Spotify Playlist Recommender System

A Spotify playlist recommender System for automatic playlist continuation using Graph Neural Networks

Heterogeneous graph data preperation

- A heterogenous graph is constructed using a subset of the Spotify million playlist Dataset.
- The graph contains playlist, tracks and artist nodes.
- There are multiple edge types that connect different types of nodes.
- Negative sampling of nodes is done to prevent all the node embeddings from being the same, i.e this enables the model to learn that the playlist is more like(contains) song A and is different(does not contain) song B.

Heterogeneous Graph transformer Model

- A Heterogeneous Graph Transformer is used to update node embeddings by aggregating node features from neighboring nodes.
- Finally to predict whether a link is present between two nodes, an inner product is done between the node embeddings to find how similar they are.
- Bayesian Personalized ranking which takes into account both the positive and negative predictions is used to train the network.

Results

The model shows good performance acheiving an AUC of 0.9 and Recall of 0.74