This is a Movie Recommendation Engine using Restricted Boltzmann Machine (RBM) approach for collaborative filtering. This engine recommends movies by trying to find users that are similar to each other based on their movie ratings given by each user.
RBM is a Generative model with two layers(Visible and Hidden) that assigns a probability to each possible binary state vectors over its visible units. Visible layer nodes have visible bias(vb) and Hideen layer nodes have hidden bias(hb). A weight matrix of row length equal to input nodes and column length equal to output nodes.
If you have CUDA enabled GPU in PyTorch then it will be selected by defualt
We can use any dataset variant from MovieLens. In load_data.py->load_dataset(size), there are 3 choices as following:
- small : http://files.grouplens.org/datasets/movielens/ml-latest.zip
- medium : https://grouplens.org/datasets/movielens/10m/
- large : https://grouplens.org/datasets/movielens/25m/
Since the movie-ids are not continous i.e. there are some gaps in numbering, we encode all movie-ids to category-ids and save it to a dictionary.
We create 2 lists : movies and ratings where each element in the list contains the following information :
- movies-list : list of ids of movies a user has rated
- ratings-list : list of ratings given by user given to movies corresponding to movies-list. Rating are normalized to be in range [0,1]
RBM has 2 layers : Visible and Hidden
- Visible layer size : Number of movies
- Hidden layer size : 1024 units
First install the required packages
git clone https://github.com/rushiv0609/Movie-Recommendation-using-RBM.git
pip install -r requirements.txt
python main.py
Each input batch is transformed into a torch tensor of shape [batch-size, number-of-movies] using movies-list & ratings-list. Then this transformed input is passed to RBM.
k-contrastive divergence is used to train the weights and bias of RBM.
We can use recommend() function in recommend.py to recommend top 10 movies by passing user-id as parameter