Competition platforms teach a particular and valuable discipline: your opinion of your model is irrelevant, the held out score decides. DrivenData applies that to problems posed by non-profits, government agencies, research institutions and international organisations, and the resulting experience differs from Kaggle in ways that matter for anyone doing applied work. The problems are the draw. Predicting which buildings will fail an inspection, estimating water pump functionality across a country, detecting disease from field collected imagery, forecasting energy consumption in constrained settings.
These come from organisations that will act on the result, and that changes the shape of the task. Constraints tend to be real, class imbalance tends to be severe, label quality tends to be imperfect, and the evaluation metric is usually chosen because it reflects an actual decision rather than because it is convenient. The data is messier and this is the point. A great deal of Kaggle experience involves tabular data that has already been cleaned into a competition-ready state, and that skips the part of applied machine learning where most of the difficulty lives.
DrivenData datasets more often arrive with the artefacts of their collection intact: inconsistent categoricals, missingness that is not random, geographic and temporal structure that will punish a naive cross validation split. Learning to handle that is worth more than another few tenths of a point of ensemble tuning. The smaller fields are a genuine advantage for learning. Kaggle's headline competitions attract thousands of entrants including full time competitors and heavily resourced teams, and the honest experience for a newcomer is placing in the middle of the pack with no clear idea why.
DrivenData fields are often in the hundreds, the gap between a solid solution and the leaders is narrower, and a careful individual can place well enough to learn something from the comparison. Winning solutions are published with write ups and code, and because the competitions are less dominated by extreme ensembling, the winning approaches are more often intelligible and more often reusable. Reading those write ups is one of the better free learning activities available. The limitations are mostly about ecosystem.
There are far fewer competitions running at any time, so if none of the current ones interest you, you wait. The community is small, which means the public notebook culture that makes Kaggle so educational barely exists here. On Kaggle you can watch a strong practitioner work through a problem in public; on DrivenData you generally cannot, and you are more on your own. There is no integrated compute, so you bring your own environment, and no equivalent of Kaggle Learn, so no built in teaching material.
This is a place to apply what you know, not to acquire it. The two to four month competition windows suit sustained part time work and do not suit a weekend of experimentation. That is arguably realistic, and it is a different commitment from what many people expect. Set against Kaggle, my recommendation is straightforward: use Kaggle to learn, because the community and the public notebooks are unmatched teaching resources, then use DrivenData to test whether what you learned survives contact with data that has not been tidied for you.
The second experience is the one that resembles a job. My four point zero is for genuinely valuable practice on genuinely realistic problems, marked down for a thin ecosystem and for the absence of the learning scaffolding that makes competition platforms accessible to beginners.