert (work / 'predictions.parquet').exists()
assert (work / 'recommendations.parquet').exists()
runs = pd.read_parquet(work / 'runs.parquet')
# 4 algorithms by 2 partitions
assert len(runs) == 8
def test_bias_global_predict():
algo = Bias(items=False, users=False)
algo.fit(simple_df)
p = algo.predict_for_user(10, [1, 2, 3])
assert len(p) == 3
assert (p == algo.mean_).all()
assert p.values == approx(algo.mean_)
def test_bias_predict_unknown_user():
algo = Bias()
algo.fit(simple_df)
p = algo.predict_for_user(15, [1, 3])
assert len(p) == 2
assert p.values == approx(
(algo.item_offsets_.loc[[1, 3]] + algo.mean_).values)
def test_bias_transform():
algo = Bias()
ratings = ml_test.ratings
normed = algo.fit_transform(ratings)
assert all(normed['user'] == ratings['user'])
assert all(normed['item'] == ratings['item'])
denorm = algo.inverse_transform(normed)
assert approx(denorm['rating'] == ratings['rating'], 1.0e-6)
def test_bias_predict_unknown_item():
algo = Bias()
algo.fit(simple_df)
p = algo.predict_for_user(10, [1, 3, 4])
assert len(p) == 3
intended = algo.item_offsets_.loc[[1, 3]] + algo.mean_ + 0.25
assert p.loc[[1, 3]].values == approx(intended.values)
assert p.loc[4] == approx(algo.mean_ + 0.25)
def test_bias_no_item():
algo = Bias(items=False)
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is None
assert algo.user_offsets_ is not None
assert algo.user_offsets_.index.name == 'user'
assert set(algo.user_offsets_.index) == set([10, 12, 13])
assert algo.user_offsets_.loc[[10, 12, 13]].values == approx(
np.array([1.0, -0.5, -1.5]))
def test_bias_user_damp():
algo = Bias(items=False, damping=5)
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is None
assert algo.user_offsets_ is not None
assert algo.user_offsets_.index.name == 'user'
assert set(algo.user_offsets_.index) == set([10, 12, 13])
assert algo.user_offsets_.loc[[10, 12, 13]].values == \
approx(np.array([0.2857, -0.08333, -0.25]), abs=1.0e-4)
def test_sweep_combine(tmp_path):
tmp_path = norm_path(tmp_path)
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path, combine=False)
ratings = ml_pandas.renamed.ratings
sweep.add_datasets(lambda: xf.partition_users(ratings, 5, xf.SampleN(5)), name='ml-small')
sweep.add_algorithms([Bias(damping=0), Bias(damping=5)],
attrs=['damping'])
sweep.add_algorithms(Popular())
sweep.persist_data()
for i in range(1, 6):
assert (work / 'ds{}-train.parquet'.format(i)).exists()
assert (work / 'ds{}-test.parquet'.format(i)).exists()
for ds, cf, dsa in sweep.datasets:
assert isinstance(ds, tuple)
train, test = ds
assert isinstance(train, pathlib.Path)
assert isinstance(test, pathlib.Path)
assert sweep.run_count() == 5 * 3
try:
sweep.run()
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert not (work / 'runs.csv').exists()
assert not (work / 'runs.parquet').exists()
assert not (work / 'predictions.parquet').exists()
assert not (work / 'recommendations.parquet').exists()
for i, (ds, a) in enumerate(sweep._flat_runs()):
run = i + 1
assert (work / 'run-{}.json'.format(run)).exists()
if isinstance(a.algorithm, Predictor):
assert (work / 'predictions-{}.parquet'.format(run)).exists()
assert (work / 'recommendations-{}.parquet'.format(run)).exists()
sweep.collect_results()
assert (work / 'runs.csv').exists()
assert (work / 'runs.parquet').exists()
assert (work / 'predictions.parquet').exists()
assert (work / 'recommendations.parquet').exists()
runs = pd.read_parquet(work / 'runs.parquet')
assert len(runs) == 5 * 3
def test_bias_clone():
algo = Bias()
algo.fit(simple_df)
params = algo.get_params()
assert sorted(params.keys()) == ['damping', 'items', 'users']
a2 = lku.clone(algo)
assert a2 is not algo
assert getattr(a2, 'mean_', None) is None
assert getattr(a2, 'item_offsets_', None) is None
assert getattr(a2, 'user_offsets_', None) is None
def test_bias_no_user():
algo = Bias(users=False)
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is not None
assert algo.item_offsets_.index.name == 'item'
assert set(algo.item_offsets_.index) == set([1, 2, 3])
assert algo.item_offsets_.loc[1:3].values == approx(
np.array([0, 1.5, -1.5]))
assert algo.user_offsets_ is None
def test_bias_check_arguments():
# negative damping is not allowed
with raises(ValueError):
Bias(damping=-1)
# negative user damping not allowed
with raises(ValueError):
Bias(damping=(-1, 5))
# negative user damping not allowed
with raises(ValueError):
Bias(damping=(5, -1))
def test_bias_transform_indexes():
algo = Bias()
ratings = ml_test.ratings
normed = algo.fit_transform(ratings, indexes=True)
assert all(normed['user'] == ratings['user'])
assert all(normed['item'] == ratings['item'])
assert all(normed['uidx'] == algo.user_offsets_.index.get_indexer(
ratings['user']))
assert all(normed['iidx'] == algo.item_offsets_.index.get_indexer(
ratings['item']))
denorm = algo.inverse_transform(normed)
assert approx(denorm['rating'] == ratings['rating'], 1.0e-6)
def test_bias_new_user_predict():
algo = Bias()
algo.fit(simple_df)
ratings = pd.DataFrame({'item': [1, 2, 3], 'rating': [1.5, 2.5, 3.5]})
ratings = ratings.set_index('item').rating
p = algo.predict_for_user(None, [1, 3], ratings=ratings)
offs = ratings - algo.mean_ - algo.item_offsets_
umean = offs.mean()
_log.info('user mean is %f', umean)
assert len(p) == 2
assert p.values == approx(
(algo.mean_ + algo.item_offsets_ + umean).loc[[1, 3]].values)
def test_bias_batch_recommend(ml_folds: MLFolds, ncpus):
algo = Bias(damping=5)
algo = TopN(algo)
recs = ml_folds.eval_all(algo, nprocs=ncpus)
ml_folds.check_positive_ndcg(recs)
def test_sweep_oneshot(tmp_path):
tmp_path = norm_path(tmp_path)
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path, combine=False)
ratings = ml_pandas.renamed.ratings
sweep.add_datasets(lambda: xf.partition_users(ratings, 5, xf.SampleN(5)), name='ml-small')
sweep.add_algorithms(Bias(damping=5))
try:
sweep.run(3)
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert not (work / 'runs.csv').exists()
assert not (work / 'runs.parquet').exists()
assert not (work / 'predictions.parquet').exists()
assert not (work / 'recommendations.parquet').exists()
assert (work / 'run-3.json').exists()
assert (work / 'predictions-3.parquet').exists()
assert (work / 'recommendations-3.parquet').exists()
with (work / 'run-3.json').open() as f:
run = json.load(f)
assert run['RunId'] == 3
def test_bias_batch_recommend_threads(ml_folds: MLFolds):
algo = Bias(damping=5)
algo = TopN(algo)
with joblib.parallel_backend('threading', n_jobs=2):
recs = ml_folds.eval_all(algo)
ml_folds.check_positive_ndcg(recs)
def test_sweep_nopreds(tmp_path):
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path, eval_n_jobs=1)
ratings = ml_test.ratings
folds = [(train, test.drop(columns=['rating']))
for (train, test) in xf.partition_users(ratings, 5, xf.SampleN(5))
]
sweep.add_datasets(folds, DataSet='ml-small')
sweep.add_algorithms(Popular())
sweep.add_algorithms(Bias(damping=0))
try:
sweep.run()
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert (work / 'runs.csv').exists()
assert (work / 'runs.parquet').exists()
assert not (work / 'predictions.parquet').exists()
assert (work / 'recommendations.parquet').exists()
runs = pd.read_parquet(work / 'runs.parquet')
# 2 algorithms by 5 partitions
assert len(runs) == 10
assert all(np.sort(runs.AlgoClass.unique()) == ['Bias', 'Popular'])
bias_runs = runs[runs.AlgoClass == 'Bias']
recs = pd.read_parquet(work / 'recommendations.parquet')
assert all(recs.RunId.isin(runs.RunId))
assert recs['score'].dtype == np.float64
def test_sweep_save(tmp_path):
tmp_path = norm_path(tmp_path)
work = pathlib.Path(tmp_path)
sweep = batch.MultiEval(tmp_path)
ratings = ml_pandas.renamed.ratings
sweep.add_datasets(lambda: xf.partition_users(ratings, 5, xf.SampleN(5)), name='ml-small')
sweep.add_algorithms(Bias(damping=5))
sweep.persist_data()
pf = work / 'sweep.dat'
with pf.open('wb') as f:
pickle.dump(sweep, f)
with pf.open('rb') as f:
sweep = pickle.load(f)
try:
sweep.run()
finally:
if (work / 'runs.csv').exists():
runs = pd.read_csv(work / 'runs.csv')
print(runs)
assert (work / 'runs.csv').exists()
assert (work / 'runs.parquet').exists()
assert (work / 'predictions.parquet').exists()
assert (work / 'recommendations.parquet').exists()
runs = pd.read_parquet(work / 'runs.parquet')
# 1 algorithms by 5 partitions
assert len(runs) == 5
def test_bias_full():
algo = Bias()
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is not None
assert algo.item_offsets_.index.name == 'item'
assert set(algo.item_offsets_.index) == set([1, 2, 3])
assert algo.item_offsets_.loc[1:3].values == approx(
np.array([0, 1.5, -1.5]))
assert algo.user_offsets_ is not None
assert algo.user_offsets_.index.name == 'user'
assert set(algo.user_offsets_.index) == set([10, 12, 13])
assert algo.user_offsets_.loc[[10, 12, 13]].values == approx(
np.array([0.25, -0.5, 0]))
def test_bias_separate_damping():
algo = Bias(damping=(5, 10))
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is not None
assert algo.item_offsets_.index.name == 'item'
assert set(algo.item_offsets_.index) == set([1, 2, 3])
assert algo.item_offsets_.loc[1:3].values == \
approx(np.array([0, 0.136364, -0.13636]), abs=1.0e-4)
assert algo.user_offsets_ is not None
assert algo.user_offsets_.index.name == 'user'
assert set(algo.user_offsets_.index) == set([10, 12, 13])
assert algo.user_offsets_.loc[[10, 12, 13]].values == \
approx(np.array([0.266234, -0.08333, -0.22727]), abs=1.0e-4)
def test_bias_damped():
algo = Bias(damping=5)
algo.fit(simple_df)
assert algo.mean_ == approx(3.5)
assert algo.item_offsets_ is not None
assert algo.item_offsets_.index.name == 'item'
assert set(algo.item_offsets_.index) == set([1, 2, 3])
assert algo.item_offsets_.loc[1:3].values == approx(
np.array([0, 0.25, -0.25]))
assert algo.user_offsets_ is not None
assert algo.user_offsets_.index.name == 'user'
assert set(algo.user_offsets_.index) == set([10, 12, 13])
assert algo.user_offsets_.loc[[10, 12, 13]].values == \
approx(np.array([0.25, -00.08333, -0.20833]), abs=1.0e-4)
def test_bias_batch_recommend(ml_folds: MLFolds, ncpus, isolate):
algo = Bias(damping=5)
algo = TopN(algo)
ml_folds.isolate = isolate
recs = ml_folds.eval_all(algo, n_jobs=ncpus)
ml_folds.check_positive_ndcg(recs)
def _build_predict(ratings, fold):
algo = Fallback(knn.ItemItem(20), Bias(5))
train = ratings[ratings['partition'] != fold]
algo.fit(train)
test = ratings[ratings['partition'] == fold]
preds = batch.predict(algo, test, n_jobs=1)
return preds
def __init__(self,
nnbrs,
min_nbrs=1,
min_sim=0,
center=True,
aggregate='weighted-average'):
algo = UserUser(nnbrs, min_nbrs, min_sim, center, aggregate)
fallback = Bias()
Fallback.__init__(self, [algo, fallback])
def test_bias_train_ml_ratings():
algo = Bias()
ratings = ml_test.ratings
algo.fit(ratings)
assert algo.mean_ == approx(ratings.rating.mean())
imeans_data = ratings.groupby('item').rating.mean()
imeans_algo = algo.item_offsets_ + algo.mean_
ares, data = imeans_algo.align(imeans_data)
assert ares.values == approx(data.values)
urates = ratings.set_index('user').loc[2].set_index('item').rating
umean = (urates - imeans_data[urates.index]).mean()
p = algo.predict_for_user(2, [10, 11, -1])
assert len(p) == 3
assert p.iloc[0] == approx(imeans_data.loc[10] + umean)
assert p.iloc[1] == approx(imeans_data.loc[11] + umean)
assert p.iloc[2] == approx(ratings.rating.mean() + umean)
def test_bias_batch_recommend_threads(ml_folds: MLFolds):
algo = Bias(damping=5)
algo = TopN(algo)
with joblib.parallel_backend('threading', n_jobs=2):
recs = ml_folds.eval_all(algo, dask=True)
assert isinstance(recs, ddf.DataFrame)
ml_folds.check_positive_ndcg(recs)
def test_bias_batch_recommend_dask(ml_folds: MLFolds):
algo = Bias(damping=5)
algo = TopN(algo)
with closing(distributed.Client()), joblib.parallel_backend('dask'):
recs = ml_folds.eval_all(algo, dask=True)
assert isinstance(recs, ddf.DataFrame)
ml_folds.check_positive_ndcg(recs)
def __init__(self,
nnbrs,
min_nbrs=1,
min_sim=1e-06,
save_nbrs=None,
center=True,
aggregate='weighted-average'):
algo = ItemItem(nnbrs, min_nbrs, min_sim, save_nbrs, center, aggregate)
fallback = Bias()
Fallback.__init__(self, [algo, fallback])