Engineer Blog Note 4: PinSage: A new graph convolutional neural network for web-scale recommender systems

2 min readSep 6, 2020

My note:

They applying Graph Convolutional Networks in graph application improve performances.

2. There are few challenges of utilizing GCNs, Pinterest provides solutions to them.

What is GCNs:

GCNs is Graph Convolutional Networks. It is useful to capture the graph information to embeddings.

Why GCNs?

In Pinterest, we can consider image-boards as a bipartite. In other applications, user-item clicks graph is also a bipartite.
In previous model at Pinterest, they only consider visual and annotations (content-based features). I feel like strictly speaking, annotations is semi-content-based features since it is user defined, which includes user information as well. However, both of them do not consider interactions information (like collaborate filtering concept). I think the graph information mainly supplement the lack of information of annotations (some images are lack of these annotations).

May we directly apply? NO!

The original model is not applied in large scale scenario.
Some assumptions, full graph Laplacian during training, don’t apply in a large scale situation.

Solution: a random-walk-based GCNs

They apply localized convolutions by sampling the neighborhood around a node and dynamically constructing a computation graph instead of multiplying feature matrices by powers of the full graph Laplacian.
They apply random walks instead of K-hop graph neighborhoods to earn importance of neighboring and memory controls.
They apply MapReduce computational to solve computations of node embeddings.

Result