@ihar Yes, I received an email today stating that I have been disqualified

did they mention the reason?!

They didn't like that I used 'id' as one of the model features.

It's against the rules of Zindi

@Ebiendele Can you show me where this rule is? I'm unable to find it.

using the ID in this competition is not considered cheating, infact it's really important to include it in the training set as it helps the model differentiate between historical patterns of each client.

It was discussed some years back I can't remember the challenge the use of id as a feature in model is not useful to the clients

It might definitely serve as a good feature in modelling but we're not meant to use it that what bring disqualification

@Ebiendele A discussion from years back can't be considered an official rule. This rule should be written in the competition description.

Also, the submission file only includes clients from the training data, so basically the competition is asking for the model to generalize for those clients. If the host wants the models to generalize for any clients, the submission should include different clients from the training data, or you should state in the description that using IDs is against the rules.

What is very unfair is telling me about this 'rule' after the competition has ended. It's very disrespectful to the person who invested a lot of time in the competition.

@Amy_Bray @Nelly43 can one of you answer here?

what was the score without the 'id' column? btw this was impressive. you really did extensive analysis

Hello @yanteixeira,

We recognise that you, and everyone else on the leaderboard, have put considerable effort into this competition, and our team understands your situation.

Please note that the ID is not a feature from the data but rather it was created through a random process that is virtually different from and unrelated to the customer-client interaction on the Adbot platform. Moreover, an important quality of a good machine learning model is generalising over new/unseen data, which includes the test set as well as any new set of data that is created after the model has been trained and validated.

For this competition, it was highlighted that winning solutions would have to:

• submit a description of the variables that had the most significant impact on engagement (number of clicks), and
• show how much the top 5 variables(predictors) affect the total number of clicks (target).

This was to primarily demonstrate the interpretability and usability of your model(s).

Using the ID as a feature in your model not only introduces a leakage (an issue that was raised in an earlier discussion), it also renders the model impractical since there is no real-world interpretation of the relationship between the IDs and number of clicks, or any other variable for that matter.

Zindi competitions are often based on real-world problems with winning solutions ultimately being handed to hosts for implementation. Therefore, to help ensure your solution wins, it would be useful to build robust models that not only adhere to competition rules but also incorporate best practices. This is because your work will always be evaluated in the same way as when delivering solutions to a client, especially when emphasised in the competition's details.

I hope this helps.

Best,

Nelson.

@Nelly43 Thanks for the answer. By the way, if this text had been the email I received, I would have understood right away.

In any case, we learn from our mistakes. See you all in the next one