Group-movie-recommender-system

In this project, we create a matrix factorization based Recommender System for a group of users. We first carry out Stochastic Gradient based Matrix Factorization of user-movie rating matrix to calculate user and movie factors.

We generate groups of users of 3 different sizes. Small (3 members), Medium(5 members) and Large (10 members) and predict group ratings using methods described below.

We try out 3 different methods.

After Factorization: where we aggregate factors of users into group factors after factorization

Before Factorization(BF): where we aggregate users' ratings into a virtual user. We calculate group factors by using simple ridge regression.

Weighted Before Factorization(WBF): Same as BF except that no. of movies watched by users are taken as weights. We solve it using weighted ridge regression method.

Finally we evaluate our project (getting roughly 80 % precision)

Paper

Project is based on the following paper:

http://www.sciencedirect.com/science/article/pii/S0020025516300196

Dataset

Dataset: https://grouplens.org/datasets/movielens/100k/

Video

Video: https://www.youtube.com/watch?v=ycf2sY2XnN8

Execution Dependencies

The notebook can be run directly. Dataset is included in the github repo. Also, the python code can be run by:

python ./GroupRec.py

We are using pandas, numpy, scipy and warnings modules. Install them by running.

pip install numpy
pip install pandas
pip install scipy

The arguments are taken via a config file config.conf that is present in the same folder. The hyperparameters for matrix factorization, group sizes and no. of generated groups can be changed through the config file.

Note: Since the notebook uses relative paths for dataset and images in res/ subdirectory, please run in the same folder when it is cloned from repo.

Flowchart