@@SebastianRaschka The model in itself is interesting but what would be cool is that if the model is unsupervised, by means let it choose the boundary index by itself. That is if you have data that is ranked. But let us say I have my data and its from 1 to 5, in level 1 there is100 samples, and in number 5 I only have 25. Will that be a problem in your model? Thank you
@@Nostalgia-futuro The reason why it can't predict the rank outside the range is that it's not clear how many out of range indexes there are. E.g., if you have labels 3, 4, 5 in the training set, it learns e.g., p > 4. If p > 4 is a large probability, you may suppose the label is 5 based on the training data. But if you have a new dataset with labels 10, 11, 12, the model would just know p > 4, but naturally not be able to distinguish between 10, 11, and 12. Btw the class imbalance issue you describe is not problem.
At 13:03, there are two yes from where extra comes? And predicted label should be 2 because that node have more probability threshold. Can you please explain?
So it's 3 because you have to add +1 to the number of "Yes"es. Why? So the 2nd output node is essentially for "rank > 2?". If that's a yes then it means 3 not 2. (Assuming here that "rank > 3" is false)
Thank you, If our dataset has imbalanced classes, can we use 'weights' (class weights) similar to how we implement them in PyTorch for cross-entropy loss, but for the CORN loss?
wonderful talk, thank you! I'm wondering if there's any performance metrics you know of for ordinal regression tasks?
can this be used for sentiment anlysis too for example?
Yes! We did include a text example in the paper using an RNN. Will probably add a transformer one in the next couple of days as well.
Great talk thank you
Glad to hear you found it useful!
I was wondering if those models are supervised or unsupervised learning for choosing the index boundary when it comes to its input data
The index boundary is based on the ranks in the training set. (It's similar to how you determine the class labels in a classification problem)
@@SebastianRaschka The model in itself is interesting but what would be cool is that if the model is unsupervised, by means let it choose the boundary index by itself. That is if you have data that is ranked.
But let us say I have my data and its from 1 to 5, in level 1 there is100 samples, and in number 5 I only have 25. Will that be a problem in your model?
Thank you
@@Nostalgia-futuro The reason why it can't predict the rank outside the range is that it's not clear how many out of range indexes there are.
E.g., if you have labels 3, 4, 5 in the training set, it learns e.g., p > 4. If p > 4 is a large probability, you may suppose the label is 5 based on the training data. But if you have a new dataset with labels 10, 11, 12, the model would just know p > 4, but naturally not be able to distinguish between 10, 11, and 12.
Btw the class imbalance issue you describe is not problem.
At 13:03, there are two yes from where extra comes?
And predicted label should be 2 because that node have more probability threshold.
Can you please explain?
So it's 3 because you have to add +1 to the number of "Yes"es. Why? So the 2nd output node is essentially for "rank > 2?". If that's a yes then it means 3 not 2. (Assuming here that "rank > 3" is false)
Hi Sebastian,
Is that ok to tune the threshold (default 0.5) when converting logits into labels?
Yes, you could do that.
Thank you, If our dataset has imbalanced classes, can we use 'weights' (class weights) similar to how we implement them in PyTorch for cross-entropy loss, but for the CORN loss?