When GraphSAGE Meets Pinterest

Mar 25 · 3 min read

In the previous stories, we introduced Introduction to Graph Embeddings, Random Walk in Node Embeddings, and 4 Graph Neural Networks. I would like to further extend to PingSAGE. It was proposed by Ying et al. in 2018. Ying is one of the authors of GraphSAGE (Hamilton et al., 2018). He and his team apply a modified version of GraphSAGE (Hamilton et al., 2018)in the Pinterest dataset.

We will use the following undirected graph to go though PinSAGE (Ying et al., 2018) algorithm. First of all, we have an input graph with Node A, B, C, D, E, and F. From node respective, and every node is a root node while other nodes are neighbor nodes.

We will take Node A is an example. The neighbor nodes of node A include nodes B, C, and D. PinSAGE (Ying et al., 2018) use depth-2 convolutions, so we need to aggregate nodes B, C, and D by their neighbor nodes as well.

In the 1-hop layer (i.e., node B, C, and D), we have aggregation function (i.e., CONVOLVE) to aggregate all neighbor nodes. After aggregated neighbor nodes, both target node’s embeddings (i.e., node A) and neighbor nodes’ embeddings (i.e., nodes B, C, and D). Finally, we get a node A’s representation.

How do we get nodes B, C, and D representation? We use the same technique in the previous step but different weighting to get embeddings. You may notice node A exists again in the second layer of neighbor. It is expected as node A is a neighbor node of nodes B, C, and D. If we remove it from aggregation, we ignore the neighbor relations.

Instead of using all neighbors, Ying et al. propose to use an important-based random walk approach to include the most important neighbors only. It not only reduces the computation footprint but also takes accounts into important neighbors.

References

Written by

More From Medium

Data Scientist Coding Exercise

Mar 19 · 6 min read

Image Filtering

Mar 28 · 5 min read

More from Towards AI

Mar 29 · 8 min read

103

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just \$5/month. Upgrade