Basic Needs Basic Rights Kenya - Tech4MentalHealth
$4,200 USD
Classify text from university students in Kenya towards a mental health chatbot
436 data scientists enrolled, 219 on the leaderboard
HealthClassificationNLPUnstructuredSDG3
Kenya
26 April—5 July
Ends in 1 month
Evaluation metric clarification
published 16 May 2020, 07:20

I understand that the evaluation for this competiotion is based on log loss, what I fail to understand is why there isn't a restriction of probabilities summing to 1. If someone gives a '1' prediction to all classes wouldn't the log loss be 0 and hence the best. Am I missing something here?

Log loss takes into account both the times the actual value is 1 and when it is 0. This means that when you mark all labels as 1, you don't get any loss from the true label, but you get a loss from the wrong labels marked as 1 (they should be closer to 0).

You could also mark everything as 1 and submit for the sake of experimenting if your daily submission quota isn't depleted :-)

Ok, i misunderstood the metric to be multiclass log loss. Thinking interms of binary log loss makes it clear. Thanks

i think they want from us that how statement are related to different labels as when you read text of depression and suicide

it seems difficult to classify them