I am curious how the public scores are calculated. I am trying to correlate the score I get with the log loss I am calculating. Sometimes I get a better score with a higher log loss and worse score with a lower log loss. My best score right now is with a model that generated a log loss of #Log_loss = 7.915164020966691 and my second best score had a log loss of #Log_loss = 6.987650697117431. These are on the modeled data, of course, not the new test batch, which generates the score. So I am not sure how to relate the two.