My Patreon : www.patreon.co... www.flaticon.c... Carrot icons created by Freepik - Flaticon www.flaticon.c... Fruit icons created by Pixel perfect - Flaticon
I recently tried to explain something of this sort to colleagues as a fishing net trying to catch some kind of medium sized fish. The size of the gaps in the net needed to guarantee you never miss a single fish you were actually trying to catch, happens to be small enough that plenty of other marine life gets caught too. This lowers the quality of the catch and introduce the costs of having to pick through when processing, potentially even ruining some batches when the wrong thing gets scooped up. On the other hand, the size of gap needed that guarantees you never catch anything besides what you want, is large enough that you will also occasionally lose out on the catch of fish you set out for. In my real work example, we can afford to have a bit of junk mixed in that gets picked out, but we can’t afford to miss anything, so I went with the smallest net-gap size I could that still never missed the legitimate targets.
Could you make a video on useful resources like blogs for inspiration? Also a kaggle series would be 😎 You are the best teacher of advanced statistics topics on YT ♥️
Absolutely love the video! Would be interested to see a code with me on the topic. How does one go about exploring the precision recall frontier? Is it just hyperparameter tuning via grid search, or a more deliberate method I'm not aware of? I have a neural net trained with decent accuracy for what I want, but it deals with stocks so I'd much rather have no signal than a false signal. Not necessarily asking for neural net, could be logistic regression or random forest, just more clarity on the question would be wonderful! Thanks for all you do man!!
“How does one go about exploring the precision recall frontier?” is an excellent question. If the question is more about how we construct the frontier, usually that is done in binary classification problems by varying the threshold for marking some example in the positive class. Low thresholds lead to strong recall but poor precision and vice versa for high thresholds.
@@ritvikmath that was my question and your answer makes good sense, thanks! I'll iterate over different classification thresholds to establish the curve and then pick the level that works best for me :)
Hi Jesse, my understanding is that while searching for the best model, you need to hyper-parameter tune for area under PRC (or equivalently AUC ie Area under the ROC curve), which is independent of the choice of threshold probability. However given a model with a fixed AUC or area under PRC, you need to find the threshold that suits your problem description. Hope that makes sense.
@@abhigyandatta2008 for sure, thanks! Make sure I've got the best model going in, then and only then test for an appropriate threshold. Makes sense to me!
I think marketing cost n market response is a better example for law of diminishing marginal returns. This carrot apple thing doesn't feel too intuitive isn't it? Also, the precision recall curve may not really be a good example of law of diminishing marginal returns too. Precision recall curve's shape will be a function of class balance as well i think
Totally a valid point. I think “most important curve in data science” depends on lots of things so this is my take on it. I feel this one’s important to understand higher level data science trade offs but bell curve is certainly crucial for understanding things like statistical behavior in the presence of large sample sizes for example