16.10 Meeting Videos

16.10.1 Cohort 1

Meeting chat log

00:24:23    Daniel Chen (he/him):   PCA maximizes the variance
00:41:01    Daniel Chen (he/him):   I guess it depends on what you're using it for? like for a visualization or using PCA for feature engineering
00:42:01    Daniel Chen (he/him):   how useful is tuning the number of PCs? I've always looked at elbow plots or something for that stuff?
00:45:45    Daniel Chen (he/him):   run pca after LASSO! :p
00:45:52    Daniel Chen (he/him):   wait. that doesn't make sense
00:45:53    Daniel Chen (he/him):   nvm
00:49:15    Daniel Chen (he/him):   kind of surprised they didn't show other MDS (multi dimensional scaling) examples since PCA is a special case of MDS
00:50:14    Daniel Chen (he/him):   if you want to give names to "loadings" you'd use factor analysis
00:52:44    Jim Gruman: thank you Jon!!
00:53:13    Daniel Chen (he/him):   bye everyone!

16.10.2 Cohort 3

Meeting chat log

00:39:34    Ildiko Czeller: do you know from where the function name plot_top_loadings comes? I have not heard the term loading in the context of PCA before. As far as I understand it plots top most contributing variables/features in each PCA component
00:40:00    Jiwan Heo:  loading is the "weights" in PCA
00:40:46    Jiwan Heo:  the coefficient of the linear combination of variables
00:40:56    Ildiko Czeller: ahh, makes sense, thanks!
00:56:04    Ildiko Czeller: I guess difference between PCA and PLS would be bigger if there were some rubbish features as well with high variance but without much predicting power
00:56:21    Jiwan Heo:  PLS is supervised, PCA is unsupervised
00:56:25    Ildiko Czeller: the first 2 components seem to be basically mirror images of each other
00:57:38    Jiwan Heo:  rubbish features would not get picked up, i'd imagine. If it doesn't impact the outcome
00:58:30    Ildiko Czeller: yeah, I think that they would not be picked in PLS at all, but might be picked up by PCA because it is unsupervised (?)
00:59:15    Jiwan Heo:  I think so. PCA just picks up any large variance, but in PLS, it has to also move the outcome in some way
01:04:16    Jiwan Heo:  sorry I have to jump off! Thank you for the presentation :)
01:05:45    Ildiko Czeller: To build an intuition about UMAP i found this interactive website very useful: https://pair-code.github.io/understanding-umap/

16.10.3 Cohort 4

Meeting chat log

01:10:50    Isabella Velásquez: Gotta go, but thank you!
01:12:12    Federica Gazzelloni:    https://docs.google.com/spreadsheets/d/1-S1UbKWay_TeR5n9LkztZY2XXrMjZr3snl1srPvTvH4/edit#gid=0