Differences between revisions 1 and 2
⇤ ← Revision 1 as of 2017-01-12 20:02:29 →
Size: 349
Editor: DavidOwen
Comment:
← Revision 2 as of 2017-01-15 00:11:07 → ⇥
Size: 1292
Editor: DavidOwen
Comment: Fleshed out notes
Deletions are marked like this. Additions are marked like this.
Line 7: Line 7:
 * Where (in the domain of the data) is the model useful?
 * Prediction versus inference. One opinion is that a good prediction model also ought to be explainable like a good inference model.
 * Parametric versus non-parametric. Parametric is great whenever you can determine the shape of your data; otherwise, non-parametric may be the only way to go.

Had a side-discussion about Lasso. David had read a paper demonstrating that using Lasso to select the best variables to include in a model uses one degree of freedom per variable, in order, within about 6 significant figures. This is contrary to the belief that selecting variables would be like conducting consecutive experiments, and require some ''p''-value adjustments (see the Green Jelly Bean XKCD comic).

Back to chapter 2:

 * ISLR's treatment of model accuracy in train versus test is top-notch.
 * Its discussion about the bias/variance trade-off is also good.

Jan 12, 2017

/Attendees

Review of Chapter 2

Chatted about various highlights of chapter 2. Went through several of the R examples that we hadn't seen before, especially contour, persp, and identify.

  • Where (in the domain of the data) is the model useful?
  • Prediction versus inference. One opinion is that a good prediction model also ought to be explainable like a good inference model.
  • Parametric versus non-parametric. Parametric is great whenever you can determine the shape of your data; otherwise, non-parametric may be the only way to go.

Had a side-discussion about Lasso. David had read a paper demonstrating that using Lasso to select the best variables to include in a model uses one degree of freedom per variable, in order, within about 6 significant figures. This is contrary to the belief that selecting variables would be like conducting consecutive experiments, and require some p-value adjustments (see the Green Jelly Bean XKCD comic).

Back to chapter 2:

  • ISLR's treatment of model accuracy in train versus test is top-notch.
  • Its discussion about the bias/variance trade-off is also good.

Next Time

MeetingNotes/2017-01 (last edited 2017-01-15 00:11:07 by DavidOwen)