Finnet Challenge 3rd place solution

The task of the competition was to find removed transactions between firms.
Our final solution is blending of several different approaches:

Recommended Systems approach
Matrix factorization. We have used several SVDs with 400, 450 and 475 components.
Link Prediction approach
Node2Vec embeddings were used as features for logistic regression.
Stacking & Blending
Boostings with combinations of categorical features, node2vec embeddings and aggregation features.
Using a leak of data

Name		Name	Last commit message	Last commit date
Latest commit History 25 Commits
img		img
Durka_Science_FinNet.pdf		Durka_Science_FinNet.pdf
README.md		README.md
Stacking&Blending.ipynb		Stacking&Blending.ipynb
creating_all_svd.ipynb		creating_all_svd.ipynb
eda_node2vec.ipynb		eda_node2vec.ipynb
ensemble.ipynb		ensemble.ipynb
new_vertices.ipynb		new_vertices.ipynb
nodes.ipynb		nodes.ipynb
svd.ipynb		svd.ipynb

Provide feedback