4th place Finnet Challenge solution

Task description

An anonymized graph of transactions between legal entities in Russia is given. Some of the edges were removed from it. The task is to restore them.

Data description

Link to the dataset: https://drive.google.com/open?id=1dRfIzUsmwD-ZmURbwPbGjbG-kfcgDsd-

vertices.csv contain information about each company:

• id - id of a legal entity
• main_okved - main OKVED of the legal entity
• region_code - region code of the legal entity
• company_type - type of the legal entity

edges.csv contain information about transactions between companies:

• id_1 - first company id
• id_2 - second company id
• value - anonymized summary value of all transactions
• n_transactions - anonymized number of transactions between id_1 and id_2

ids.csv contain ids we have to predict edges for

• id - id of the company

Solution description

My top 4 solution is just Node2Vec(walklen=4, dimensions=100) and a lot of different data mixed in one ensemble (9 CatBoost models).The most important things in my models were squared matrix of interactions each main_okved to each main_okved and the same case for region_code. I tried a lot of things including common NN with different ids embeddings, but it was absolutely bad for the public leaderboard. Also I tried to predict number of real edges, but it often predict value that less than real value, so it is useless.

Notes

During this competition I found wonderful python module to speedup Node2Vec computation which is based on CSR sparse matrices: nodevectors

Additional data

• Pretrained Node2Vec model
• Result of the work of each model from solution.ipynb

Results