A simple Recommendation System on MovieLens dataset developed with PyTorch and served with TorchServe REST API.
The MovieLenst-1M dataset that contains 1 million ratings from 6000 users on 4000 movies.
The model learns an embedding for each user and each movie and concatenates the embeddings with user features (gender and age) and movie features (genre) to build a rich feature space. On top of that, a 3-layer MLP learns to predict users' ratings on the movies in the dataset.
Result on 10% of the data held out as the validation set: RMSE= 0.848
The trained model and related files are located in ./model directory.
To use the service, first we need to build a docker image:
$ docker build -t recom .
And then run it:
$ docker run -it -p 8080:8080 recom
Now, we can use req.py to send a request to the service for a (user_id, product_id) pair, and receive the prediction of the model for the user's rating on the movie.
We can then recommend movies to the user based on the predicted ratings.