rtrec 0.2.4

A realtime recommendation system supporting online updates

rtrec: Realtime Recommendation Library in Python

A realtime recommendation system supporting online updates.

Highlights

• ❇️ Supporting online updates.
• ⚡️ Fast implementation (>=190k samples/sec training on laptop).
• ◍ efficient sparse data support.
• 🕑 decaying weights of user-item interactions based on recency.
• 🔄 Hybrid CF(Collaborative-Filtering)/CB(Content-based) model for cold-start users.
• experimental Rust implementation

Supported Recommendation Algorithims

• Sparse SLIM with time-weighted interactions.
• Factorization Machines using LightFM
• Hybrid model of SLIM (CF) and Factorization Machines (CB). Based on the number of user-item interactions, balances prediction weights between CF and CB.

Installation

# using pip
pip install rtrec
pip install rtrec[serving]

# using uv
uv add rtrec
uv sync --no-extra serving

uv add rtrec[serving]
uv sync --extra serving

Usage

Find usages in notebooks/examples.

Examples using Raw-level APIs

# 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"]

Examples using high level DataFrame APIs

# 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