matrix-factorization

Goal

Understand matrix factorization and singular value decomposition (SVD), and how this can be used to create a recommender system. Also, understand why this approach could be better than using the user-based or item-based collaborative filtering.

Matrix Factorization 101

Factorize a matrix to find out two matrices such that when you multiply them, you will get back the original matrix.
Use to discover latent features underlying the interactions between two different kinds of entities.
Use to predict ratings (sparse values, or zeros 0 in matrices). Most of the time, the matrix will be sparse because not all users will rate an item. With matrix factorization, we can predict the rating that the user will give on items they have not rated yet.

Input

Sample input for the algorithm:

Output

Sample output for the algorithm:

TODO

Run it parallel
Hashing the string and store it instead of the whole object (?)

Keywords

SVD (Singular Value Decomposition) - TODO: Definition

Reference

Mining of Massive Datasets. This book is available for free online and contains in-depth chapters on Singular Value Decomposition.

Matrix factorization

what

is matrix factorization
is the input
is the output why
use matrix factorization when
to use it
not to use it
does it work/does it not work who how
to calculate the matrix factorization
to use it for recommendation

Factorizing a matrix

We have a rating matrix R with n rows and m columns.
The columns is for user
The rows is for items
So we have a n x m matrix (read n by m).
We can decompose R into U x V
U will be n x d matrix
V will be d x m matrix
U is the user-feature matrix
V is the item-feature matrix

Singular value decomposition (SVD) is an algorithm commonly used for matrix factorization. We can use it to find items to recommend to users.

M - A matrix you want to decompose; in your case, it’s the rating matrix U - user feature matrix Sigma - weights diagonal Vt - item feature matrix

Questions

how to scale collaborative filtering/matrix factorisation in production for large number of users?

how to update them real-time or close to real-time?

Name		Name	Last commit message	Last commit date
Latest commit History 46 Commits
data		data
docs		docs
movie_helper		movie_helper
.gitignore		.gitignore
00_setup.ipynb		00_setup.ipynb
01_matrix_factorization_vowpal.ipynb		01_matrix_factorization_vowpal.ipynb
02_stochastic_gradient_descent.ipynb		02_stochastic_gradient_descent.ipynb
03_alternating_least_squares.ipynb		03_alternating_least_squares.ipynb
04_ml100k.ipynb		04_ml100k.ipynb
05_recommendation_system_with_tensorflow.ipynb		05_recommendation_system_with_tensorflow.ipynb
06_matrix_factorization.ipynb		06_matrix_factorization.ipynb
07_collaborative_filtering.ipynb		07_collaborative_filtering.ipynb
08_knn_collaborative_filtering.ipynb		08_knn_collaborative_filtering.ipynb
09_matrix_factorization.ipynb		09_matrix_factorization.ipynb
10_matrix_factorization_with_bias.ipynb		10_matrix_factorization_with_bias.ipynb
11_non_negative_matrix_factorization.ipynb		11_non_negative_matrix_factorization.ipynb
12_neural_netowrk.ipynb		12_neural_netowrk.ipynb
13_torch.ipynb		13_torch.ipynb
14_matrix.ipynb		14_matrix.ipynb
Makefile		Makefile
README.md		README.md
lda.ipynb		lda.ipynb
main.py		main.py
matrix_factorization.ipynb		matrix_factorization.ipynb
matrix_factorization.py.ipynb		matrix_factorization.py.ipynb
matrix_factorization_101.ipynb		matrix_factorization_101.ipynb
mf-run.sh		mf-run.sh
mf-setup.sh		mf-setup.sh
mf-test.sh		mf-test.sh
poetry.lock		poetry.lock
pyproject.toml		pyproject.toml
signature.py		signature.py
singular_value_decomposition.ipynb		singular_value_decomposition.ipynb

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

matrix-factorization

Goal