A realtime recommendation system supporting online updates.
- ❇️ Supporting online updates.
- ⚡️ Fast implementation (>=190k samples/sec training on laptop).
- ◍ efficient sparse data support.
- 🕑 decaying weights of user-item interactions based on recency.
experimental Rust implementation
- Sparse SLIM with time-weighted interactions.
- Factorization Machines using LightFM
pip install rtrecFind usages in notebooks/examples.
# Dataset consists of user, item, tstamp, rating
import time
current_unixtime = time.time()
interactions = [('user_1', 'item_1', current_unixtime, 5.0),
('user_2', 'item_2', current_unixtime, -2.0),
('user_2', 'item_1', current_unixtime, 3.0),
('user_2', 'item_4', current_unixtime, 3.0),
('user_1', 'item_3', current_unixtime, 4.0)]
# Fit SLIM model
from rtrec.models import SLIM
model = SLIM()
model.fit(interactions)
# can fit from streams using yield as follows:
def yield_interactions():
for interaction in interactions:
yield interaction
model.fit(yield_interactions())
# Recommend top-5 items for a user
recommendations = model.recommend('user_1', top_k=5)
assert recommendations == ["item_4", "item_2"]# load dataset
from rtrec.experiments.datasets import load_dataset
df = load_dataset(name='movielens_1m')
# Split data set by temporal user split
from rtrec.experiments.split import temporal_user_split
train_df, test_df = temporal_user_split(df)
# Initialize SLIM model with custom options
from rtrec.recommender import Recommender
from rtrec.models import SLIM
model = SLIM(min_value=0, max_value=15, decay_in_days=180, nn_feature_selection=50)
recommender = Recommender(model)
# Bulk fit
recommender.bulk_fit(train_df)
# Partial fit
from rtrec.experiments.split import temporal_split
test_df1, test_df2 = temporal_split(test_df, test_frac=0.5)
recommender.fit(test_df1, update_interaction=True, parallel=True)
# Evaluation
metrics = recommender.evaluate(test_df2, recommend_size=10, filter_interacted=True)
print(metrics)
# User to Item Recommendation
recommended = recommender.recommend(user=10, top_k=10, filter_interacted=True)
assert len(recommended) == 10
# Item to Item recommendation
similar_items = recommender.similar_items(query_items=[3,10], top_k=5)
assert len(similar_items) == 2