Optimizing evaluation metrics
Summing up what we have discussed so far, an objective function is a function inside your learning algorithm that measures how well the algorithm’s internal model is fitting the provided data. The objective function also provides feedback to the algorithm in order for it to improve its fit across successive iterations. Clearly, since the entire algorithm’s efforts are recruited to perform well based on the objective function, if the Kaggle evaluation metric perfectly matches the objective function of your algorithm, you will get the best results.
Unfortunately, this is not frequently the case. Often, the evaluation metric provided can only be approximated by existing objective functions. Getting a good approximation, or striving to get your predictions performing better with respect to the evaluation criteria, is the secret to performing well in Kaggle competitions. When your objective function does not match your evaluation metric,...
