Meeting chat log
00:04:25 Janita Botha: I am gatcrashing, if that's oke?
00:04:56 Jim Gruman: Welcome Janita
00:05:45 Jon Harmon (jonthegeek): Everyone is welcome!
00:05:54 Tan Ho: R song when
00:07:26 Conor Tompkins: We can’t risk not recording this one
00:14:21 Jon Harmon (jonthegeek): When you do a 5-fold CV on say a dataset of 100 elements (80 for testing and 20 for the estimation for each fold), do you create 5 sets of 20 elements and then just randomize the way you create the resampling CV datasets?
00:17:05 Conor Tompkins: Out of sample can mean different things for continuous vs. categorical variables
00:17:53 Julia Silge: https://applicable.tidymodels.org/
00:19:56 Jordan Krogmann: stats::filter
00:19:59 Jordan Krogmann: for time series
00:20:53 Emil Hansen: check_range() might also be nice to check for range of new observations https://recipes.tidymodels.org/reference/check_range.html
00:21:42 Tony ElHabr: I thought we were about to get a max v Julia fight
00:21:47 Tony ElHabr: get your popcorn ready
00:26:34 Tyler Grant Smith: you might use transformed forms of the original set of predictors
00:27:00 arjun paudel: you could use portion of original predictors and pca components
00:30:07 Daniel Chen (he/him): that sounds like the python 2 -> 3 transision xD
00:30:25 Jordan Krogmann: tidypredict??
00:32:06 Conor Tompkins: shinyapps.io works for basic stuff
00:33:06 Jordan Krogmann: @julia/max how do you handle the pipeline aspect of feature transformations in a sql database
00:33:25 Tony ElHabr: maybe use GitHub actions?
00:33:35 Asmae Toumi: I will blog the hell out of my process when the time comes (soon!)
00:33:39 Jordan Krogmann: it is an extreme pain to do some transformations as a sql f unction for new data
00:33:41 Conor Tompkins: Max what is your dog’s name?
00:35:26 Asmae Toumi: shaaaaaade
00:35:45 Jonathan Leslie: We use AWS to host shiny apps inside ECS (elastic container service?). Happy to discuss with anyone off-line if you’re curious.
00:35:52 Asmae Toumi: Bye Julia!!!! Thank you!!!!
00:35:56 Conor Tompkins: Thanks!
00:35:58 Daniel Chen (he/him): bye Julia!
00:35:59 rahul bahadur: Thanks Julia
00:35:59 Jonathan Trattner: Thank you!!
00:36:08 Jordan Krogmann: lator tator
00:36:13 pavitra: thanks Julia!
00:36:19 Tyler Grant Smith: 42!
00:36:21 Daniel Chen (he/him): 42!
00:36:25 Tan Ho: It's not 42
00:36:25 Daniel Chen (he/him): :D
00:36:29 Tony ElHabr: 6669
00:36:31 Daniel Chen (he/him): 4242
00:36:41 Tan Ho: 834 I think?
00:36:45 Emil Hansen: I would not trust 2020!
00:37:05 Max Kuhn: https://twitter.com/TechRonic9876/status/1247743194038546432
00:37:40 Tan Ho: lolol I was in the area code
00:38:02 Jordan Krogmann: birth year
00:38:05 Jordan Krogmann: everytime
00:38:40 Tyler Grant Smith: but what do you seed your random seed generator with...
00:39:00 Jordan Krogmann: birth year number... inception
00:39:07 Conor Tompkins: @max what is the simplest feature engineering step that has given you the most performance gain?
00:39:43 Tan Ho: ^ bonus points for something aside from "log transform everything"
00:40:31 Jordan Krogmann: https://datachef.co/
00:41:20 pavitra: normalize
00:41:47 Tyler Grant Smith: preprocess
00:41:49 Emil Hansen: My mental model is:
recipe() -> define
prep() -> estimate
bake() -> apply
00:42:58 Joe Sydlowski: Are the MARS models the only models with feature selection in tidymodels today? Do you have any concerns with using the feature selection in MARS when resampling?
00:43:15 pavitra: once I discovered workflow I was delighted to not mess with bake and prep
00:43:48 Jon Harmon (jonthegeek): Can you talk about the design process of metric sets as functionals, e.g. ames_metrics <- metric_set(rmse, rsq, mae), then calling ames_metrics(...)? The idea of using functionals is interesting, but what are the actual practical advantages?
00:47:05 Tony ElHabr: it is quite elegant
00:49:26 Jon Harmon (jonthegeek): https://github.com/tidymodels/workflows/issues
00:53:19 Tyler Grant Smith: I like both modeltime and fable...
00:54:13 Jon Harmon (jonthegeek): Three questions really. Firstly what package system do you think works better for time series in a tidymodels workflow for time series, Modeltime, fable, or other? Second question, this is possibly just my experience, but I often feel time series is poorly taught/documented by comparison (excluding FPP) to other techniques. I was wondering how you might think this could be improved upon, and how tidy models system may contribute to better understanding and comparison? A third question (cheeky sorry) a lot of time-series focuses on model fit, however, sometime we want to investigate potential uplifts based on the presence or absence of features across multiple future forecasts. In which case causal inference can be a useful option, is this something we would be able to do in tidy models?
00:57:13 arjun paudel: is there ever a reason to normalize dummy variables?
00:57:37 Conor Tompkins: Kenneth can you mute
00:58:20 Tan Ho: there was a mars question
00:58:30 Tan Ho: > Are the MARS models the only models with feature selection in tidymodels today? Do you have any concerns with using the feature selection in MARS when resampling?
- Joe
00:59:06 Jordan Krogmann: I think workflows is the future
00:59:07 Tyler Grant Smith: is there a way to do mean (target) encoding in recipes?
01:00:35 Conor Tompkins: Is that like putting tune() inside a step_?
01:00:59 Jordan Krogmann: anything on deployment @#max
01:02:33 arjun paudel: I saw "workflowsets" mentioned in tidyAPM repo...is it making to CRAN anytime soon?
01:02:37 Tyler Grant Smith: @conor no, it wouldn't really be tuned. It is learned from the data
01:03:21 Tyler Grant Smith: :thumbsup:
01:04:22 Tyler Grant Smith: does workflowsets process in parallel?
01:05:36 Jordan Krogmann: thanks max!
01:05:37 Jim Gruman: Thank You!!!
01:05:39 Conor Tompkins: Thanks max!
01:05:40 Joe Sydlowski: Thanks Max!
01:05:41 Jonathan Trattner: Thank you so much!!
01:05:43 rahul bahadur: Thanks Max
01:05:47 Tan Ho: Thank you :D
01:05:50 Asmae Toumi: Thanks max!!!!!!!!!!!!!!
01:05:51 pavitra: this was amazing! thanks max and Julia!
01:05:55 Jonathan Leslie: Thanks, Max
01:05:56 arjun paudel: Thank you!!
01:05:56 Andy Farina: Thank you!