I have few questions:
-
How many sites should be in submission? We have 18130 samples in grid_cells.geojson, 10878 in train labels and 9066 in sample submission.
-
How far ahead should we forecast every week? 1 week, 2 weeks, all weeks till the 1st of July.
-
How you will evaluate on a weekly basis. You will evaluate on your servers(in this case we need hardware specs, evaluation timeout and so on. It’s really very important) or we will evaluate it by ourselves and submit fresh submission every week(I have no idea how you would control “no changes” and “used sources” parts from rules in this case)
-
Could you please provide more details about the evaluation process. Like: Jan 11 6:00-7:00 UTC we will execute all submissions and every submission should output 18130*2 values for dates Jan 18 and Jan 25. Evaluation time is also important, as it may have influence on data availability for some sources, and this may have influence on model and used sources.
-
What will you do if submission doesn’t produce output for some source and gt for this source would be null as well? What will you do if some submission will work fine for most execution dates, and fall without valid output once(or twice)?
-
I’ve checked cell_id = 01be2cc7-ef77-4e4d-80ed-c4f8139162c3 and it’s not 1 * 1km, it’s 0.8 * 0.8km. Is it fine?
-
Only 218 rows in train_labels have at least 10% non-null values. The other one(10.5k or 18k) may have few zeros, or nothing, or 2 measures 0 and 110. These random values may have a random and significant impact on the final score during testing. For example, 0dabd62e-e70a-4d05-8e2c-22aedebc94ab have known measures for 3 dates 2018-03-31=4.8, 2018-05-24=0.1 and 2019-04-07=110.1 and this last 110 may be more important than “fair” RMSE for main 218 cells with well known data. How will you deal with this?