Solution of team "Двоеточие" for Finnet Challenge
So, the best score was achieved with an ensemble of 7 boostings using 256-dim node2vec embeddings + some aggregation features.
We also tried things like graph neural networks, recommendation based on adjacency matrix factorization and nearest neighbors, but these solutions had bad public results.
Mixing some solutions up sometimes improved results
