Upon learning that the evaluation system for this competition is based on log loss, my initial inclination was to explore the common practice of utilizing threshold values or rounding up probabilities to either extreme (0 or 1). However, a deeper dive into the contest guidelines revealed a discouragement of such approaches. Alongside this cautionary note, the promotion of leveraging large ensembles for improved efficiency was somehow discouraged.

"The values can be between 0 and 1. Do not set thresholds (or round your probabilities) to improve your place on the leaderboard. In order to ensure that the client receives the best solution, Zindi will need the raw probabilities. This will allow the clients to set thresholds to their own needs."

In essence, this competition challenges us to navigate the uncharted territories of model evaluation, emphasizing the significance of delivering raw probabilities for the client's flexibility in setting thresholds. It underscores the pursuit of optimal solutions that cater to individual client needs rather than a generic, leaderboard-driven approach.

Wishing all my fellow participants the best of luck in this journey of innovation and efficiency.

Warm regards,

Ntem, Kenyor K