Computing desk
< August 21	<< Jul \| August \| Sep >>	August 23 >

Welcome to the Wikipedia Computing Reference Desk Archives
The page you are currently viewing is an archive page. While you can leave answers for any questions shown below, please ask new questions on one of the current reference desk pages.

August 22

ModelsPredictive models that are not built through machine learning?

How do you call ~~computer~~ predictive models that are not obtained through machine learning? In the sense that they are not built based on examples that are generalized by a computer program. Llaanngg (talk) 17:38, 22 August 2016 (UTC)[reply]

Why not start at our main article, Model, and read through the many articles that are linked? If that's not sufficient, Wiktionary has entries and a thesaurus.

Nimur (talk) 18:28, 22 August 2016 (UTC)[reply]

In my opinion, this is like asking "what do you call an animal that does not have wings?" But maybe I'm missing something. I'm also not clear on what you mean by "computer model". For example, I can set up a model of the solar system with an appropriate set of differential equations. I may choose to use a computer to calculate certain solutions to the model, but the model ultimately has nothing to do with computers. There isn't really any kind of model that intrinsically depends on a computer. There are models that aren't that useful if we don't have the assistance of a computer, but they are still models, even if they only exist as scribbles on paper.

Anyway, if you can clarify what you're after, we might be able to help you better. Machine learning is ultimately a very small class of techniques, while the concept of modeling is vast. SemanticMantis (talk) 19:11, 22 August 2016 (UTC)[reply]

Ok, that was too broad.

I get from the model article linked above, that lots of things are models. I am more specifically interested in predictive modelling. According to the article " predictive modelling is synonymous with, or largely overlapping with, the field of machine learning," Would your claim that ML is a "a very small class of techniques" still hold true here?

If you want to predict behavior - loan defaults, the weather, fraud, future or combined purchases - what could you use when not an ML model?

BTW, we don't have a word for an animal without wings, but we could make a list of them - moas, almost all mammals, fishes, and so on.Llaanngg (talk) 21:14, 22 August 2016 (UTC)[reply]

Minor point: In the Maori language, the plural of Moa is Moa. One waka, twenty waka. There is nothing to indicate the plural. New Zealanders have retained the distinction to this day, although some, especially the young, are not aware of it and use the plural with the English 's'. No criticism of your post intended, just pointing this out as a curiosity. Akld guy (talk) 21:51, 22 August 2016 (UTC)[reply]

Sorry about that. This -- " predictive modelling is synonymous with, or largely overlapping with, the field of machine learning," is completely and utterly false. There are zillions of models that are designed to predict something that have nothing to do with machine learning. For one category of examples, see e.g. General_circulation_model. Those are usually based on mechanistic principles of physics, lots of partial differential equations, then some clever bits to semi-parallelize the problem.

Honestly, I think that predictive modelling is a terrible article. Or maybe the article is fine, but it has a the wrong name. It only talks about a relatively limited category of statistical methods. There are many, many things that are models and predictive that are not even mentioned in the article. I don't think the first sentence is properly supported by its reference. Notice the prominent neutrality warning. I think it was written by someone who doesn't really know much about modeling, or somehow got swept away into some niche terminilogy. I stand by my claim that machine learning is a relatively narrow class of methods, even when restricted to models that are designed to predict something about the world, and are designed with the intention of using machines to perform computations. Our article on machine learning is decent, and mostly does not give the impression that it is the only game in town.

Consider that modeling a good pendulum with a set of equations describing simple harmonic motion is a very good predictive model, and it can be refined with friction, air pressure, mass of string, etc., etc. Not only does that have nothing to do with machine learning, it also has nothing much to do with statistics. So I really don't know what to do about that article, that's a mess that I see no easy way out of. The fact that a few authors may have a very narrow sense of "predictive modeling" doesn't mean our article should follow a similarly narrow path. For some more sane description of different model classifications, see Mathematical_model#Classifications.

There are lots of ways to model e.g. loan defaults that are not ML. That have nothing to do with ML. E.g. you could set up a markov chain for each loan, then look at long term behavior in aggregate as a function of various model parameters.

The main problem is that you apparently got rightfully confused by an article that has decent content but horrible terminology that is not consistent with the real world, or even the basic tenets of how words work.

Back to your original question: models are often split into categories of phenomenological and mechanistic, and that is also a split that puts ML on one side and generally not the other. A statistical ML model of weather would look at e.g. correlations of when it rained today vs. rained yesterday. You could try to predict the weather that way, but you're probably better off thinking about physics. Some climate models use data assimilation, but they are mostly based on physics, not phenomenological associations. SemanticMantis (talk) 22:00, 22 August 2016 (UTC) (P.S. bats are mammals with wings, and perhaps flying fish are fishes with wings :)[reply]

For a brief scattering of models and techniques that can be used for prediction and don't use ML: linear regression, general linear model, Bayesian inference, solid modeling, Monte Carlo markov chains, Bayesian networks, Ecosystem_model, soft body dynamics, Computational_fluid_dynamics, population model Hydrological_transport_model, Models of DNA evolution... and yeah, it's almost like trying to list animals without wings. I can do it, but not exhaustively, and I'm not sure when to stop, or what specific part you might be interested in. But hopefully that helps a little bit, and I can try to help more tomorrow if you can further refine your interest. SemanticMantis (talk) 22:12, 22 August 2016 (UTC)[reply]

That surely helps in this form. There are enough tracks to follow.

My interest is mainly broadening my view, finding alternatives to ML, which appeared to me to be over-hyped around me. Llaanngg (talk) 22:27, 22 August 2016 (UTC)[reply]

User:Llaanngg, I see, thanks. Yes, certain crowds would love for you to think that ML is the best/main tool for these purposes. One parting comment: ML can do all sorts of useful things, but ML models seldom have any explanatory power. That is, they can help us determine what might happen, but can't help us figure out why, or how, or in what manner the outcome came to pass. For that, you need another set of modeling tools ;) SemanticMantis (talk) 14:08, 23 August 2016 (UTC)[reply]