Section: New Results

Compétitions d’apprentissage automatique avec le package R rchallenge

The following result has been obtained by Adrien Todeschini (CQFD member) Robin Genuer.

In machine learning, empirical performance on real data are crucial in the success of a method. Recent years have seen the emergence of a large number of machine learning competitions. These challenges are motivated by industrial (Netflix prize) or academic (HiggsML challenge) applications and put in competition researchers and data scientists to obtain the best performance. We wanted to expose students to this reality by submitting a challenge in the context of the machine learning course. The leaderboard is displayed on an automatically updated web page allowing emulation among students. The history of the results also allows them to visualize their progress through the submissions. In addition, the challenge can continue outside of the supervised sessions promoting independence and exploration of new learning techniques and computer tools. The system we have implemented is available as an R package for reuse by other teachers. Building on the R Markdown and Dropbox tools, it requires no network configuration and can be deployed very easily on a personal computer.