Post AQL3DoSdeEAl63G29Q by tiago@social.skewed.de
(DIR) More posts by tiago@social.skewed.de
(DIR) Post #AQKkFbyl806hpoQTBo by jacomyma@mas.to
2022-12-06T13:24:29Z
0 likes, 0 repeats
Nice point by @GrandjeanMartin : one of Moreno's sociogram has something hidden that you see if you use a modern layout.Moreno separates boys and girls and indeed they are poorly connected. But there is a group of girls that bridges them, and is equally connected to boys and other girls.A better visualization requires less assumptions.
(DIR) Post #AQKkFcXr1WBDaeqUtM by tiago@social.skewed.de
2022-12-06T15:09:49Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin Mathieu, this is the kind of thing I'm talking about.Why is a three group division “better”? What criterion is being used?The 2D layout? Do the nodes really live in this space? And if they did, would the force-directed algorithm be able to recover it accurately?If I use a statistical inference method for this network, I detect only 2 communities.How can you be sure that you are not overfitting, and Moreno was right all along?Btw, this dataset is now available on @netzschleuder for people to play with:https://networks.skewed.de/net/moreno_sociograms
(DIR) Post #AQKpOQFAd1qVoc7fyi by jacomyma@mas.to
2022-12-06T16:07:25Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder Moreno chose to use the gender for his visualization. As a result, we can see the relative lack of links between boys and girls, but that is all.The modern layout allows to see that there two groups of girls, one of which has no connections with the boys.This is an objective description of the network, that does not depend on the visualization. But not every visualization allows to see it.
(DIR) Post #AQKqMVlZE05eJmve8O by jacomyma@mas.to
2022-12-06T16:11:27Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder The groups I picked here are such that the kids are more connected to the kids of the same group than to the kids of other groups.So you could see those groups as assortative but it does not mean that other groupings are invalid. I only use the groups to write a description.
(DIR) Post #AQKqMWIXFQSfy2LyWO by tiago@social.skewed.de
2022-12-06T16:18:16Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin @netzschleuder What is the objective of your description? Are you not trying to draw an interpretation of some kind? Do you expect the girls that belong to the third group to have different characteristics from the other ones?You are comparing it to Moreno's, and arguing it is better. But what was Moreno trying to do with his?If Moreno was interested in the probabilities of links forming — homophily — then arguably his layout would have been more faithful, and yours would be less.
(DIR) Post #AQKrNG3TDxmTpnDs6i by jacomyma@mas.to
2022-12-06T16:29:38Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder A pretty big if! Facts don't care about Moreno's intent (or mine). The description is what it is. It does not exist as a function of a goal.Then when the description is convoked in an interpretation, what happens depends on that interpretation. It's on the interpreter, not the description.I only described here. It's a reasonable fact that some things are not visible in Moreno's viz.
(DIR) Post #AQKtimg3KXMF04f4oS by tiago@social.skewed.de
2022-12-06T16:55:57Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin @netzschleuder I appreciate the delineation you are making — and see why you are making it — but I actually don't think this a reasonable stance.I'm assuming we are attempting to do science, where we do not merely describe the facts, we have also to explain them.In fact, something I learned in my very first semester in the lab sessions as a physics undergrad is that we don't even have access to facts, only to measurements, which are always uncertain. So, in physics (and in the natural sciences by extension), we are always forced to interpret to some degree — there is no such thing as a purely objective description. I have a hard time believing this is different in the social sciences and humanities.Also, its always possible to find more complex descriptions hiding in simpler ones. I could describe Moreno's network according to preference patterns between 7 groups of girls and 3 of boys, and claim these were hidden in your layout, which only cares about assortativity. It would be a true fact. But so what?I think your description is suggestive of an alternative *interpretation* of the data — and this is its true appeal.
(DIR) Post #AQL0ZqANgh6otWWuIa by jacomyma@mas.to
2022-12-06T18:12:46Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder Objectivity is never pure: agreed.Multiple equally objective descriptions coexist: agree.Descriptions come with suggestions: agreed. One cannot extirpate the suggestion out of the expression.But that does not mean that there is a hidden agenda. I'd say we just live with the annoyance of the implicits of the layout, that we use to describe because it begs questions, yet not to answer them.
(DIR) Post #AQL2HYLfOaPOKSDSoS by tiago@social.skewed.de
2022-12-06T18:31:53Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin @netzschleuder I know I'm just pestering you at this point, but what if we don't have to live with the annoyance? What if there is a methodology that allows the implicits to become explicit, so we can reason about them, and make informed comparisons? What if some of the methods are already available?Relatedly: There is a kind of hidden functionality in @graph_tool that I intend to expand and explore: start with an arbitrary clustering, and provide a force-directed layout that respects it. 🤯
(DIR) Post #AQL2gV1ct8ZvGHjnqy by jacomyma@mas.to
2022-12-06T18:36:23Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder @graph_tool The problem for me is that the clustering unnecessarily enforces boundaries that do not exist. The Alps exist without a border, so do topological clusters. The limit of the Mont Blanc is an artifact.
(DIR) Post #AQL3DoSdeEAl63G29Q by tiago@social.skewed.de
2022-12-06T18:42:23Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin @netzschleuder @graph_tool This is *also* a hypothesis that can be reasoned and verified under the very same framework. There are even papers about it! And written code! And even some visualization techniques — although they are not quite complete. Lot's more to do.Remember: If I give you a network where the hard clusters *do* exist, because I've put them there, you will still think the network looks like the Alps if you view it via a force-directed layout — it can't represent anything else.
(DIR) Post #AQL3sbaCQtuN1rFHY8 by jacomyma@mas.to
2022-12-06T18:49:44Z
0 likes, 0 repeats
@tiago @GrandjeanMartin @netzschleuder @graph_tool Oh but remember that I disagree on that very last point. Just mentioning it for completeness. I'm too tired to continue today!
(DIR) Post #AQL44d5qdutFBkmZoO by tiago@social.skewed.de
2022-12-06T18:51:57Z
0 likes, 0 repeats
@jacomyma @GrandjeanMartin @netzschleuder @graph_tool Then when you are well rested, come back and explain why you think hard clusters in networks are an impossibility.
(DIR) Post #AQM7NP4XUVuYE2m2dc by jacomyma@mas.to
2022-12-07T07:03:40Z
0 likes, 0 repeats
@tiago You give me a network. Although I need to know where it comes from to interpret it, I do not to describe it.In the version of Dune written by random typing monkeys, the Emperor still betrays Leto Atreides. The text is the same, and we can describe it.
(DIR) Post #AQMFUHU4tvlILsVg2K by jacomyma@mas.to
2022-12-07T07:03:56Z
0 likes, 0 repeats
@tiago Borges has a story where a dude named Pierre Menard rewrites the Quixote to the exact word. It's funny because the exact Quixote text in that version means something different, precisely because it is written by someone else.The text is the same and the meaning is different.The description is the same and the interpretation is different.
(DIR) Post #AQMFUI26rOz43QQr56 by jacomyma@mas.to
2022-12-07T07:04:41Z
0 likes, 0 repeats
@tiago You give me a network that comes from a generative process where you have planted a certain structure. We must interpret it a certain way. You've made that case and I agree.
(DIR) Post #AQMFUIYiu94VgZgtuq by jacomyma@mas.to
2022-12-07T07:05:16Z
0 likes, 0 repeats
@tiago I find a network and you come to me with the exact same network. You come and you tell me that my interpretation is wrong because in your exact same network come from a place where my interpretation makes no sense. But in fact, your interpretation only applies if my exact same network comes from the same place as yours, which is does not. Your network is written by Cervantes and mine by Pierre Menard, and they mean the same thing.
(DIR) Post #AQMFUJ4czWanHWcNe4 by jacomyma@mas.to
2022-12-07T07:05:52Z
0 likes, 0 repeats
@tiago Just because you plant hard clusters and generate a blended network does not mean that blended networks necessarily hide hard clusters. Even the exact same networks.
(DIR) Post #AQMFUJZ7AAyko4sjAG by tiago@social.skewed.de
2022-12-07T08:34:31Z
0 likes, 0 repeats
@jacomyma I'm confused by your argumentation. We already agreed that we can describe many different facts about data, independent of any explanation for them. But you said:“The problem for me is that the clustering unnecessarily enforces boundaries that do not exist.”How can the boundaries not exist at purely descriptive level? I can divide any network into groups and point to them — the division is as factual as any other representation, including the 2D-layout.And what necessity makes them unnecessary?I had assumed you were making a point about interpretation, because at a purely descriptive level we have a near complete anarchy: we can choose any of the trillion facts that our data contains.Also, since we already agreed that descriptions are suggestive of interpretations, I don't understand your staunch phobia to making them explicit, or at least owning up to them.The example you gave — monkeys typing Dune — is unnatural, and at some point we need to account for this.In the past, you have said that the examples I usually give of the face on mars, Jesus on toast, shape of clouds, etc, are an attempt of ridiculing, but they are not: they illustrate exactly the argument you are making. Just like seeing structure in random text, clusters in networks, we can argue that a cloud looks like a pirate ship. And no one can disagree at an objective level.I see no substance in mere descriptions. We need theory. And to quote Francis Crick: “A theory should not attempt to explain all the facts, because some of the facts are wrong.”
(DIR) Post #AQMFUKzNrzaPDqBDkW by jacomyma@mas.to
2022-12-07T07:04:17Z
0 likes, 0 repeats
@tiago You give me a network. Two given nodes are, or are not, connected. That is just a fact, an objective description. I do not need to know where the network comes from.With that network either I can make groups of densely connected groups, or I cannot. Objective description. But what it means depends on where the network comes from.
(DIR) Post #AQMYd6OdOQ8dhOspvs by jacomyma@mas.to
2022-12-07T12:09:00Z
0 likes, 0 repeats
@tiago Groups are a useful trick to describe, but not inherently necessary. Layout-based descriptions have one less unnecessary thing than group-based ones. At least where no groups are given a priori, and although they have other issues.A Bayesian inference with a group-less model is possible, I assume. But even just formulating this model is hard. I've tried and failed. So far.
(DIR) Post #AQMa8cKWoVvLffYQUq by tiago@social.skewed.de
2022-12-07T12:25:58Z
0 likes, 0 repeats
@jacomyma Descriptions that are more general — i.e. have fewer constraints — are more prone to overfitting. Constraints are *necessary* to interpretation.If you do a force-directed layout of a sparse ER network you will always see lots of interesting structures — all of which are spurious from a generative perspective, since the nodes are all identical.There is a big literature of inferential latent-space models. The most canonical (but rather old) reference is from Hoff et al: https://doi.org/10.1198/016214502388618906I even have one with Mark Newman from several years back, which combines it with some ideas from the SBM, while preserving continuous positions: https://dx.doi.org/10.1103/PhysRevLett.115.088701Bayesian model selection between SBM-like descriptions and latent space models is possible.
(DIR) Post #AQMfXbaFE44WMldjxA by jacomyma@mas.to
2022-12-07T13:26:29Z
0 likes, 0 repeats
@tiago The drawbacks of the layout do not invalidate its benefits.