mallewellyn
commented
8 months ago Task list:
-
[x] 1. Line 22/Key points: I can't see where this last key point re computational speed is discussed. Potentially omit or add a brief comment about this in the text?
-
[x] 2. Line 46/Introduction: The end of this sentence could clarify the differences between the approach of this episode and the previous one. Something like: "that instead [what regularisation does and how it differs from information sharing]".
-
[x] 3. Line 79/Introduction: I like the example to show what happens when you fit a model using all features. Could add a sentence like "We explain what singularities are and why they appear when fitting high-dimensional data below" to just clarify that this is addressed!
-
[x] 4. Line 79/Introduction: Also, a brief explanation of why some effect sizes are very high as this doesn't seem to be addressed.
- [x] 5. Line 84/Singularities: I think it needs to be clearer why high-dimensional data result in singularities (i.e., explicitly saying that determinant of matrix is zero for high-dimensional data, thus singularities are always present when fitting models naively to high-dimensional data and R often cannot fit the model).
-
[x] 6. Line 87/Singularities: remove comma before "and why are they..."
-
[x] 7. Line 141/Correlated features -- common in high-dimensional data: Perhaps this summary of the challenges could be placed outside of the pinned box. I think it deserves more attention as it succinctly summarises both of the above points and motivates the need for regularisation.
-
[x] 8. Lin 141/Correlated features: Also, the fact that p>n is problematic is discussed in a lot of detail (resulting in singularities), and so I think should be added to this point to justify the use of regularisation (or else the discussion on high-dimensional data being problematic by its very size could be removed and just discussion of correlations retained)!
Something like: "Regularisation can help us to deal with correlated features." -> "Regularisation can help us to deal with correlated features, as well as effectively reduce the number of features (dimension) in our model, and thus addresses these issues".
-
[x] 9. Line 152/Challenge 1: "Discuss in groups:" could be "Consider or discuss in groups:" to make suitable for an individual learner.
-
[x] 10. Lines 175 & Line 311: Query - it feels, from the section above, as though you are about to describe regularisation as this is now fully motivated. These regression sections therefore confused me a bit. There's a similar interlude in the previous episode. Could it be possible to cover this as preliminary material at the start of the episode, reference another tutorial on linear regression + training and testing models and reference this throughout all episodes, or use this as a "bridging" episode at the start of the formal set of episodes and back-reference?
Could then provide the information re restricting the model after the example (where we try to fit a linear model on the whole data set) as further motivation and explain how this relates to generalisability.
-
[x] 11. Line 441/Using regularisation to impove generalisability: "This can be done with regularisation. The idea to add another condition to the problem we’re solving with linear regression." -> "This can be done with regularisation: adding another condition to the problem we’re solving with linear regression."
-
[x] 12. Line 512/Using regularisation to impove generalisability: "punished" -> "punishes"
-
[x] 13. Line 515/Using regularisation to improve generalisability: A brief few words about the trade-off here with the extent of regularisation would help to clarify.
-
[x] 14. Line 516/Why would we want to restrict our model?: "This type of regularisation is called ridge regression" makes it sound as though ridge regression=OLS. Slight re-wording to tie ridge regression to the non-zero penalty may help.
-
[x] 15. Line 518: Somewhere in this section, could possibly add a title to improve signposting of the ridge method (since there's one for lasso).
-
[x] 16. Line 520: These problems already feel explained and thus I would propose a minor re-wording to make this a recap and demonstration of the fact that regularisation actually works.
-
[x] 17. Line 548/Why would we want to restrict our model?: "trend" -> "tend".
-
[x] 18. Line 742/Cross-validation: Haven't defined what a 'good' value for lambda looks like (to define what the 'best' looks like)
-
[x] 19. Line 749: Cross-validation: "Cross-validation is a really deep topic that we're not going to cover in more detail today, though!" possibly a slight re-wording as the subsequent analysis seems to cover cross-validation in a bit more detail.
-
[x] 20. Line 756/Cross-validation: "We can use this new idea to choose a lambda value, by" -> "We can use this new idea to choose a lambda value by"
-
[x] 21. Line 795/Blending ridge regression and the LASSO - elastic nets: Would possibly re-order this title to put elastic nets first if also using lasso and ridge titles (just for signposting).
-
[x] 22. Line 798/Blending ridge regression and the LASSO - elastic nets: "So far, we've used ridge regression, where alpha = 0, and LASSO regression, where alpha = 1." -> "So far, we've used ridge regression (where alpha = 0) and LASSO regression (where alpha = 1)."
-
[x] 23. Line 1019/Other types of outcomes: It is a little unclear to me that the intercept-only model is selected. Could annotate the code to demonstrate how this happened.
alanocallaghan
Episode 3
Another nice episode. Although it's quite long, I think it covers what are often quite challenging ideas in a very approachable way. Most of the comments I have relate to how regularisation is motivated and some minor re-wording to clarify. I do, however, have a more challenging query about the placement of the linear regression background information. I have highlighted this in bold below!
Again, I will submit pull requests where possible and very happy to discuss anything.
Line 22/Key points: I can't see where this last key point re computational speed is discussed. Potentially omit or add a brief comment about this in the text?
Line 46/Introduction: The end of this sentence could clarify the differences between the approach of this episode and the previous one. Something like: "that instead [what regularisation does and how it differs from information sharing]".
Line 79/Introduction: I like the example to show what happens when you fit a model using all features. Could add a sentence like "We explain what singularities are and why they appear when fitting high-dimensional data below" to just clarify that this is addressed!
Also, a brief explanation of why some effect sizes are very high as this doesn't seem to be addressed.
Line 84/Singularities: I think it needs to be clearer why high-dimensional data result in singularities (i.e., explicitly saying that determinant of matrix is zero for high-dimensional data, thus singularities are always present when fitting models naively to high-dimensional data and R often cannot fit the model).
Line 141/Correlated features -- common in high-dimensional data: Perhaps this summary of the challenges could be placed outside of the pinned box. I think it deserves more attention as it succinctly summarises both of the above points and motivates the need for regularisation.
Also, the fact that p>n is problematic is discussed in a lot of detail (resulting in singularities), and so I think should be added to this point to justify the use of regularisation (or else the discussion on high-dimensional data being problematic by its very size could be removed and just discussion of correlations retained)!
Something like: "Regularisation can help us to deal with correlated features." -> "Regularisation can help us to deal with correlated features, as well as effectively reduce the number of features (dimension) in our model, and thus addresses these issues".
Line 152/Challenge 1: "Discuss in groups:" could be "Consider or discuss in groups:" to make suitable for an individual learner.
Lines 175 & Line 311: Query - it feels, from the section above, as though you are about to describe regularisation as this is now fully motivated. These regression sections therefore confused me a bit. There's a similar interlude in the previous episode. Could it be possible to cover this as preliminary material at the start of the episode, reference another tutorial on linear regression + training and testing models and reference this throughout all episodes, or use this as a "bridging" episode at the start of the formal set of episodes and back-reference?
Could then provide the information re restricting the model after the example (where we try to fit a linear model on the whole data set) as further motivation and explain how this relates to generalisability.
Line 515/Using regularisation to improve generalisability: A brief few words about the trade-off here with the extent of regularisation would help to clarify.
Line 516/Why would we want to restrict our model?: "This type of regularisation is called ridge regression" makes it sound as though ridge regression=OLS. Slight re-wording to tie ridge regression to the non-zero penalty may help.
Line 518: Somewhere in this section, could possibly add a title to improve signposting of the ridge method (since there's one for lasso).
Line 520: These problems already feel explained and thus I would propose a minor re-wording to make this a recap and demonstration of the fact that regularisation actually works.
Line 742/Cross-validation to find the best value of $\lambda$: Haven't defined what a 'good' value for lambda looks like (to define what the 'best' looks like)
Line 749: Cross-validation to find the best value of $\lambda$: "Cross-validation is a really deep topic that we're not going to cover in more detail today, though!" possibly a slight re-wording as the subsequent analysis seems to cover cross-validation in a bit more detail.
Line 795/Blending ridge regression and the LASSO - elastic nets: Would possibly re-order this title to put elastic nets first if also using lasso and ridge titles (just for signposting).
Line 1019/Other types of outcomes: It is a little unclear to me that the intercept-only model is selected. Could annotate the code to demonstrate how this happened.
Minor comments
Line 87/Singularities: remove comma before "and why are they..."
Line 441/Using regularisation to impove generalisability: "This can be done with regularisation. The idea to add another condition to the problem we’re solving with linear regression." -> "This can be done with regularisation: adding another condition to the problem we’re solving with linear regression."
Line 512/Using regularisation to impove generalisability: "punished" -> "punishes"
Line 548/Why would we want to restrict our model?: "trend" -> "tend".
Line 756/Cross-validation to find the best value of $\lambda$: "We can use this new idea to choose a lambda value, by" -> "We can use this new idea to choose a lambda value by"
Line 798/Blending ridge regression and the LASSO - elastic nets: "So far, we've used ridge regression, where
alpha = 0, and LASSO regression, wherealpha = 1." -> "So far, we've used ridge regression (wherealpha = 0) and LASSO regression (wherealpha = 1)."