Creating ROC curves and ensembling models in R with "caret" | R Tutorial (2021)

[@fullsurr3465]

The part of the video with ensembles was really helpful to me. Thanks!

[@RichardOnData]

Yes, it's very powerful - and also a great way to land yourself knee deep in some code that takes 4 hours to run!

[@djangoworldwide7925]

I watched the whole series. Fantastisch!

[@shaoru]

very helpful series. Thanks!

[@HaiLeQuang]

It's a really pity your channel doesnt get more views. But keep up the good work

[@spencerantoniomarlen-starr3069]

I think most people either dramatically overthink (or perhaps underthink) the tradeoffs for when to favor TPR, TNR, PPV, NPV, and Accuracy (not counting the more advanced/synthetic metrics like Kappa, AUC, or F1). It is actually relatively simple. Is the potential upside benefit of getting your positive predictions correct larger than the potential (the range) downside risk of getting your positive predictions wrong, i.e. the payoffs of the TPR vs the TNR. If for instance, you are trying to predict future stock prices to assess which stocks you should buy naked short positions on, then the TNR (aka specificity) is WAY more important than the TPR (aka sensitivity) because if you are wrong, the potential losses from the stock price going up instead of down are unbounded, that is, there is no upper limit to the price a stock can reach. Similarly in the other direction, if you essentially have a known discrete or known limited potential downside or cost, and the potential upsides are unbounded (or even just have a much wider range), then the TPR is more important. How to choose between PPV and TPR is even easier by the way! It just depends what you want to do with your output and in what form it comes. If you are taking some sort of disease screening test you bought or administered to you by your doctor and you get a positive reading and you'd like to calculate the probability that you have that disease (assuming you can find out the population base rate and somehow adjust for how much of an amplification the fact that you have symptoms should impose on that base rate), always go with the Positive Predictive Value (aka Precision). That is literally the common language definition of it actually!

[@minhaoling3056]

Hi, is it possible to combine three layers of feature extraction method and one model layer into only a model object ?