💻 Trending Now: CV vs LB (without using future...

AI/ML for 5G-Energy Consumption Modelling by ITU AI/ML in 5G Challenge

20 000 CHF

Completed (over 2 years ago)

Skills you will learn

Prediction

1029 joined

277 active

Info Data Chat Leaderboard

Start

Jul 26, 23

Oct 13, 23

Reveal

Oct 17, 23

Krishna_Priya

CV vs LB (without using future data)

Connect · 18 Sep 2023, 14:51 · 9

Hey guys, How is the CV vs LB of the model according to competition rules? Will help us raise the bar and deliver quality models in general.

NOTE: The only rule stated by the host, the model must NOT use future data for modeling.

My latest: CV - 0.739, LB - 0.9933

Discussion 9 answers

University of Yaoundé I

My best: CV - 2.16, LB - 2.13

18 Sep 2023, 15:38

Upvotes 0

newbee

What's your CV strategy? How many folds do you use? Thanks.

19 Sep 2023, 00:29

Upvotes 0

Krishna_Priya

custom kfolds (based on competition test set), 10 folds

replied to newbee19 Sep 2023, 11:57

Upvotes 0

nooby

CV - 1.64, LB - 2.15

Did you use historical data, especially historical data of the Energy?

19 Sep 2023, 14:15

Upvotes 0

Krishna_Priya

Yes, I have. There is no rule from the organizers stating not to use it until now . But yes, we will have to explain the approach and it's real world application.

replied to nooby19 Sep 2023, 14:22

Upvotes 0

ahmedattia

I am a bit confused about the use of future data.

Does the restriction considers only the use of future values of the energy consumption field (lag features for example ) or the entire future data ?

If the response is the latter, how is it possible to predict energy consumption for the first hours of the first day ( in that case only few training samples can be used ) to train the model.

20 Sep 2023, 13:54

Upvotes 0

Krishna_Priya

Hi @ahmedattia,

In the past discussions, it was stated not to use the future values as features in the model which should include any KPI + target variable.

However, we can train on the complete data, as training on complete data will capture the instantaneous physical relationships between the target and independent KPIs, so for the test samples of the first hours your prediction will be based on these instantaneous relationships.

PS: The above thoughts are solely mine and do not represent the organizer's point of view.

replied to ahmedattia26 Sep 2023, 16:00

Upvotes 2