Utilize your graph structure throughout the pipeline.

[kfoynt]

Issue

Your network structure is utilized only at the beginning of the pipeline in a pre-processing step. This is classical ML where you manually construct features for your users. This is known not to provide the best performance, see for example node2vec vs graph neural networks.

Solution

Use graph neural networks to utilize the network structure through the ML pipeline.

Some Details

Currently you use PageRank on your network to pre-process the network and vectorize it. PageRank is giving you an embedding for each node. Then the pre-processed embeddings are fed to a neural network.

PageRank features are usually of good quality, however, it is best if you use them in combination with other individual features that you might have and feed them to graph neural network. This way you will allow the model to decide what part of the graph structure to use and which features.

[aloiscochard]

I'm sorry @kfoynt but besides giving a technical definition graph neural networks you seem to have completely failed at motivating its use case.

We will consider this as duplicate of https://github.com/twitter/the-algorithm/issues/1760 for now.

(I use youtube sometimes, with the command line interface)

We are happy to revisit our decision in future when you'll be ready to share your motivations.