Ratings and Rankings -- Using Deep Learning When Class Labels Have A Natural Order

Ғылым және технология

Deep learning offers state-of-the-art results for classifying images and text. Common deep learning architectures and training procedures focus on predicting unordered categories, such as recognizing a positive and negative sentiment from written text or indicating whether images contain cats, dogs, or airplanes. However, in many real-world problems, we deal with prediction problems where the target variable has an intrinsic ordering. For example, think of customer ratings (e.g., 1 to 5 stars) or medical diagnoses (e.g., disease severity labels such as none, mild, moderate, and severe). This talk will describe the core concepts behind working with ordered class labels, so-called ordinal data. We will cover hands-on PyTorch examples showing how to take existing deep learning architectures for classification and outfit them with loss functions better suited for ordinal data while only making minimal changes to the core architecture.
Slides: sebastianraschka.com/pdf/slid...
Code: raschka-research-group.github...
0:00 Introduction
0:32 Many Real-World Predictions Problems Have Ordered Labels
0:57 Ordered Labels? Tell Me More!
3:59 Can't we just use regular classifiers for ordered labels?
5:47 How? Let's (Re)Use What We Already know: An Extended Binary Classification Framework
8:07 Problem: rank inconsistency
10:53 Converting a Classifier into a CORN Model in 3 Lines of Code
13:09 Acknowledgements

Пікірлер: 9

  • @nguyenhuuuc2311
    @nguyenhuuuc23112 жыл бұрын

    Now ! CORAL mlp model for tabular data is one of my choices besides LGBM and XGBoost. When the talk ended, I saw a few tweets about that. I didn't think It'll be uploaded on KZread. Thus, I reached CORAL documentation and played around with it and made a few comparisons with LGBM. It's better than LGBM in that problem (both are tuned). - Thanks for your uploading so that I can learn and understand the architecture of CORAL/CORN more! 😊

  • @SebastianRaschka

    @SebastianRaschka

    2 жыл бұрын

    Wow I am glad to hear that it's that useful. Amazing to hear that it also outperforms LGBM on your problem task! :)

  • @736939
    @7369392 жыл бұрын

    Interesting. Did you separated the classes according the proportion of logits ? Like in formula below (for the linear scaling cases) class_interval = (argmin(logits) + argmax(logist)) / num_classes

  • @SebastianRaschka

    @SebastianRaschka

    2 жыл бұрын

    Thanks for the comment! Do you mean whether I tried this method for comparison?

  • @736939

    @736939

    2 жыл бұрын

    @@SebastianRaschka Yes, I mean this or it gives rank inconsistency? Also, is it possible to classify a hierartchical classes by using similar approach? Idea: the class that on the top of another class, intuitively should have higher rank logits.

  • @SebastianRaschka

    @SebastianRaschka

    2 жыл бұрын

    @@736939 Interesting, what loss function would you use with that though (e.g. considering that argmax/argmin are not differentiable)?

  • @SebastianRaschka

    @SebastianRaschka

    2 жыл бұрын

    @@736939 I do like the idea with the hierarchical classes. I think that esp with the CORN approach and conditional probabilities, that could be something interesting to build (if I ever find the time)! Thanks for suggesting :)

  • @736939

    @736939

    2 жыл бұрын

    @@SebastianRaschka No, I mean at the end stage when we produce classifcation, to use proportion between number of classes and the range between the min and max logits. To separate decision onto equal ranges between classes from min to max logit.

Келесі