Kaggle turned machine-learning problems into public competitions

Companies accumulate data they want to turn into value, but data scientists are rare and organizations struggle to acquire talent. Kaggle.com offers an alternative: companies host machine-learning competitions and participants build prediction models for prizes.

A 2020 HICSS study used ten expert interviews plus data crawled from Kaggle to understand how organizations use competitions. It found real benefits — discussing with participants and learning state-of-the-art approaches — but noted competitions can only cover a fraction of the tasks in a typical data science project.

The study identified 12 factors within three categories that influence an organization's perceived success when hosting a data science competition, providing a framework for when crowdsourcing beats hiring.

Crowdsource the model via competition

The crowd is far larger and more diverse than any in-house team.

A clear evaluation metric lets organizations compare models objectively.

Competitions expose internal teams to state-of-the-art approaches.

Hosting costs a prize, not salaries and headcount.

If expertise is scarce and expensive, buy solutions instead of headcount: a well-scoped competition with a clear metric lets the crowd do the work and teaches your team in the process.

Kaggle became the reference marketplace for data-science competitions, used by companies and researchers worldwide to source models, benchmark approaches and recruit talent, before being acquired by Google in 2017.