Talk:Local regression

Technicality question
Since the mathematics behind this method of statistical analysis may not be understandable to anyone who has not done an undergraduate maths or statistics degree, is it perhaps appropriate to "dumb down" the introduction paragraphs slightly? 2A01:388:205:311:0:0:1:2 (talk) 16:58, 19 January 2018 (UTC)

"Computationally Expensive" is unscientific
it is frequently mentioned that loess is computationally expensive. this is a bit unscientific. it would be nice to have either asymptotic complexities (big o), or specific examples of problem sizes/computations. — Preceding unsigned comment added by 192.124.26.250 (talk) 14:24, 5 September 2007

Copyright warning
Much of the introduction is taken word-for-word from the [NIST page]. I'm not sure if this is a copyright problem or not, since the NIST site seems to be a Federal government publication (there is no copyright notice). —Preceding unsigned comment added by 136.142.169.66 (talk) 17:26, 14 July 2008 (UTC)
 * Yes, that is the whole point of the NIST-PD at the bottom of the page, which produces
 * [[Image:NIST_logo.svg|30px]] This article incorporates text from a public domain publication of the National Institute of Standards and Technology, a U.S. government agency.
 * Btyner (talk) 23:14, 14 July 2008 (UTC)
 * That note is rather useless, since although it says that the material comes from the NIST website, but does not state what web site. Geoffrey.landis (talk) 13:40, 14 April 2017 (UTC)

LOESS or loess?
Cleveland named his method "loess", not "LOESS"; the name is not an acronym, but a reference to the geologic material (cf. the cited articles). I see that the NIST page writes it in all caps, but I believe this is erroneous. This wikipedia article is currently a bit inconsistent, writing it "LOESS" some places and "loess" others. Is there any reason not to edit all instances to lower-case? Thomas Tvileren (talk) 15:24, 22 February 2011 (UTC)

Having heard Bill talk about LOWESS before he started calling it LOESS and again after, the story was to me not quite so clear. LOWESS is the original acronym, and when he did a presentation to geologists, they suggested that he drop the "W" from the acronym because loess did often have a visual similarity to nonparametric regression curves and was still suggestive of the original word. The exact details of whether it should still be thought of as an acronym were a little vague at that time - I don't recall Bill saying "it's not an acronym any more" (this would have been before the 1988 paper, his views may have become more set later). I'd say it's at least arguably still a kind of acronym and you do see "LOcally wEighted Scatterplot Smoother" written out. You're right that Cleveland nearly always writes it as "loess" but fairly common usage has "LOESS" (while others write "Loess"). It's not completely clear that we should say 'LOESS is wrong', whatever Bill's view of it. We should have consistency in the article, and on that basis it's perhaps easy enough make an argument for following Cleveland's practice on this, but "erroneous" is perhaps too strong. Glenbarnett (talk) 22:48, 28 January 2018 (UTC)

From the Cleveland and Devlin article (1988), "The shortened name loess has some semantic substance. A loess (pronounced "1ō' is") is a deposit of fine clay or silt along river valleys; in a vertical cross-section of earth, a loess would appear as a narrow, curve-like stratum running through the section." I would vote for both "no w" and "lower case" as the rules to be followed. Ehusman (talk) 04:09, 11 February 2018 (UTC)

"LOESS"/"LOWESS" is correct, "loess"/"Loess" is wrong. Acronyms should be written in all capitals. Languages live by their own rule, no matter what the method's creator personal preference is. Lower-case spelling is a marketing gimmick used by researchers to attract attention and promote their work (akin fancy abbreviations like RoBERTa, ERNIE, WEASEL). It is also possible that the original author made an honest mistake, after all they are an authority in statistics, not in linguistics. We should write LOESS, unless we mean the geological structure. Similarly, we should write LASSO, unless we mean a type of a rope. AVM2019 (talk) 13:09, 31 May 2022 (UTC)


 * Acronyms should be written in all capitals.
 * Not according to:
 * Paul Kennedy:
 * University of Arizona:
 * Cambridge Dictionary:
 * In fact, you seem to be confusing acronyms with initialisms, which, strictly speaking, loess is not (not only it constitutes a pronounceable abbreviation, but it is composed of fragments of its composite words, not their initials: cally ighted catterplot moother).
 * — Guarapiranga ☎ 14:05, 31 May 2022 (UTC)
 * Thanks for the references, it is good to have them at hand and also to learn something new. You are correct that I ignored the difference between acronyms and initialisms, and that LoESS is of the former type. Nevertheless, I protest against writing all-small letters (loess) or capitalising it like most proper nouns (Loess). Let me further quote the Cambridge Dictionary reference you used
 * One may argue that LOESS is established well enough, or that there are other situations in which acronyms can be written in small letters, but I do not buy that. Common sense suggests that there are practical advantages of writing LOESS or LoESS (as compared to loess or Loess), which should be prioritised over the inventor's preference (and even over formal rules, had any of them contradicted this). First, it is easier to scan a technical text for model names when they have many capitals. Second, this serves as an unambiguous indication (in the beginning of a sentence or elsewhere) that a reader is dealing with a name of some method rather than a common noun. This is especially helpful to readers who are not very proficient in English or statistics (sounds like a good chunk of Wikipedia's target audience).AVM2019 (talk) 19:54, 31 May 2022 (UTC)
 * One may argue that LOESS is established well enough, or that there are other situations in which acronyms can be written in small letters, but I do not buy that. Common sense suggests that there are practical advantages of writing LOESS or LoESS (as compared to loess or Loess), which should be prioritised over the inventor's preference (and even over formal rules, had any of them contradicted this). First, it is easier to scan a technical text for model names when they have many capitals. Second, this serves as an unambiguous indication (in the beginning of a sentence or elsewhere) that a reader is dealing with a name of some method rather than a common noun. This is especially helpful to readers who are not very proficient in English or statistics (sounds like a good chunk of Wikipedia's target audience).AVM2019 (talk) 19:54, 31 May 2022 (UTC)

Variables Not Defined
"n" is not defined (probably means number of observations), and "x" is not defined. It's not acceptable to write equations and not define the terms--very poor practice. Would the author please include them? Thanks! Chafe66 (talk) 19:47, 4 February 2016 (UTC)
 * Back I come, one year later, and the variables are still not defined. Perhaps no one is watching the talk page. If anyone is feeling like they own this page, please step up. Otherwise...section should be rewritten to normal standards (defining variables that appear in equations). Chafe66 (talk) 01:23, 9 February 2017 (UTC)
 * There are a lot of places where variables are not defined. This article needs work. The NIST page referenced does, however, define n, so I'll add that to the page. Geoffrey.landis (talk) 13:42, 14 April 2017 (UTC)

Span in LOESS
On the local regression page, under "Localized subsets of data" it is stated that

"Useful values of the smoothing parameter typically lie in the range 0.25 to 0.5 for most LOESS applications."

while on the Bootstrap aggregating page, under the Example section, it is stated

"The lines are clearly very wiggly and they overfit the data - a result of the span being too low."

Are they contradictory statements?

Astrobob.tk (talk) 21:14, 14 May 2016 (UTC)

External links modified
Hello fellow Wikipedians,

I have just modified one external link on Local regression. Please take a moment to review my edit. If you have any questions, or need the bot to ignore the links, or the page altogether, please visit this simple FaQ for additional information. I made the following changes:
 * Added tag to http://www.stat.purdue.edu/~wsc/papers/localregression.principles.ps
 * Added archive https://web.archive.org/web/20050912090738/http://www.stat.purdue.edu/~wsc/localfitsoft.html to http://www.stat.purdue.edu/~wsc/localfitsoft.html

When you have finished reviewing my changes, you may follow the instructions on the template below to fix any issues with the URLs.

Cheers.— InternetArchiveBot  (Report bug) 22:59, 4 January 2018 (UTC)

LOWESS/loess is not all of local regression
The article reads as if Cleveland's loess is all there is in local regression. I think there's plenty of presentations of local regression without the "robustness" parts that loess has in there. I think it would make for a cleaner, more understandable presentation to have a more generic article on local regression (referencing any number of sources) and within that a substantial section on LOWESS/loess (which is important but not the whole area). Glenbarnett (talk) 23:03, 28 January 2018 (UTC)

Locally Weighted Learning listed at Redirects for discussion
An editor has asked for a discussion to address the redirect Locally Weighted Learning. Please participate in the redirect discussion if you wish to do so. Widefox ; talk 15:05, 28 July 2019 (UTC)

Weights
I'm not a maths graduate, but I was exploring non-parametric locally-weighted trend-lines, and I was just wondering if anyone uses weights based on the formula for the bottom-left corner of a circle of radius one centered at (1,1), ie Weight (w) = 1 - (2|d| - d^2)^0.5, where d is the distance along the horizontal axis of the data-point from the point you want to estimate, scaled so the maximum distance is 1? Below I've included an example of its application on Australian youth unemployment data with the trend-line being based on a radius of nine data-points. It does not use regression techniques, nevertheless, it sure fits the data a lot better than the simple asymmetric (lagged) and unweighted moving averages built into Excel. I've also included a second version with the same data, basing the trend-line weights on Weight (w) = (1 - (2|d| - d^2)^0.5)^0.5, extending the smoothing range to 19 data-points either side, and adding interpolated points based on the average of the adjacent trendline points plus or minus an eighth of the difference between the average of the immediately adjacent trendline points and the average of the next closest trendline points (increasing it if the far points are lower than the near points, and vice versa). It produces a result very similar to Lowess smoothing. I was hoping to learn what I'd have to do to get from the weights to a Lowess regression, but this article doesn't really specify the method in a way that can be easily followed. The article mentions Cleveland (1979), but provides no links to the text. Would a circular-function-based weighting be consistent with Cleveland (1979)? I notice that it would have more of a spiked-central weighting than any of the weighting functions mentioned in Kernel (statistics) Alternatively, it would be useful if the article included a brief summary of the recommended conditions for a weighting system. MathewMunro (talk) 15:51, 12 March 2021 (UTC)

LOWESS
"LOWESS (locally weighted scatterplot smoothing)" --In the abbreviation "LOWESS", where does the capital E come from? It can't come from "estimated", because that isn't part of the expansion. Why is it not known as LOWSS, rhyming with "dose"? — Preceding unsigned comment added by 82.209.132.154 (talk) 08:46, 29 January 2023 (UTC)


 * I'm pretty sure there's an E in it for the same reason there's an O in it. It's not a true acronym. It's " LO cally WE ighted S catterplot S moothing." Columbo2014 (talk) 20:03, 15 June 2023 (UTC)