def _setup_obs_handler(self):
ratings_df = pd.read_csv(rating_csv_path)
obs = ObservationsDF(ratings_df, uid_col=self.user_id_col, iid_col=self.item_id_col)
obs = obs.sample_observations(n_users=1000, n_items=1000)
self.state.train_obs, self.state.test_obs = obs.split_train_test(ratio=0.2, users_ratio=1.0)
# add some fake data for sanity tests
self.state.train_obs.df_obs = self._add_testing_obs_data(self.state.train_obs.df_obs)
def test_data(self):
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data(
movielens_dir)
ratings_df = pd.read_csv(rating_csv_path)
self.assertListEqual(list(ratings_df.columns),
['rating', 'timestamp', 'itemid', 'userid'])
self.assertEqual(len(ratings_df), 1000209)
users_df = pd.read_csv(users_csv_path)
self.assertListEqual(list(users_df.columns), [
'user_ind', 'gender', 'age', 'occupation', 'zipcode', 'index',
'occupation_name', 'userid'
])
self.assertEqual(len(users_df), 6040)
movies_df = pd.read_csv(movies_csv_path)
self.assertListEqual(list(movies_df.columns),
['item_ind', 'itemid', 'genres'])
self.assertEqual(len(movies_df), 3883)
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
obs = ObservationsDF(df_obs=ratings_df)
info = obs.data_info()
self.assertEqual(info['len'], 989539)
self.assertEqual(info['n_unique_items'], 3706)
self.assertEqual(info['n_unique_users'], 5796)
self.assertEqual(info['ratings_20_pctl'], 3.0)
self.assertEqual(info['ratings_80_pctl'], 5.0)
def test_data(self):
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data(
movielens_dir)
ratings_df = pd.read_csv(rating_csv_path)
self.assertListEqual(list(ratings_df.columns),
['rating', 'timestamp', 'itemid', 'userid'])
self.assertEqual(len(ratings_df), 1000209)
users_df = pd.read_csv(users_csv_path)
self.assertListEqual(list(users_df.columns), [
'gender', 'age', 'occupation', 'zipcode', 'index',
'occupation_name', 'userid'
])
self.assertEqual(len(users_df), 6040)
movies_df = pd.read_csv(movies_csv_path)
self.assertSetEqual(
set(movies_df.columns), {
'itemid', 'Adventure', 'FilmNoir', 'Comedy', 'SciFi',
'Fantasy', 'Crime', 'Mystery', 'Action', 'Thriller', 'Horror',
'Musical', 'Drama', 'Western', 'War', 'Animation', 'Romance',
'Childrens', 'Documentary'
})
self.assertEqual(len(movies_df), 3883)
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
obs = ObservationsDF(df_obs=ratings_df)
info = obs.data_info()
self.assertEqual(info['len'], 989539)
self.assertEqual(info['n_unique_items'], 3706)
self.assertEqual(info['n_unique_users'], 5796)
self.assertEqual(info['ratings_20_pctl'], 3.0)
self.assertEqual(info['ratings_80_pctl'], 5.0)
def fit(self, train_obs: ObservationsDF, *args, **kwargs):
factors_obs, reg_obs = train_obs.split_train_test(
ratio=self.stacking_split)
self._set_item_features_df(train_obs)
self._set_data(factors_obs)
self._fit_factorizer(factors_obs)
self._fit_regressor(reg_obs)
def test_splits(self):
ratings_df = pd.read_csv(rating_csv_path)
obs_params = dict(uid_col='userid',
iid_col='itemid',
timestamp_col='timestamp')
obs = ObservationsDF(ratings_df, **obs_params)
obs = obs.sample_observations(n_users=1000, n_items=1000)
self._split_tester(obs)
items_df = pd.read_csv(movies_csv_path)
obs_feat = ObsWithFeatures(df_obs=ratings_df,
df_items=items_df,
item_id_col='itemid',
**obs_params)
obs_feat = obs_feat.sample_observations(n_users=1000, n_items=1000)
self._split_tester(obs_feat)
def test_splits(self):
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
ratings_df = pd.read_csv(rating_csv_path)
obs = ObservationsDF(ratings_df,
uid_col='userid',
iid_col='itemid',
timestamp_col='timestamp')
obs = obs.sample_observations(n_users=1000, n_items=1000)
ratio = 0.2
# regular split
train_obs, test_obs = obs.split_train_test(ratio=ratio)
self._obs_split_data_check(obs, train_obs, test_obs)
self.state.train_obs, self.state.test_obs = train_obs, test_obs
# split for only some users
user_ratio = 0.2
train_obs, test_obs = obs.split_train_test(ratio=ratio,
users_ratio=user_ratio)
self._obs_split_data_check(obs, train_obs, test_obs)
post_split_ratio = test_obs.df_obs['userid'].nunique(
) / train_obs.df_obs['userid'].nunique()
self.assertAlmostEqual(user_ratio, post_split_ratio, places=1)
# split by timestamp
time_col = obs.timestamp_col
train_obs, test_obs = obs.split_train_test(ratio=ratio,
time_split_column=time_col)
self._obs_split_data_check(obs, train_obs, test_obs)
self.assertGreaterEqual(test_obs.df_obs[time_col].min(),
train_obs.df_obs[time_col].max())
"""
This is an example on datasets-1M demonstrating:
- More advanced fitting features: fit, evaluation, early stopping, hyper-param search
"""
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
import pandas as pd
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
from ml_recsys_tools.recommenders.lightfm_recommender import LightFMRecommender
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data()
ratings_df = pd.read_csv(rating_csv_path)
obs = ObservationsDF(ratings_df, uid_col='userid', iid_col='itemid')
train_obs, test_obs = obs.split_train_test(ratio=0.2)
lfm_rec = LightFMRecommender()
# train LightFM with early stopping and print evaluation results
lfm_rec.fit_with_early_stop(train_obs,
epochs_max=30,
epochs_step=1,
stop_patience=1,
valid_ratio=0.2,
metric='n-MRR@10',
refit_on_all=True)
print(
lfm_rec.eval_on_test_by_ranking(test_obs.df_obs, prefix='lfm early stop '))
# perform a hyperparameter search on LightFM recommender
space = lfm_rec.guess_search_space()
hp_results = lfm_rec.hyper_param_search(
"""
This is an example on datasets-1M demonstrating recommenders from spotlight library
"""
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
import pandas as pd
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data()
ratings_df = pd.read_csv(rating_csv_path)
obs = ObservationsDF(ratings_df, uid_col='userid', iid_col='itemid', timestamp_col='timestamp')
train_obs, test_obs = obs.split_train_test(ratio=0.2, time_split_column=obs.timestamp_col)
from ml_recsys_tools.recommenders.spotlight_recommenders import EmbeddingFactorsRecommender
emb_rec = EmbeddingFactorsRecommender(model_params=dict(loss='adaptive_hinge', n_iter=1))
# emb_rec.fit(train_obs)
emb_rec.fit_with_early_stop(train_obs, epochs_max=5, epochs_step=1)
print(emb_rec.eval_on_test_by_ranking(test_obs, prefix='implicit embeddings '))
# trying to reproduce this:
# https://github.com/maciejkula/spotlight/tree/master/examples/movielens_sequence
from ml_recsys_tools.recommenders.spotlight_recommenders import SequenceEmbeddingRecommender
seq_rec = SequenceEmbeddingRecommender(
model_params=dict(n_iter=15, embedding_dim=32, batch_size=32, learning_rate=0.01),
fit_params=dict(max_sequence_length=200, timestamp_col='timestamp'))
seq_rec.fit(train_obs)
# emb_rec.fit_with_early_stop(train_obs, epochs_max=30, epochs_step=3)
print(seq_rec.eval_on_test_by_ranking(test_obs, prefix='lstm ', include_train=False))
"""
using multiple test sets
"""
# dataset: download and prepare dataframes
import pandas as pd
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
from ml_recsys_tools.recommenders.lightfm_recommender import LightFMRecommender
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data()
# read the interactions dataframe and create a data handler object and split to train and test
ratings_df = pd.read_csv(rating_csv_path)
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF
obs = ObservationsDF(ratings_df)
train_obs, test_obs = obs.split_train_test(ratio=0.2, users_ratio=0.2)
def construct_multiple_test_sets(test_df, train_df):
# by user history - active and inactive users
user_hist_counts = train_df.userid.value_counts()
user_hist_counts.hist(bins=100, alpha=0.5)
active_users = user_hist_counts[user_hist_counts >= 300].index.tolist()
test_df_act_us = test_df[test_df.userid.isin(active_users)]
test_df_nonact_us = test_df[~test_df.userid.isin(active_users)]
# by item popularity- popular and unpopular items
item_hist_counts = train_df.itemid.value_counts()
item_hist_counts.hist(bins=100, alpha=0.5)
popular_items = item_hist_counts[item_hist_counts >= 1000].index.tolist()
