Showing posts with label statistics. Show all posts
Showing posts with label statistics. Show all posts

Sunday, February 19, 2012

Emergent Similarity

Computers and humans seem to have a different notion of similarity. Here are a couple of examples that appeared almost simultaneously on Twitter.

@socialtechno "Twitter told me @dansabbagh is 'similar to' @rupertmurdoch. Computers do not make skilful editors."

@markhillary "Got my Spotify tuned to music similar to the Pogues and Meat Loaf just came on..?"


When statistical analysis throws up a surprising similarity or juxtaposition, we humans naturally wish to explain this by finding something in common. It is quite possible that there is a common characteristic that nobody has previously noticed, and statistical analysis then leads us to a new way of classifying things: scientific progress has sometimes taken this route. But statistical analysis is also perfectly capable of throwing up meaningless similarities - or what we humans with our limited intelligence are unable to find meaningful.

However, computers usually infer similarity between two items not by their intrinsic characteristics (affinity clustering) but by their relationship to external activities (interaction clustering). People who liked A also liked B. People who searched for A also searched for B. We just need to get used to the fact that this notion of "similarity" is not the same as ours.

 

 

See also Rhyme or Reason - the Logic of Netflix (June 2017)

Friday, November 21, 2008

Missouri loses bellwether status

As was widely predicted, Missouri has lost its bellwether status [BBC News, 21 November 2008]. For only the second time since 1904, Missouri voted for the losing candidate in the presidential election. Aw, shucks! Immediately after the election, David McCoy pointed out that the concept of bellwetherhood was Bad Statistics. It is also bad inductive reasoning to suggest that something is unlikely because it hasn't happened before, or that something is remarkable because it is unlikely. J.R. Lucas (author of several books of philosophy including The Concept of Probability) used to call this the Dover Fallacy.
There was a young curate of Dover Who bowled twenty-five wides in an over, Which had never been done By a clergyman’s son On a Tuesday in August in Dover.
The BBC reports Larry Sabato, Professor of Political Science at the University of Virginia, as saying that trying to identify bellwethers may not be a worthwhile enterprise. But if you search for "Larry Sabato bellwether" you can find what appear to be many attempts by Professor Sabato to do just that for various state and national elections including the most recent. Glad to see a professor willing to change his mind about something. Meanwhile, there is apparently some status involved in being a bellwether state. Now I don't know a lot about herding sheep, but I think the way it works is that if you are herding sheep you only need one bell, but if you are herding goats you need one bell for each goat. Perhaps having a bell around its neck made Missouri feel more important than the other sheep. Now people are already discussing how to get Missouri its bell back.

Listen Missouri, you don't need this. Did you know that the bellwether was originally a castrated ram? Or you could always try astrology instead: Reclaiming our Revolutionary Roots.

 

See also You Don't Need to be Smart Here (October 2008)

Monday, October 09, 2006

Regression Fallacy

If you criticize people after an exceptionally bad performance, they tend to get better next time. If you praise people after an outstandingly good performance, they tend to get worse next time. Therefore criticism is more effective than praise as a way of improving performance.

WRONG

After an exceptional performance - either exceptionally good or exceptionally bad - the next performance is likely to be closer to the average. This is a well-known phenomenon in statistics known as regression toward the mean.

So the short-term effect of praise and criticism may be a lot less significant than it appears. And if we want to look at the longer-term effect of praise and criticism, we need to take a much longer view than the next performance. 

The purpose of a system is what it does. But which system, over which time horizon? If we are overly concerned with short-term results, we may easily come to believe that criticism is more effective (fit-for-purpose) than praise. The same may be true of all sorts of other meddlesome interventions. People learn to meddle and criticize because these actions appear to have short-term effects, and so these actions are reinforced. 

System thinking invites us to pay attention to the scoping and time horizon of the system in question. It may be difficult to look beyond the narrow short-term effects, but it may be important to try.