2
$\begingroup$

Let’s assume that four participants (A, B, C and D) take on five sport-challenges (e.g. swimming, running, ...). Our goal is to predict the placement of each participant for each challenge. Moreover, let’s assume we have appropriate predictors. We know that each placement (1 to 4) is unique for each challenge (only one winner ...).

My questions:

I think this prediction task is a multi-label multi-output classification, right?

Are there any algorithms, which provide unique forecasts (for each class). In other words, the algorithm should classify each person in one unique class (1 to 4). Obviously, we know a priori that same placements are very unlikely.

Thank you!

Greets

$\endgroup$
2
$\begingroup$

2 Questions.

  1. I think this prediction task is a multi-label multi-output classification, right?

Yes. But you could consider splitting these joint classification tasks into seperate ones (i.e. for swimming whats the placement of A,B,C,D etc) IF you data allows it.

  1. Are there any algorithms, which provide unique forecasts (for each class)?

Think in terms of probabilities. The strongest (most discriminatory one) is the one that has the highest-not-so-much-overlapping probabilities to different outputs. So taking these probabilities and "rounding" (possibly forcing some predictions if model is highly uncertain---meaning, for example, that he predicts 25% for every placement, here some other heuristics should be used. For example create a new variable that says how did competitor fare in other competitions etc) that up will give unique classes.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you for the response! I am not sure if your procedure is really what I am looking for. What about this approach: as we receive the probabilities for each class and for each participant (as well as for each sport-challenge) we can calculate a weighted score. For instance a forecast for the first challenge and participant "A" might be ... 1=25%, 2=25%, 3=25% and 4=25%, which yields a weighted average of 2.5. Calculating these scores for each participant (and each challenge) provides the possibility to re-classify based on it. Might such a post-prediction arrangement bias the results? $\endgroup$ – Blackout Jan 7 at 12:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.