[x] [Mila] metric NDCG vs MRR in the context of RecSys (can be shown that NDCG is almost equivalent to MRR under the circumstances that only 1 hotel is clicked)
[x] [PJ] methodology: transposition of data (item per row), quick validation process, iteratively check the idea
[x] [PJ] features: numerical features, ranking transformations within the clickout group, lag features - compare items to next/prev items, accumulators (prevents leaks)
[x] additional features extracted from properties (all the data_prep scripts)
[ ] feature importance - both meta features and particular features (numeric)
[ ] too revealing data - what would happen if there was no time and/or rank
[x] [Mila] model (LightGBM) - parameter tuning, C++ custom objective
[ ] blend - list of models used (with parameters)
[ ] possible improvements - don't include the time of the click in the dataset
Things to write about