changhyeonnam / MF

Matrix Factorization with MovieLens dataset in Pytorch

MFML (Matrix Factorization MovieLens in torch)

This repository is implementation about Matrix Factorization Techniques for Recommender Systems

Dataset

The MovieLens Dataset is used as a bench mark dataset in journals, conference papers, and State-of-the-Art papers in the Recommendation System field. The MovieLens dataset consists of columns including userId, movieId, rating, timestamp, tag, title, genres, etc. Among the columns, the ratings.csv file consisting of userId, movieId, rating, and timestamp is used. Based on a 100k dataset, the dataset consists of 610 users, 9742 movies, and 100836 ratings. MovieLens data is basically a user-based dataset, including only users who have evaluated at least 20 movies for each user. The file size ranges from 100k, 1M to a maximum of 20M.

Model explanation

• A Matrix Factorization model was implemented based on the Matrix Factorization Techniques for Recommender Systems paper.

• The Matrix Factorization model completes the matrix for the target by inner product (dot product) of latent factors for user-item interaction. MF can be implemented considering both implicit feedback and explicit feedback, and the model was trained in consideration of only explicit feedback in the task.

• In this task, in addition to dot product of features, terms such as bias and confidence score described in the paper were added to the prediction.

  • The bias term was added to prevent the bias of ratings for specific users or specific items from being included in the model.
  • The confidence score described in the paper is an indicator of how reliable the user is, and it can also be judged by the frequency of ratings for each user.

  This formula is a cost function with the addition of bias and confidence score terms. RSME (Root Squared Mean Error) was used as the cost function.

Matrix Factorization model directory tree

.
├── dataset
├── evaluation.py
├── main.py
├── model
│   └── MF.py
├── optimize.py
├── train.py
└── utils.py

development enviroment

• OS: Max OS X
• IDE: pycharm, gpu sever computer vim
• GPU: NVIDIA RTX A6000

Quick start

python main.py -e 30 -b 32 -f 30 -lr 0.001 -down True

Matrix Factorization Result

MovieLens 100K	average cost(RMSE)	epoch	# latent factor
Matrix Factorization	0.9108	30	30
Matrix Factorization with confidence score	0.9242	30	30
Matrix Factorization with bias	0.9668	30	30
Matrix Factorization with confidence score and bias	1.4973	30	30

The number of factors showed the best performance at 19.

epoch =30, batch=32, learning rater = 0.001, best factor =19. (RMSELoss per epoch)

reference : Matrix Factorization Techniques for Recommender Systems review written in korean : Review

Matrix Factorization with MovieLens in torch