To do - Lasso approach
bniebuhr opened this issue · 4 comments
Preparation:
- function to add strata in RSF setup, for sampling purposes
- Implement function
explore_blocks()
Fitting:
3. Predictor standardization: implement unstandardization of coefficients in bag_models()
or bag_fit_net_(c)logit()
if standardize = "external"
4. put cumulative type (nearest, cum), zoi shape, infrastructure type, and radius as input
5. Concordance: add continuous Boyce index/SomersD - for no strata present
6. Go for the clogit implementation to be coherent with the logit - all functions
After
7. Variable importance: what if removing variables leads to higher validation score than the full model for drop = TRUE?
8. Separate bag_plot_response
into two - prediction and then plotting - more customizable, and also a fast plot like it is
9. weighted quantile prediction is not working
10. Variable importance vs weighted coefficients - how to interpret? What does the variable importance mean?
11. How to plot effects of all ZOI (both cumulative and nearest) by incorporating the number of features?
General:
12. Define final datasets in the package (rsf, SSF) and document them - check data contributors as well
plot_coef:
- add possibility to rescale the coefficients (standardize them) within the function.
- add different types of plots (histogram, bar plots, barplot side-by-side)
- add order to terms, both type of infra/variable and ZOI radius
variable_importance:
- add recognition of type of validation (habitat or not), and habitat only prediction for evaluating variable importance
Add function to make a spatial prediction
- Function to add strata was included in a separate issue: #21
- Function
explore_blocks()
included in 0de9212 and improved in 8a5c499 - Predictor standardization: implement unstandardization of coefficients in bag_models() or bag_fit_net_(c)logit() if standardize = "external": included in a separate issue: #22
- Not sure what was that, ignored for now.
- Boyce and SomersD without strata - not included now, include if required
- Correspondence of logit - clogit functions: they were kept correspondent in a series of different commits. One function to be added to the logit context: bag_load_net_logit (could be the same?). Opened in a different issue in #23
- Variable importance: ignored for now, check later
- Separate
bag_plot_response
into two: it is now possible to return either the plot or the dataset ifggplot = FALSE
. - Weighted quantile prediction now working, but it is highly variable
- Variable importance is proportional to the standardized coefficients, to be shown in a supplementary material in the methods paper.
- How to plot effects of all ZOI (both cumulative and nearest) by incorporating the number of features? Good question to be discussed, opened in a new issue now in #24
- Documentation of SSF and RSF datasets opened in a new issue in #25
- plot_coef:
- variable_importance: add recognition of type of validation (habitat or not), and habitat only prediction for evaluating variable importance: opened in a different issue in #26
- bag_predict_spat: first implementation for
data.frame
in 0d4257c