davharris / rosalia

Exact inference for small binary Markov networks

rosalia: Exact inference for small binary Markov networks

Santa Rosalia, the patron saint of biodiversity, became a controversial symbol in the 1970's and 1980's for ecological inferences regarding species' interactions. This package, which provides tools from statistical physics for drawing such inferences, is named in her honor.

Installation:

The easiest way to install rosalia is with the devtools package, which can be installed from the R command line with install.packages("devtools").

Once devtools is available, rosalia can be downloaded and installed with devtools::install_github("davharris/rosalia")

rosalia's Issues

Editor 2: change focus from partial correlations to joint distribution

Editor 3: remove lines 84-96

Make sure alpha and beta have sensible (dim)names

Editor 10: clarify behavior at MLE

Was insufficiently clear about why MLE yields model with "exactly the observed occurence frequencies and co-occurrence frequencies". The answer is that the model matches these sufficient statistics at the MLE.

DESCRIPTION fails to declare an Imports dependency on package progress

In NAMESPACE you have

importFrom(progress,progress_bar)

Hence installing via devtools results in:

Downloading GitHub repo davharris/rosalia@master
from URL https://api.github.com/repos/davharris/rosalia/zipball/master
Installing rosalia
'/home/gavin/R/build/3.2-patched/bin/R' --no-site-file --no-environ --no-save  \
  --no-restore CMD INSTALL  \
  '/tmp/Rtmpx2g1CO/devtools7d27161d2ed3/davharris-rosalia-b6ad42f'  \
  --library='/home/gavin/R/build/3.2-patched/library' --install-tests 

* installing *source* package ‘rosalia’ ...
** R
** inst
** preparing package for lazy loading
Error in loadNamespace(j <- i[[1L]], c(lib.loc, .libPaths()), versionCheck = vI[[j]]) : 
  there is no package called ‘progress’
ERROR: lazy loading failed for package ‘rosalia’
* removing ‘/home/gavin/R/build/3.2-patched/library/rosalia’
Error: Command failed (1)

Reviewer 1-1: wants a more mechanistic simulation model

Might be able to address this by pointing out model misspecification when environmental conditions vary from site to site. Author seems to want abundances, though. Might be able to do that.

editor 5: fix paragraph indentation in latex template

Editor 4: minimize physics terminology

• Explain why "normalizing constant" doesn't work and we need to say "partition function." Explained a little bit in http://mlg.eng.cam.ac.uk/zoubin/papers/doubly_intractable.pdf

Editor 18: relevant literature

One reviewer wanted me to address JSDMs, the other wanted me to address some papers I hadn't seen from Colin Beale's group

rewrite README

Editor 9: prior distributions on alpha and beta

Editor is concerned that my priors are too informative.

One possibility is to use Empirical Bayes, but I'm not sure how that would work with this optimizer.

Need to find a better reference for logistic priors, or use something else. The Gelman paper mimics them with t-distributed priors, which isn't as on-point as it could be.

check .Rbuildignore

Reviewer 2-6: analysis of sensitivity and uncertainty

Reviewer 2-2: good references

The second one looks especially relevant

Aderhold, A, Husmeier, D, Lennon, JJ, Beale, CM & Smith, VA 2012. Hierarchical Bayesian models in ecology: reconstructing species interaction networks from non-homogeneous species abundance data. Ecological informatics, 11, 55-64.

Faisal M.A., Dondelinger F., Husmeier D. & Beale C. 2010. Inferring species interaction networks from species abundance
data: a comparative evaluation of various statistical and machine learning methods. Ecological Informatics, 5, 451-464.

Editor 13: clarify lines 194-202

can move to appendix if needed

Reviewer 1-2: missing references to the JSDM literature

Reviewer 2-4: The beta symbol in Fig 2A should be beta_12

Reviewer 2-1: use real data

Might not be feasible in 20 pages, but definitely worth looking into.

Editor 11: more "estimation" versus "inference"

Reviewer 2-5 (duplicate of editor 12?)

ridge parameter: switching to auto-estimation

Editor 17: conditioning simulations on all species present --> biases

Editor 16: add equations to Appendix section 1

Editor 15: "Markov"

Other names are as bad or worse:

• "Markov random fields" also have "Markov" in the name.
• "Boltzmann machine" & "Ising model" also aren't great fits. Boltzmann machines usually have latent variables and Ising models are usually on lattices. Neither of these has much currency outside of machine learning and physics, respectively.

Editor 7: 24 simulations is insufficient

Need to find a way to run these simulations without ignoring antivirus warnings or spending another $200 on EC2

Editor 14: clarify significance in 3-species results

Editor 8: doesn't like my simulations

Wants me to use the same parameters for all simulations. Says that only Bayesians would like my current approach? Need to investigate this further.

Editor 6: distinguish the two correlation approaches?

Editor asked for something about line 137, which seems to be a typo? Need to track down which line they meant.

Editor 1: re-order methods

• explicitly state the data needed, model structure, and estimation method
• estimation versus inference
• simulations at the end of the methods

Reviewer 1-3: clarify partial correlation's usefulness

Can get something like it from inverting the matrix in a JSDM, for instance, which means we can use existing tools and get a full posterior distribution.

Reviewer 2-3: Markov networks in spatial ecology

already mentioned, but should note it in response to reviewers

Editor 12: why ridge lambda value

just use the automatic estimation from corpcor instead