-
Basic GNN architecture.
Inspiration
Friend suggestions are a staple of social network apps. This feature is not yet implemented in Lens, so we've decided to leverage GNNs to extend what the protocol can do.
What it does
Our app is able to suggest Lens profiles that could be interesting to you, based on the profiles you are already following.
How we built it
A decentralised, open social graph is a great boon for ML developers. We've scraped the entire follows network using the Lens API. Then we've trained a graph neural network in PyTorch to do link prediction. The trained model calculates a matrix of scores, giving a probability that some profile will want to follow another. This data is fed into our React application which enables you to view the profile recommendations.
Challenges we ran into
Training message-passing neural networks on data with a large node count is always tricky and we ran into computational constraints quickly. We were able to overcome those by leveraging cloud computing.
Accomplishments that we're proud of
Using cutting-edge ML techniques to build a useful service for an exciting protocol.
What's next for Friend Suggestions on Lens
The next step is to turn the application into a DAPP that will allow users to get recommendations for their profiles using micro-transactions.
Log in or sign up for Devpost to join the conversation.