- Itai's comments
- Proper English title pages
- Abstract
- Final formatting
- Include Hebrew title pages and abstract
- Introduction
- Further Research
- Conclusion
- changes to theory and method based on comments
- Literature: first draft
- Theory: (and everywhere) lambda_i used for intensity AND for index: change intensity to lambda_I/lambda_P
- Theory: 2.2 describe N(ds) distribution as Poisson Process; define PP
- Theory: 2.2 lambda(x,x)ds approx prob: derive (from Bernoulli/Poisson)
- Theory: 2.3 do 1D before 2D and move 1D plots after 1D
- Theory: 2.3 1D plot - specify which kernel
- Theory: 2.3 for biweight explain radius is 1 but can be changed by bandwith
- Theory: 2.5 explain K has mean 0 and 3X differentiable with positive variance and radial symmetry
- Theory: 2.5 assumptions on f
- Theory: 2.5 replace section A2 with ref to Silverman and clean up derivations
- Theory: 2.6 define oracle bandwidth and add to glossary
- Theory: 2.7 implications of inconsistency (error increase) so we normalize
- Theory: 2.7 verify n^{1/3} as others state O(1)
- Theory: 2.8 remove this section and mention Dalenius in method
- Theory: 2.9 mention rejection sampling and give reference
- Method: better describe centroid
- Method: 4.1 describe the study area
- Method: 4.1 put in a picture of pop and incident points
- Method: 4.1 Discuss the real-life underlying story of the data with examples
- Method: 4.1 Discuss how the real-life underlying story is modeled with SPP or Poisson
- Method: 4.1 Describe the kernel method of intensity estimation, assumes some data, "if someone has incident location data, they can use this method"
- Method: 4.1 mention 2 kernel methods, incidents vs population
- Method: 4.2 define lambda := lambda(., .) to show lambda is a function
- Method: 4.2.5: motivation for centroid
- Method: 4.2.5: formulas
- Method: 4.3: "and scale it" - example
- Method: 4.4: Add formulas and a reference for CV
- Method: 4.5: Explain buffer with respect to edge effects
- Method: 4.5: use \texttt for variable names
- Method: 4.5: explain that the oracle knows the true lambda and so creates a baseline that approximates mise-optimal b/w
- Method: 4.5: add a step to compute lambda_p
- Method: 4.6: reword last sentence
- Method: Add computing and technical issues and solutions: time, AWS, etc. move from wherever.
- Theory: finished draft
- Discussion: finished draft
- Derivations: bias-variance
- Results: account for comments
- Results: all subsections: compare peak, centroid, etc in each section briefly
- Results: all subsections: compare selectors.
- Results: all subsections: make more clear MISE chart discussion is for all selectors, etc
- Theory: continue major rewrite
- Theory: major rewrite
- Literature search: start
- Results: text mentions colours
- Discussion: major rewrite
- Theory: major rewrite
- Method: double integral with (x1, x2) in W as limits
- Method: 4.2.1: "in many cases": when & how?
- Method: 4.4: "A common method for bandwidth selection..." to start
- Results: (and everywhere) in population and incidents scatter use a different size/symbol instead of colour
- Results: make all charts b/w or grayscale (not colour)
- Results: only 2 charts per row
- Results: bandwidth histograms: lighten fill; why shadow in printed?
- Discussion: Create Boxplot of MISE distribution comparing Oracle to Silverman to CV selection, possibly OTHER accuracy measures
- Discussion: Create Boxplot of MISE difference between S-O and CV-O (MISE only)
- Discussion: overall plot of everything e.g. NMISE of CV, Silv, Oracle vs. experiment number
- Appendix: Section A.10 title should say peaks are NOT in same place
- Appendix: Section A.8 mention in subsection title that peaks are in same place
- Results: major rewrite
- Results: 5.1 mentions 5.6-7 but not 5.2-4
- Appendix: some tables are squished. Paragraph indentation?
- For NMIAE and NSUP - use mu not mu^2
- Results: Figure 5.10 "empirical distribution of
MISEandRMISE" - Define h_opt
- Results: 5.2 sample size: n = actual number of incidents; state that we fix the expected but observe actual
- "error fell with increasing
the multiplication expected number of incidents" + "mu" - 5.2/5.3 add CV bandwidths to tables
- 5.2/5.3 split titles to 2 rows
- 5.2/5.3 add "mean" to h_o, etc.
- Method: Describe subsections 4.1-4.4 and 4.6.
- Results: 5.1 drift needs to be seen in relation to square
- Results: convergence rates
- 5.2/5.3 describe what's in the tables and make a plot (log-log?)
- "negative polynomial order" <-- check
- Method: explain the units of distance
- Method: mention no edge effect compensation
- Method: mention parallelization and randomization algorithms and R packages, and AWS including which instance types and other details of execution
- Method: more rigorous math
- Method: add "relative" and "normalized" error measures
- Results: 5.1 mention that population and incident bivariate normal is independent with equal variances (and move to Method)
- Results: 5.1 mention the actual size of the study area (and move to Method)
- Method: describe the data generation process
- replace "decay rate" with sigma_p, sigma_i, etc and "spread"
- No headers on pages
- Method: use "approximate" instead of "estimate" for MISE, MIAE etc.
- Method: use \widetilde instead of \hat on MISE, MIAE, etc.
- Discussion: move 5.8 to 6
- Conclusion: when is Silverman better than CV?
- use (x_1, x_2) instead of vector x
- Results: How does spread affect bandwidth?
- Appendix: table 36 oracle is worse than Silverman?
- Explain error measures with formulas
- subcaption titles are too wide
- appendix: subcaptions on top of tables
- subcaption: "true risk function
distribution" - factor see if another term is common otherwise use expected number of incidents