def test_gym_environment_builder(self):
env = movie_lens.create_gym_environment(self.env_config)
env.seed(100)
env.reset()
# Recommend some manual slates and check that the observations are as
# expected.
for slate in [[0], [0], [2]]:
observation, _, _, _ = env.step(slate)
for field in ['doc', 'response', 'user']:
self.assertIn(observation[field], env.observation_space.spaces[field])
def test_user_order_is_consistent(self):
self.env.reset_sampler()
first_list = []
for _ in range(100):
observation = self.env.reset()
first_list.append(observation['user']['user_id'])
self.env.reset_sampler()
other_list = []
for _ in range(100):
observation = self.env.reset()
other_list.append(observation['user']['user_id'])
self.assertEqual(first_list, other_list)
# Also check that changing the seed creates a new ordering.
config = copy.deepcopy(self.env_config)
config.seeds.user_sampler += 1
env = movie_lens.create_gym_environment(config)
other_list = []
for _ in range(100):
observation = env.reset()
other_list.append(observation['user']['user_id'])
self.assertNotEqual(first_list, other_list)
def generate_data(env_config,
slate_type,
n_samples,
seed,
intercept,
rating_coef,
div_seek_coef,
diversity_coef,
recommender_agent=None,
shift=0.,
user_pool='train'):
"""Generate synthetic data.
Args:
env_config: Environment Configuration.
slate_type: Type of slate, should be one of ['all', 'top1genre',
'top2genre', 'test'].
n_samples: Number of samples to generate.
seed: Random seed.
intercept: Intercept of the ground truth logReg model P(Y).
rating_coef: Coefficient for rating in the ground truth model.
div_seek_coef: Coefficient for diversitySeeker in the ground truth model.
diversity_coef: Coefficient for diversity in the ground truth model.
recommender_agent: If None, will make recommendations uniformly at random.
shift: Increase the rating (prior to clipping) by this amount for each genre
in [Comedy, Drama, Childrens] and decrease by this amount for each genre
in [Sci-Fi, Fantasy, War]'))
user_pool: Train/Eval/Test pool of users.
Returns:
Pandas DataFrame containing the samples.
"""
if not env_config.embeddings_path.endswith(('.json', '.pkl')):
raise ValueError('Embedding path should end in .json or .pkl')
if not env_config.genre_history_path.endswith(('.json', '.pkl')):
raise ValueError('Genre history path should end in .json or .pkl')
if (env_config.user_config.accept_prob < 0
or env_config.user_config.accept_prob > 1):
raise ValueError('Accept probability should be in [0, 1]')
if slate_type == 'all':
top_movie_slate = np.arange(utils.NUM_MOVIES)
elif slate_type == 'top1genre':
# Most popular movie of each genre, with de-duplication
top_movie_slate = utils.find_top_movies_per_genre(
movies_per_genre=1, data_path=env_config.data_dir)
elif slate_type == 'top2genre':
# Top 2 most popular movies of each genre, with de-duplication
top_movie_slate = utils.find_top_movies_per_genre(
movies_per_genre=2, data_path=env_config.data_dir)
elif slate_type == 'top20':
# Top 20 most popular movies
top_movie_slate = utils.find_top_movies_overall(
num_movies=20, data_path=env_config.data_dir)
elif slate_type == 'test':
top_movie_slate = np.arange(5)
else:
raise ValueError('Slate type not recognized')
env_config.slate_size = len(top_movie_slate)
# Comedy and drama are the most commonly watched (and therefore least
# "diverse") genres
pos_shift = ['Comedy', 'Drama', 'Children\'s']
# These happen to be more "diverse" genres, and in our top1 from each
# genre tends to hit Star Wars movies in particular.
neg_shift = ['Sci-Fi', 'Fantasy', 'War']
genre_shift = [0.] * len(utils.GENRES)
for genre in pos_shift:
genre_shift[utils.GENRE_MAP[genre]] = shift
for genre in neg_shift:
genre_shift[utils.GENRE_MAP[genre]] = -shift
env_config.genre_shift = genre_shift
# TODO(moberst): All randomization should be through rngs, but currently the
# multinomial choice model (imported from RecSim) does not accept an rng,
# so this is a work-around to ensure consistent user choices.
np.random.seed(seed)
rng = np.random.default_rng(seed)
if recommender_agent is None:
recommender_agent = agent.RandomAgent(movies=top_movie_slate)
ml_env = movie_lens_simulator.create_gym_environment(env_config)
ml_env.environment.set_active_pool(user_pool)
res = []
for i in range(n_samples):
if i % 100 == 0:
logging.info('Iteration: %d / %d', i, n_samples)
# Generate data, one row per user interaction
row = {}
initial_obs = ml_env.reset()
user_id = initial_obs['user']['user_id']
row['user'] = user_id
row['diversity_seeker'] = (
ml_env.environment.user_model._user_state.diversity_seeking) # pylint: disable=protected-access
slate = recommender_agent.make_recommendation(user_id)
obs, _, _, _ = ml_env.step(slate)
[response] = obs['response']
row['rec'] = slate[0]
row['watched'] = response['doc_id']
row['rating'] = response['rating']
row['diversity'] = response['diversity']
res.append(row)
df = pd.DataFrame.from_dict(res).sort_values('user')
logits = intercept + df['rating'] * rating_coef + df[
'diversity_seeker'] * div_seek_coef + df['diversity'] * diversity_coef
# Calculate probability of the long-term reward (p_ltr) and sample it.
df['p_ltr'] = scipy.special.expit(logits)
df['ltr'] = rng.binomial(1, df['p_ltr'])
return df
