My Problem With Kaggle Competitions

For those who don't know Kaggle Competitions: they are machine-learning competitions where multiple teams (or individuals) compete for the best score.

My problem is shown in the following three screenshots of the final ranking (the leaderboard) for three Kaggle competitions:

Predicting Red Hat Business Value:

PUBG Finish Placement Prediction

Google Analytics Customer Revenue Prediction competition:

What do you notice in these screenshots?

The difference in score that allows somebody to win a competition is tiny and not practical in my opinion. Let's take the Predicting Red Hat Business Value competition as an example. In this competition, participants are asked to identify which customers have the most potential business value for Red Hat based on their characteristics and activities; they are required to assign a probability (between 0 and 1) to each customer. The metric used for evaluation is area under the ROC curve.

We can see in the leaderboard that the score of the third team is 0.99459 while the score of the fourth team is 0.99400. The small difference between these two scores ( = 0.00059) qualified one to win the competition and deprived the other of that.

Moreover, the difference between the first-ranked participant and the 300th-participant is only 0.00343. Does a difference of 0.00059 really reflect a practical difference between the two solutions? I doubt that.

So is there an alternative?

I think that a better alternative would be to define score ranges for evaluation. For example, for the Predicting Red Hat Business Value competition, consider those whose scores are higher than 0.992 as the winners of the competition, those whose scores are higher than 0.9917 as gold-medal winners, and so on. This will leave us with 94 competition winners and 150 gold-medal winners for example (the number of participants in this competition was 2271).

And for money prize, the current approach can be followed where the first three get the money since it is not reasonable to distribute it to 94 teams.

I think that this alternative encourages the development of more generalized models, and it will allow people with similar skill levels to have similar rank in the competition as well. This is what I think at the moment based on my understanding of Kaggle competitions.

Published on: 24 Mar 2019   |   Last modified: 24 Mar 2019

Tags data-science kaggle

Categories Data Science