📦

Neural Graph Collaborative Filtering

Key ideas

The collaborative signal is typically not encoded in regular embeddings for users or items.
Proposal to capture bipartite graph into the embedding process.

2 key components in learnable CF models:
- Embeddings which transform user/items into vectorised representations
- Interaction modeling which reconstructs historical interactions based on the embeddings.
Matrix Factorisation for example directly embeds user/item ID as a vector and models interactions via inner product.
Neural CF replaces this inner product with non-linear NN.
Translation CF replaces inner product with distance metrics like Euclidean.

Diagram showing the interaction graph expanded as a tree structure for collaborative filtering.

Instead of expanding the interaction graph as a tree, which is complex and expensive:
- Devise an embedding propagation layer which aggregates embeddings of interacted items or users.
- Stack multiple embedding propagation layers.

Embedding layer: builds a parameter matrix as an embedding lookup table. This is the initial state of user and item embeddings to be optimised.

Embedding layer parameter matrix diagram.