🚨 New paper on ArXiv: “Uncertainty quantification and posterior sampling for network reconstruction” TL;DR; We present an efficient method to sample the entire ensemble of possible network reconstructions that are compatible with an indirect observation, e.g. a dynamics. Short thread: 1/N - ThreadSky

tiago.skewed.de • 84 days ago

🚨 New paper on ArXiv:

“Uncertainty quantification and posterior sampling for network reconstruction”

TL;DR; We present an efficient method to sample the entire ensemble of possible network reconstructions that are compatible with an indirect observation, e.g. a dynamics.

Short thread: 1/N

Comments

aminsense.bsky.social•59 days ago

📌

jrgant.bsky.social•84 days ago

📌

tiago.skewed.de•84 days ago

Like a rookie, I forgot to put a link to the paper in the first post!

https://arxiv.org/abs/2503.07736

ekernf01.bsky.social•84 days ago

This is really cool. I work on causal networks governing gene activity and I have been wanting to be able to sample from a posterior like this. Any idea how your posterior inclusion probabilities would compare to q-values from GeneNet? https://cran.r-project.org/web/packages/GeneNet/index.html

tiago.skewed.de•84 days ago

Posterior probabilities and p/q-values are fundamentally distinct. The later refer to the rejection of null models, while the former refers to uncertainty of inferred models.

I haven't compared directly w/ GeneNet, but what's in this paper should be more interesting and robust in a variety of ways.

ekernf01.bsky.social•84 days ago

ok, thanks for explaining.

tiago.skewed.de•84 days ago

Network reconstruction is needed when we do not have direct measurements on the network structure, only on the dynamics that it generates, or some other indirect data. It's an important problem, with direct relevance to ecology, neuroscience, epidemiology, and others.

https://dx.doi.org/10.1038/s41467-022-34267-9 2/N

tiago.skewed.de•84 days ago

The literature on network reconstruction is large and heterogeneous, including many problematic heuristics like correlation thresholding, and more principled approaches based on generative models and Bayesian inference. 3/N

tiago.skewed.de•84 days ago

One central aspect of this inverse problem is that it is ill-posed: in general, there are many networks that are compatible with an observation.

This can be partially dealt with regularization, i.e. finding the most parsimonious hypothesis from the set of possibilities: https://arxiv.org/abs/2405.01015 4/N

tiago.skewed.de•84 days ago

However, this approach, like the vast majority proposed for this inverse problem, still only produces a “point estimate” — a single reconstructed network — which does not convey any uncertainty or insight on the set of possible competing answers, leaving out a major part of the reconstruction.

5/N

tiago.skewed.de•84 days ago

One of the obstacles for network reconstruction is algorithmic complexity. Even for point estimates, existing methods could not run faster than O(N²) for a network of N nodes. Recently, we have been able to improve that to O(N log² N) with a more efficient algorithm: https://arxiv.org/abs/2401.01404 6/N

tiago.skewed.de•84 days ago

The current work leverages on those developments to produce *samples* from the posterior distribution of reconstructed networks, according to their plausibilities, in a manner that also works for larger problem instances. 7/N