thakurrajanand / avito2016 Goto Github PK

On the premise that one of the main differences between train and test is that train has ~25% of records with network_size > 200. So experiment first with earlier faster model that has both local CV and LB scores. First, filter validation for only network_size <= 200 to see impact on validation score. Next, train and valid on network_size <= 200. Last, do full train with network_size <= 200 and submit to see impact on LB.