Project Name | Stars | Downloads | Repos Using This | Packages Using This | Most Recent Commit | Total Releases | Latest Release | Open Issues | License | Language |
---|---|---|---|---|---|---|---|---|---|---|
Implicit | 3,174 | 22 | 10 | 2 days ago | 43 | January 29, 2022 | 80 | mit | Python | |
Fast Python Collaborative Filtering for Implicit Feedback Datasets | ||||||||||
Librec | 3,113 | 9 months ago | 79 | other | Java | |||||
LibRec: A Leading Java Library for Recommender Systems, see | ||||||||||
Cornac | 714 | a month ago | 37 | February 19, 2022 | 15 | apache-2.0 | Python | |||
A Comparative Framework for Multimodal Recommender Systems | ||||||||||
Recommendation | 324 | 9 months ago | 19 | mit | Jupyter Notebook | |||||
Recommendation System using ML and DL | ||||||||||
Recsys_course_at_polimi | 307 | 5 months ago | 3 | agpl-3.0 | Jupyter Notebook | |||||
This is the official repository for the Recommender Systems course at Politecnico di Milano. | ||||||||||
Recommend | 269 | 5 years ago | 2 | Python | ||||||
recommendation system with python | ||||||||||
Cofactor | 155 | 6 years ago | apache-2.0 | Jupyter Notebook | ||||||
CoFactor: Regularizing Matrix Factorization with Item Co-occurrence | ||||||||||
Deep_matrix_factorization_models | 140 | 5 years ago | 6 | Python | ||||||
Implementation of the "Deep Matrix Factorization Models for Recommender Systems" | ||||||||||
Kitabe | 127 | 6 months ago | 2 | mit | HTML | |||||
Book Recommendation System built for Book Lovers📖. Simply Rate ⭐ some books and get immediate recommendations🤩 | ||||||||||
Simple_mf | 123 | a year ago | 2 | mit | Jupyter Notebook | |||||
Simple but Flexible Recommendation Engine in PyTorch |
Simple recommendatnion system implementation with Python
# clone repoisitory
git clone [email protected]:chyikwei/recommend.git
cd recommend
# install numpy & scipy
pip install -r requirements.txt
pip install .
A jupyter notbook that compares PMF and BPMF model can be found here.
To run BPMF with MovieLens 1M dataset:
First, download MovieLens 1M dataset and unzip it (data will be in ml-1m
folder).
Then run:
>>> import numpy as np
>>> from recommend.bpmf import BPMF
>>> from recommend.utils.evaluation import RMSE
>>> from recommend.utils.datasets import load_movielens_1m_ratings
# load user ratings
>>> ratings = load_movielens_1m_ratings('ml-1m/ratings.dat')
>>> n_user = max(ratings[:, 0])
>>> n_item = max(ratings[:, 1])
>>> ratings[:, (0, 1)] -= 1 # shift ids by 1 to let user_id & movie_id start from 0
# fit model
>>> bpmf = BPMF(n_user=n_user, n_item=n_item, n_feature=10,
max_rating=5., min_rating=1., seed=0).fit(ratings, n_iters=20)
>>> RMSE(bpmf.predict(ratings[:, :2]), ratings[:,2]) # training RMSE
0.79784331768263683
# predict ratings for user 0 and item 0 to 9:
>>> bpmf.predict(np.array([[0, i] for i in xrange(10)]))
array([ 4.35574067, 3.60580936, 3.77778456, 3.4479072 , 3.60901065,
4.29750917, 3.66302187, 4.43915423, 3.85788772, 4.02423073])
examples/
folder. The scripts will download MovieLens 1M dataset automatically, run PMF(BPMF) model and show training/validation RMSE.python setup.py test
or run test with coverage:
coverage run --source=recommend setup.py test
coverage report -m
pip uninstall recommend
Old version code can be found in v0.0.1
. It contains a Probabilistic Matrix Factorization model with theano implementation.
The previous version (0.2.1
) did not implement correctly MCMC sampling in the BPMF algorithm. In fact, at every timestep it computed the predictions basing on the current value of the feature matrices, and used it to estimate the RMSE. This has no meaning from the MCMC point of view, whose purpose is to sample the feature matrices from the correct distributions in order to estimate the integral through which the rating ditribution is computed.
Instead, the correct approach (see Eq. 10 in reference [2]) entails averaging the predictions at every time step to get a final prediction and compute the RMSE. Essentially, the predicted value itself does not depend only on the last extracted value for the feature matrices, but on the whole chain. Having modified this, the RMSE for both the train and test set with BPMF improves (you can see it in this notebook). (Thanks LoryPack's contribution!)