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
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
def demo_recs():
"""
A demo set of train, test, and recommendation data.
"""
train, test = simple_test_pair(ml_test.ratings, f_rates=0.5)
users = test['user'].unique()
algo = PopScore()
algo = PlackettLuce(algo, rng_spec='user')
algo.fit(train)
recs = recommend(algo, users, 500)
return train, test, recs
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
def objective_fn(params: Dict[str, Any]):
algo = als.BiasedMF(
features=params["features"],
iterations=params["iteration"],
reg=0.1,
damping=5,
)
model = util.clone(algo)
model = Recommender.adapt(model)
model.fit(train_df)
recs = batch.recommend(model, test_users, recsize)
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test_df)
target_metric = -results.ndcg.mean()
return {"loss": target_metric, "status": STATUS_OK}
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
# See that the users are ranked differently in the matrix
print(modelAls.user_index_[0:10])
first = modelAls.user_index_[0]
print(type(first))
firstUser = np.array([first], np.int64)
# Get recommendation for a user
# As als.BiasedMF is not implementing Recommend
# We need to first adapt the model
# https://lkpy.lenskit.org/en/stable/interfaces.html#recommendation
# rec = Recommender.adapt(modelAls)
# recs = rec.recommend(first, 10) #Gives error
# Get 10 recommendations for a user (pandas dataframe)
recs = batch.recommend(modelAls, firstUser,
10, topn.UnratedCandidates(train), test)
print(recs)
# Get the first recommended item
firstRec = recs.iloc[0, 0]
firstRecScore = recs.iloc[0, 1]
print(firstRec)
# Get the explanation of the recommendation
# Get the index of the items
items = modelAls.item_index_
# Find the index of the first item
indexFirstRec = items.get_loc(firstRec)
# Get the feature values of the user
userProfile = modelAls.user_features_[0]
itemProfile = modelAls.item_features_[indexFirstRec]
def eval(aname, algo, train, test, n):
fittable = util.clone(algo)
fittable = Recommender.adapt(fittable)
fittable.fit(train)
# predict ratings
ratings_est = fittable.predict(test[['user', 'item']])
print(len(ratings_est))
print(len(test['rating']))
# now we run the recommender
users = test.user.unique()
recs = batch.recommend(fittable, users, n)
# add the algorithm name for analyzability
recs['Algorithm'] = aname
y_true = []
for i in range(len(recs)):
row = recs.iloc[i]
user_id = row['user']
item_id = row['item']
boolen_ls = (test['user'] == user_id)
chosen_rows = [i for i, x in enumerate(boolen_ls) if x]
focs_test = test.iloc[chosen_rows]
focs_test = focs_test[focs_test['rating'] >= 4]
if (item_id in focs_test['item']):
y_true.append('1')
else:
y_true.append('0')
def coverage(preds, items, num_items):
rec_item = []
for i in range(len(preds)):
# for beer
if preds[i] >= 4:
# for jester
# if preds[i] > 0:
rec_item.append(items[i])
return len(set(rec_item)) / num_items
def hit_rate(preds, labels, users, topk=10):
user_pred_dict = {}
hit_rates = []
for i in range(len(preds)):
if users[i] not in user_pred_dict:
user_pred_dict[users[i]] = []
user_pred_dict[users[i]].append((preds[i], labels[i]))
for user in user_pred_dict:
user_res = sorted(user_pred_dict[user],
key=lambda x: x[0])[-topk:]
hit_rates.append(
np.sum([int(x[1]) > 0 for x in user_res]) / topk)
return np.mean(hit_rates)
all_df = pd.concat([train, test])
item_count = len(all_df['item'].unique())
auc_score = roc_auc_score(y_true, recs['score'])
cov = coverage(recs['score'], recs['item'], item_count)
hit = hit_rate(recs['score'], recs['item'], recs['user'])
return score, cov, hit
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, nprocs=ncpus)
return recs
model = Recommender.adapt(model)
model.fit(train_df)
# COMMAND ----------
# have a subset of test users with equal distribution between longtail and shorthead preference
df1 = test_df.merge(user_pref.toPandas(),
"left",
left_on="user",
right_on="userId")
df2 = df1.query("longtail_pref >= 0.5").sample(n=250, random_state=123)
df3 = df1.query("longtail_pref < 0.5").sample(n=250, random_state=123)
test_users = pd.concat([df2, df3]).user.unique()
recs = batch.recommend(model, test_users, 100)
rla = topn.RecListAnalysis()
rla.add_metric(topn.ndcg)
results = rla.compute(recs, test_df)
print(f"NDCG: {results.ndcg.mean()}")
# COMMAND ----------
# MAGIC %md
# MAGIC Save the results of `recs`:
# COMMAND ----------