Continuing development of ZIPLN models

[jchiquet]

• Exact optimization with Lambert function for $R$ (parameter of the conditional distirbution of the ZI component)
• Network-like methods for ZIPLNfit_sparse
• ZIPLNnetwork, a
  • collection with PLNfamilly collecting a series of ZIPLNfit_sparse object with automatic grid
  • merge/share classes between PLNnetworkfamily and ZIPLNnetwork

Still missing

• use correct weights in estimation of $\pi$ when using exact distributino of $W$ (with Lambert function)
• complete test for ziplnnetwork

Note that I won't add postTreatment yet zo ZIPLN since I think we need to rethink a bit all the psotTreatment made on PLN-lie objects(in particular, all the variance estimators...)

[mahendra-mariadassou]

Apart from the 3 small comments:

• inception_cov to constrain the shape of the covariance matrix in the inception model
• the remark about lapply / future.lapply
• number of parameters for ZIPLNnetwork (the extra p is indeed present in PLNnetwork)
  It's all good for me.

[mahendra-mariadassou]

Introduced in ctrapnell's PR. Allows the user to:

• specify a custom inception with inception
• constrain the shape inception_cov of the covariance in the (yet-to-be fitted) inception model used to initialize members of the family among full (default, same as before), diagonal, spherical
  Useful for large dataset as the "full" model used for initialization (corresponding to $\lambda = 0$ can take a long time to fit.
  inception_cov is ignored is inception is provided.

[jchiquet]

noted, thx.

[mahendra-mariadassou]

stability_selection() uses:

• future::lapply() for PLNnetwork but
• lapply() for ZIPLNnetwork
  Should we switch to lapply() everywhere ?

[jchiquet]

I used lapply in place of future_lapply for debugging (the later suppress the output which is not very convenient for debuging). For the moment, I suggest we use future_lapply for all parallel computing, and we will update this particular instance once we will have our complete rethinking of all parallel computing stuff.

[mahendra-mariadassou]

With this formula, we don't count diagonal coefficients of $\Omega$. Although they are not (always) penalized, they should still be treated as parameters when computing e.g. BIC ?

[mahendra-mariadassou]

Now applies to

PLNmodels/R/ZIPLNfit-class.R

Lines 804 to 806 in 36123fd

	nb_param_pln = function() {
	as.integer(self$p * self$d + self$n_edges)
	},

[mahendra-mariadassou]

All good for me (and I should indeed have looked at it sooner...)