This Repository holds the notebooks which we used to train our model, iMult-VAE.
Our model is variation of variational autoencoders and collaborative filtering, and is used to perform Top N predictions for songs recommendations, which is a task in the field of Music Recommendation Systems.
Our model is based on the Mult-VAE model which was published in the paper: "Variational autoencoders for collaborative filtering" by Dawen Liang, Rahul G. Krishnan, Matthew D. Hoffman, and Tony Jebara, in The Web Conference (aka WWW) 2018.
Our model differs in two ways from the original paper:
-
The original paper uses only implicit user feedback to learn its weights, we incorporate item-based side information to better tune the model’s weights.
-
Different beta selection: The authors of the original paper describe a parameter named which is used to control the strength of regularization of the objective function, the original paper performs a simple heuristic to optimize the value. We suggest to use user-specific values.
Dataset:
For the evaluation of our models, we used the Million Song Dataset. Specifically, we used two datasets from it:
-
The echo nest dataset
This dataset contains (user id, song id, play count) triplets for one million users.
This dataset was used as implicit user feedback.
http://millionsongdataset.com/sites/default/files/challenge/train_triplets.txt.zip
-
The Last.FM Dataset
This dataset contains metadata for the million songs in the MSD datset.
This dataset was used as complementary side information for the echo nest’s data.
http://labrosa.ee.columbia.edu/millionsong/sites/default/files/AdditionalFiles/track_metadata.db