def test_tf_bmf_batch_accuracy(tf_session):
from lenskit.algorithms import basic
from lenskit.algorithms import bias
import lenskit.crossfold as xf
from lenskit import batch
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
algo = lktf.BiasedMF(25,
damping=10,
batch_size=1024,
epochs=20,
rng_spec=42)
algo = basic.Fallback(algo, bias.Bias(damping=10))
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return batch.predict(algo, test)
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
preds = pd.concat(eval(train, test) for (train, test) in folds)
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.83, abs=0.025)
user_rmse = preds.groupby('user').apply(
lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(1.03, abs=0.05)
def test_als_implicit_batch_accuracy():
import lenskit.crossfold as xf
from lenskit import batch
from lenskit import topn
ratings = lktu.ml100k.load_ratings()
algo = als.ImplicitMF(25, iterations=20)
def eval(train, test):
_log.info('running training')
train['rating'] = train.rating.astype(np.float_)
algo.fit(train)
users = test.user.unique()
_log.info('testing %d users', len(users))
candidates = topn.UnratedCandidates(train)
recs = batch.recommend(algo, users, 100, candidates)
return recs
folds = list(xf.partition_users(ratings, 5, xf.SampleFrac(0.2)))
test = pd.concat(te for (tr, te) in folds)
recs = pd.concat(eval(train, test) for (train, test) in folds)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
_log.info('nDCG for users is %.4f', results.ndcg.mean())
assert results.ndcg.mean() > 0
def test_implicit_als_batch_accuracy():
import lenskit.crossfold as xf
from lenskit import batch, topn
ratings = lktu.ml100k.ratings
algo_t = ALS(25)
def eval(train, test):
_log.info('running training')
train['rating'] = train.rating.astype(np.float_)
algo = util.clone(algo_t)
algo.fit(train)
users = test.user.unique()
_log.info('testing %d users', len(users))
recs = batch.recommend(algo, users, 100)
return recs
folds = list(xf.partition_users(ratings, 5, xf.SampleFrac(0.2)))
test = pd.concat(f.test for f in folds)
recs = pd.concat(eval(train, test) for (train, test) in folds)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
dcg = results.ndcg
_log.info('nDCG for %d users is %.4f', len(dcg), dcg.mean())
assert dcg.mean() > 0
def test_als_batch_accuracy():
from lenskit.algorithms import basic
import lenskit.crossfold as xf
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
lu_algo = als.BiasedMF(25, iterations=20, damping=5, method='lu')
cd_algo = als.BiasedMF(25, iterations=25, damping=5, method='cd')
# algo = basic.Fallback(svd_algo, basic.Bias(damping=5))
def eval(train, test):
_log.info('training LU')
lu_algo.fit(train)
_log.info('training CD')
cd_algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return test.assign(lu_pred=lu_algo.predict(test), cd_pred=cd_algo.predict(test))
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
preds = pd.concat(eval(train, test) for (train, test) in folds)
preds['abs_diff'] = np.abs(preds.lu_pred - preds.cd_pred)
_log.info('predictions:\n%s', preds.sort_values('abs_diff', ascending=False))
_log.info('diff summary:\n%s', preds.abs_diff.describe())
lu_mae = pm.mae(preds.lu_pred, preds.rating)
assert lu_mae == approx(0.73, abs=0.025)
cd_mae = pm.mae(preds.cd_pred, preds.rating)
assert cd_mae == approx(0.73, abs=0.025)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.lu_pred, df.rating))
assert user_rmse.mean() == approx(0.91, abs=0.05)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.cd_pred, df.rating))
assert user_rmse.mean() == approx(0.91, abs=0.05)
def test_ii_batch_accuracy():
from lenskit.algorithms import basic
import lenskit.crossfold as xf
from lenskit import batch
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
ii_algo = knn.ItemItem(30)
algo = basic.Fallback(ii_algo, basic.Bias())
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return batch.predict(algo, test, n_jobs=4)
preds = pd.concat((eval(train, test)
for (train, test)
in xf.partition_users(ratings, 5, xf.SampleFrac(0.2))))
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.70, abs=0.025)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(0.90, abs=0.05)
def test_uu_implicit_batch_accuracy():
from lenskit import batch, topn
import lenskit.crossfold as xf
ratings = lktu.ml100k.ratings
algo = knn.UserUser(30, center=False, aggregate='sum')
folds = list(xf.partition_users(ratings, 5, xf.SampleFrac(0.2)))
all_test = pd.concat(f.test for f in folds)
rec_lists = []
for train, test in folds:
_log.info('running training')
rec_algo = Recommender.adapt(algo)
rec_algo.fit(train.loc[:, ['user', 'item']])
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(rec_algo, test.user.unique(), 100, n_jobs=2)
rec_lists.append(recs)
recs = pd.concat(rec_lists)
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, all_test)
user_dcg = results.ndcg
dcg = user_dcg.mean()
assert dcg >= 0.03
def test_ii_batch_recommend(ncpus):
import lenskit.crossfold as xf
from lenskit import batch, topn
if not os.path.exists('ml-100k/u.data'):
raise pytest.skip()
ratings = pd.read_csv('ml-100k/u.data', sep='\t', names=['user', 'item', 'rating', 'timestamp'])
def eval(train, test):
_log.info('running training')
algo = knn.ItemItem(30)
algo = Recommender.adapt(algo)
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(algo, test.user.unique(), 100, n_jobs=ncpus)
return recs
test_frames = []
recs = []
for train, test in xf.partition_users(ratings, 5, xf.SampleFrac(0.2)):
test_frames.append(test)
recs.append(eval(train, test))
test = pd.concat(test_frames)
recs = pd.concat(recs)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
dcg = results.ndcg
_log.info('nDCG for %d users is %f', len(dcg), dcg.mean())
assert dcg.mean() > 0.03
def split_dataset(ratings, user_fraction=.1):
"""Split a dataset in train/test data"""
n_users = len(ratings['user'].unique())
# There are many ways to separate a dataset in (train, test) data, here are two:
# - Row separation: the test set will contain users that the model knows.
# The performance of the model will be its ability to predict "new"
# tastes for a known user
# - User separation: the test set will contain users that the model has
# never encountered. The performance of the model will be its abiliy to
# predict new users behaviours considering the behaviour of other
# known users.
# see [lkpy documentation](https://lkpy.readthedocs.io/en/stable/crossfold.html)
# Here the sampling is as follow:
# - Sample test_fraction * n_total users
# - Randomly select half of their listenings for the test set
result = list(
xf.sample_users(ratings[['user', 'item', 'rating']],
partitions=1,
size=int(n_users * user_fraction),
method=xf.SampleFrac(.5)))[0]
print(f'n test users: {len(result.test["user"].unique())}')
return result.train, result.test
def test_bias_batch_predict(ncpus):
from lenskit.algorithms import bias
import lenskit.crossfold as xf
from lenskit import batch
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
algo = bias.Bias(damping=5)
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.predict(algo, test, n_jobs=ncpus)
return recs
preds = pd.concat(
(eval(train, test)
for (train,
test) in xf.partition_users(ratings, 5, xf.SampleFrac(0.2))))
_log.info('analyzing predictions')
rmse = pm.rmse(preds.prediction, preds.rating)
_log.info('RMSE is %f', rmse)
assert rmse == pytest.approx(0.95, abs=0.1)
def test_fsvd_batch_accuracy():
from lenskit.algorithms import basic
from lenskit.algorithms import bias
import lenskit.crossfold as xf
from lenskit import batch
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
svd_algo = svd.FunkSVD(25, 125, damping=10)
algo = basic.Fallback(svd_algo, bias.Bias(damping=10))
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return batch.predict(algo, test)
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
preds = pd.concat(eval(train, test) for (train, test) in folds)
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.74, abs=0.025)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(0.92, abs=0.05)
def test_bias_batch_recommend():
from lenskit.algorithms import basic
import lenskit.crossfold as xf
from lenskit import batch, topn
if not os.path.exists('ml-100k/u.data'):
raise pytest.skip()
ratings = pd.read_csv('ml-100k/u.data',
sep='\t',
names=['user', 'item', 'rating', 'timestamp'])
algo = basic.Bias(damping=5)
algo = TopN(algo)
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(algo, test.user.unique(), 100)
return recs
folds = list(xf.partition_users(ratings, 5, xf.SampleFrac(0.2)))
test = pd.concat(y for (x, y) in folds)
recs = pd.concat(eval(train, test) for (train, test) in folds)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
dcg = results.ndcg
_log.info('nDCG for %d users is %f (max=%f)', len(dcg), dcg.mean(),
dcg.max())
assert dcg.mean() > 0
def test_hpf_batch_accuracy():
import lenskit.crossfold as xf
from lenskit import batch, topn
import lenskit.metrics.topn as lm
ratings = lktu.ml100k.load_ratings()
algo = hpf.HPF(25)
def eval(train, test):
_log.info('running training')
train['rating'] = train.rating.astype(np.float_)
algo.fit(train)
users = test.user.unique()
_log.info('testing %d users', len(users))
candidates = topn.UnratedCandidates(train)
recs = batch.recommend(algo, users, 100, candidates, test)
return recs
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
recs = pd.concat(eval(train, test) for (train, test) in folds)
_log.info('analyzing recommendations')
dcg = recs.groupby('user').rating.apply(lm.dcg)
_log.info('dcg for users is %.4f', dcg.mean())
assert dcg.mean() > 0
def test_tf_bpr_batch_accuracy(tf_session):
from lenskit.algorithms import basic
import lenskit.crossfold as xf
from lenskit import batch, topn
ratings = lktu.ml100k.ratings
algo = lktf.BPR(20, batch_size=1024, epochs=20, rng_spec=42)
algo = Recommender.adapt(algo)
all_recs = []
all_test = []
for train, test in xf.partition_users(ratings, 5, xf.SampleFrac(0.2)):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(algo, np.unique(test.user), 50)
all_recs.append(recs)
all_test.append(test)
_log.info('analyzing results')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
rla.add_metric(topn.recip_rank)
scores = rla.compute(pd.concat(all_recs, ignore_index=True),
pd.concat(all_test, ignore_index=True),
include_missing=True)
scores.fillna(0, inplace=True)
_log.info('MRR: %f', scores['recip_rank'].mean())
_log.info('nDCG: %f', scores['ndcg'].mean())
assert scores['ndcg'].mean() > 0.1
def test_ii_batch_recommend(ncpus):
import lenskit.crossfold as xf
from lenskit import topn
ratings = lktu.ml100k.ratings
def eval(train, test):
_log.info('running training')
algo = knn.ItemItem(30)
algo = Recommender.adapt(algo)
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(algo, test.user.unique(), 100, n_jobs=ncpus)
return recs
test_frames = []
recs = []
for train, test in xf.partition_users(ratings, 5, xf.SampleFrac(0.2)):
test_frames.append(test)
recs.append(eval(train, test))
test = pd.concat(test_frames)
recs = pd.concat(recs)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
dcg = results.ndcg
_log.info('nDCG for %d users is %f', len(dcg), dcg.mean())
assert dcg.mean() > 0.03
def recommend(algo_wrappers, ratings):
all_recs = []
test_data = []
for train, test in xf.partition_users(ratings[['user', 'item', 'rating']],
5, xf.SampleFrac(0.2)):
test_data.append(test)
for algo_wrapper in algo_wrappers:
all_recs.append(do_recommend(algo_wrapper, train, test))
return all_recs, test_data
class LegMedLensKit():
def loadData():
ratings = pd.read_csv('/Users/josse/Desktop/ratings.dat',
sep='::',
names=['user', 'item', 'rating', 'timestamp'])
print(ratings.head())
return (ratings)
#print ("test")
ratings = loadData()
data_matrix = np.array(
ratings.pivot(index='item', columns='user', values='rating'))
print(data_matrix)
data_matrix_rev = np.nan_to_num(data_matrix)
print(data_matrix_rev)
algo_ii = knn.ItemItem(20)
algo_als = als.BiasedMF(50)
def eval(aname, algo, train, test):
print("test")
fittable = util.clone(algo)
fittable = Recommender.adapt(fittable)
fittable.fit(train)
users = test.user.unique()
# now we run the recommender
recs = batch.recommend(fittable, users, 100)
# add the algorithm name for analyzability
recs['Algorithm'] = aname
print("recs")
print(recs.head())
return recs
all_recs = []
test_data = []
for train, test in xf.partition_users(ratings[['user', 'item', 'rating']],
1, xf.SampleFrac(0.2)):
test_data.append(test)
#print(test.head(10))
all_recs.append(eval('ItemItem', algo_ii, train, test))
all_recs.append(eval('ALS', algo_als, train, test))
print("test2")
print(all_recs.head())
all_recs = pd.concat(all_recs, ignore_index=True)
print(all_recs.head())
test_data = pd.concat(test_data, ignore_index=True)
#print(test_data.head)
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(all_recs, test_data)
results.head()
results.groupby('Algorithm').ndcg.mean()
results.groupby('Algorithm').ndcg.mean().plot.bar()
def test_sample_frac():
ratings = lktu.ml_test.ratings
users = np.random.choice(ratings.user.unique(), 5, replace=False)
samp = xf.SampleFrac(0.2)
for u in users:
udf = ratings[ratings.user == u]
tst = samp(udf)
trn = udf.loc[udf.index.difference(tst.index), :]
assert len(tst) + len(trn) == len(udf)
assert len(tst) >= math.floor(len(udf) * 0.2)
assert len(tst) <= math.ceil(len(udf) * 0.2)
samp = xf.SampleFrac(0.5)
for u in users:
udf = ratings[ratings.user == u]
tst = samp(udf)
trn = udf.loc[udf.index.difference(tst.index), :]
assert len(tst) + len(trn) == len(udf)
assert len(tst) >= math.floor(len(udf) * 0.5)
assert len(tst) <= math.ceil(len(udf) * 0.5)
def run(self, strategy_context: RecommenderAlgorithmStrategyContext) -> np.ndarray:
data_set_source = strategy_context.data_set_source
data_frame_reader: DataFrameReaderStrategy = self.data_frame_reader_factory.create(data_set_source)
data_set: DataFrame = data_frame_reader.parse(DataFrameReaderStrategyContext(data_set_source))
partition = list(partition_users(data=data_set, partitions=1, method=crossfold.SampleFrac(0.2)))[0]
test, train = partition.test, partition.train
number_of_recommendations = strategy_context.number_of_recommendations
algorithm = Recommender.adapt(Bias())
trained_algorithm = algorithm.fit(train)
recommendations = lenskit.batch.recommend(trained_algorithm, test['user'].unique(), number_of_recommendations)
return recommendations.groupby('user')['item'].apply(lambda x: x).to_numpy().reshape(
(-1, number_of_recommendations))
def test_global_metric():
import lenskit.crossfold as xf
import lenskit.batch as batch
from lenskit.algorithms.bias import Bias
train, test = next(
xf.sample_users(lktu.ml_test.ratings, 1, 200, xf.SampleFrac(0.5)))
algo = Bias()
algo.fit(train)
preds = batch.predict(algo, test)
rmse = pm.global_metric(preds)
assert rmse == pm.rmse(preds.prediction, preds.rating)
mae = pm.global_metric(preds, metric=pm.mae)
assert mae == pm.mae(preds.prediction, preds.rating)
def test_tf_isvd(ml20m):
algo = lenskit_tf.IntegratedBiasMF(20)
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return batch.predict(algo, test)
folds = xf.sample_users(ml20m, 2, 5000, xf.SampleFrac(0.2))
preds = pd.concat(eval(train, test) for (train, test) in folds)
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.60, abs=0.025)
user_rmse = preds.groupby('user').apply(
lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(0.92, abs=0.05)
def test_uu_batch_accuracy():
from lenskit.algorithms import basic
import lenskit.crossfold as xf
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.ratings
uu_algo = knn.UserUser(30)
algo = basic.Fallback(uu_algo, basic.Bias())
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
preds = [__batch_eval((algo, train, test)) for (train, test) in folds]
preds = pd.concat(preds)
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.71, abs=0.028)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(0.91, abs=0.055)
def test_partition_users_frac():
ratings = lktu.ml_test.ratings
splits = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
splits = list(splits)
assert len(splits) == 5
ucounts = ratings.groupby('user').item.count()
uss = ucounts * 0.2
for s in splits:
tucs = s.test.groupby('user').item.count()
assert all(tucs >= uss.loc[tucs.index] - 1)
assert all(tucs <= uss.loc[tucs.index] + 1)
assert all(s.test.index.union(s.train.index) == ratings.index)
assert len(s.test) + len(s.train) == len(ratings)
# we have all users
users = ft.reduce(lambda us1, us2: us1 | us2,
(set(s.test.user) for s in splits))
assert len(users) == ratings.user.nunique()
assert users == set(ratings.user)
def test_user_metric():
import lenskit.crossfold as xf
import lenskit.batch as batch
from lenskit.algorithms.bias import Bias
train, test = next(
xf.sample_users(lktu.ml_test.ratings, 1, 200, xf.SampleFrac(0.5)))
algo = Bias()
algo.fit(train)
preds = batch.predict(algo, test)
rmse = pm.user_metric(preds)
u_rmse = preds.groupby('user').apply(
lambda df: pm.rmse(df.prediction, df.rating))
assert rmse == approx(u_rmse.mean())
mae = pm.user_metric(preds, metric=pm.mae)
u_mae = preds.groupby('user').apply(
lambda df: pm.mae(df.prediction, df.rating))
assert mae == approx(u_mae.mean())
def test_als_implicit_batch_accuracy():
import lenskit.crossfold as xf
from lenskit import batch
from lenskit import topn
ratings = lktu.ml100k.ratings
def eval(train, test):
train['rating'] = train.rating.astype(np.float_)
_log.info('training CG')
cg_algo = als.ImplicitMF(25, iterations=20, method='cg')
cg_algo = Recommender.adapt(cg_algo)
cg_algo.fit(train)
_log.info('training LU')
lu_algo = als.ImplicitMF(25, iterations=20, method='lu')
lu_algo = Recommender.adapt(lu_algo)
lu_algo.fit(train)
users = test.user.unique()
_log.info('testing %d users', len(users))
cg_recs = batch.recommend(cg_algo, users, 100, n_jobs=2)
lu_recs = batch.recommend(lu_algo, users, 100, n_jobs=2)
return pd.concat({
'CG': cg_recs,
'LU': lu_recs
}, names=['Method']).reset_index('Method')
folds = list(xf.partition_users(ratings, 5, xf.SampleFrac(0.2)))
test = pd.concat(te for (tr, te) in folds)
recs = pd.concat((eval(train, test) for (train, test) in folds),
ignore_index=True)
_log.info('analyzing recommendations')
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test)
results = results.groupby('Method')['ndcg'].mean()
_log.info('LU nDCG for users is %.4f', results.loc['LU'].mean())
_log.info('CG nDCG for users is %.4f', results.loc['CG'].mean())
assert all(results > 0.28)
assert results.loc['LU'] == approx(results.loc['CG'], rel=0.05)
def test_pop_batch_recommend(ncpus):
from lenskit.algorithms import basic
import lenskit.crossfold as xf
from lenskit import batch, topn
import lenskit.metrics.topn as lm
if not os.path.exists('ml-100k/u.data'):
raise pytest.skip()
ratings = pd.read_csv('ml-100k/u.data',
sep='\t',
names=['user', 'item', 'rating', 'timestamp'])
algo = basic.Popular()
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
cand_fun = topn.UnratedCandidates(train)
recs = batch.recommend(algo,
test.user.unique(),
100,
cand_fun,
test,
nprocs=ncpus)
return recs
recs = pd.concat(
(eval(train, test)
for (train,
test) in xf.partition_users(ratings, 5, xf.SampleFrac(0.2))))
_log.info('analyzing recommendations')
_log.info('have %d recs for good items', (recs.rating > 0).sum())
dcg = recs.groupby('user').rating.agg(lm.dcg)
_log.info('DCG for %d users is %f (max=%f)', len(dcg), dcg.mean(),
dcg.max())
assert dcg.mean() > 0
def test_ii_batch_recommend(ncpus):
import lenskit.crossfold as xf
from lenskit import batch, topn
import lenskit.metrics.topn as lm
if not os.path.exists('ml-100k/u.data'):
raise pytest.skip()
ratings = pd.read_csv('ml-100k/u.data',
sep='\t',
names=['user', 'item', 'rating', 'timestamp'])
algo = knn.ItemItem(30)
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
cand_fun = topn.UnratedCandidates(train)
recs = batch.recommend(algo,
test.user.unique(),
100,
cand_fun,
nprocs=ncpus)
# combine with test ratings for relevance data
res = pd.merge(recs, test, how='left', on=('user', 'item'))
# fill in missing 0s
res.loc[res.rating.isna(), 'rating'] = 0
return res
recs = pd.concat(
(eval(train, test)
for (train,
test) in xf.partition_users(ratings, 5, xf.SampleFrac(0.2))))
_log.info('analyzing recommendations')
dcg = recs.groupby('user').rating.apply(lm.dcg)
_log.info('DCG for %d users is %f', len(dcg), dcg.mean())
assert dcg.mean() > 0
def all_movie_recommends(ratings, optionList):
all_recs = []
test_data = []
#Declare algorithm models
basic_bias_model = basic.Bias()
knn_model = iknn.ItemItem(20)
knn_u_model = uknn.UserUser(20)
als_b_model = als.BiasedMF(50)
als_i_model = als.ImplicitMF(50)
funk_model = funksvd.FunkSVD(50)
for train, test in xf.partition_users(ratings[['user', 'item', 'rating']],
5, xf.SampleFrac(0.2)):
test_data.append(test)
for option in optionList:
if option == 1:
all_recs.append(
batch_eval('BasicBias', basic_bias_model, train, test))
if option == 2:
all_recs.append(batch_eval('ItemItem', knn_model, train, test))
if option == 3:
all_recs.append(
batch_eval('UserUser', knn_u_model, train, test))
if option == 4:
all_recs.append(
batch_eval('ALS-Biased', als_b_model, train, test))
if option == 5:
all_recs.append(
batch_eval('ALS-Implicit', als_i_model, train, test))
if option == 6:
all_recs.append(batch_eval('FunkSVD', funk_model, train, test))
all_recs = pd.concat(all_recs, ignore_index=True)
test_data = pd.concat(test_data, ignore_index=True)
return all_recs, test_data
def test_als_batch_accuracy():
from lenskit.algorithms import basic
import lenskit.crossfold as xf
import lenskit.metrics.predict as pm
ratings = lktu.ml100k.load_ratings()
svd_algo = als.BiasedMF(25, iterations=20, damping=5)
algo = basic.Fallback(svd_algo, basic.Bias(damping=5))
def eval(train, test):
_log.info('running training')
algo.fit(train)
_log.info('testing %d users', test.user.nunique())
return test.assign(prediction=algo.predict(test))
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
preds = pd.concat(eval(train, test) for (train, test) in folds)
mae = pm.mae(preds.prediction, preds.rating)
assert mae == approx(0.73, abs=0.025)
user_rmse = preds.groupby('user').apply(lambda df: pm.rmse(df.prediction, df.rating))
assert user_rmse.mean() == approx(0.91, abs=0.05)
def test_uu_implicit_batch_accuracy():
from lenskit import batch, topn
import lenskit.crossfold as xf
import lenskit.metrics.topn as lm
ratings = lktu.ml100k.load_ratings()
algo = knn.UserUser(30, center=False, aggregate='sum')
folds = xf.partition_users(ratings, 5, xf.SampleFrac(0.2))
rec_lists = []
for train, test in folds:
_log.info('running training')
algo.fit(train.loc[:, ['user', 'item']])
cands = topn.UnratedCandidates(train)
_log.info('testing %d users', test.user.nunique())
recs = batch.recommend(algo, test.user.unique(), 100, cands, test)
rec_lists.append(recs)
recs = pd.concat(rec_lists)
user_dcg = recs.groupby('user').rating.apply(lm.dcg)
dcg = user_dcg.mean()
assert dcg >= 0.1
def test_sample_users_frac():
ratings = lktu.ml_test.ratings
splits = xf.sample_users(ratings, 5, 100, xf.SampleFrac(0.2))
splits = list(splits)
assert len(splits) == 5
ucounts = ratings.groupby('user').item.count()
uss = ucounts * 0.2
for s in splits:
tucs = s.test.groupby('user').item.count()
assert len(tucs) == 100
assert all(tucs >= uss.loc[tucs.index] - 1)
assert all(tucs <= uss.loc[tucs.index] + 1)
assert all(s.test.index.union(s.train.index) == ratings.index)
assert len(s.test) + len(s.train) == len(ratings)
# no overlapping users
for s1, s2 in it.product(splits, splits):
if s1 is s2:
continue
us1 = s1.test.user.unique()
us2 = s2.test.user.unique()
assert len(np.intersect1d(us1, us2)) == 0
