i came across a cool talk where mariia bulycheva from zalando breaks down how they moved away from classic deep learning models to using gnn for their platform. she dives into the nitty-grity of turning messy logs like clicks and searches (aka user activities)
into useful graphs with different types of nodes & edges.
she explains that "message passing" is key - think: each node sending info only if it's relevant, kinda like a whispering game but in code. there's also the pesky issue she faced where graph data leakage could mess up privacy and model accuracy - basically your personal stuff getting used without consent .
what really blew my mind was how they tackled latency issues with
a hybrid architecture that allowed faster inference times, pumping out those context-rich embeddings for use in other models. it's like having a supercharged version of yourself ready to share all the cool info you've learned!
anyone else tried this approach and had similar successes or faced different challenges? i'm curious if anyone is using gnn on their social media platform, especially when dealing with complex user interactions
article:
https://www.infoq.com/presentations/graph-neural-networks/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global