So now I am confused. Google seeks to do personalization – which is a reduction in diversity (and a reduction in entropy). And they also seek to increase diversity in their search results, which is an increase in entropy.

How can they be doing both at the same time?!

I think you're monitoring [google blog] to see how Google Blogoscoped is faring, right? :-)

Cluster search is good

Mux: I don't think that personalization always results in reduction in diversity (on the first page, anyway).

For example, an apple grower would find some fruit results in their personalized search for "apple", increasing the diversity compared to the computer-only results that would otherwise be returned for a search on "apple".

> I think you're monitoring [google blog] to see how
> Google Blogoscoped is faring, right? :-)

I use it to test if there's big shuffles in this blog's ranking, but also to test in general if there's bigger shuffles in how Google ranks stuff. It's also a good sample query to compare different search engines because "google blog" is a rather open query and there's nothing too simple telling search engines that e.g. Google Blogoscoped or Google Operating System or Googlified should rank well for this. Try [google blog] in Yahoo and Microsoft's Live.com – IMO, Google has the best overall result (looking at all 10 items), Yahoo comes in second, and Microsoft is third.

Both Yahoo's and Microsoft's results have some "hiccups" like showing the default Google entry on Wikipedia, which is not an appropriate result for the query [google blog] but does have "...wiki/Google_Blog" in the URL, probably for historic reasons as someone set up that entry and it then got merged with the main Google entry on Wikipedia. Yahoo also only shows 1 independent blog (disregarding Matt Cutts' blog, which is sort of in-between), and Yahoo shows google.com, which is not really relevant for [google *blog*].

Live.com is worst with some distance. They find blog.google.com which redirects to google.com and is irrelevant. They find an entry about Google blog search pinging help, which is only remotely relevant in the top 10. They find a Google Group which is irrelevant. They find an "add gadgets" page which shows some entity encoding problem in the URL (&), and is also irrelevant.

Google's result is the only where each of the top 10 entry is at least somewhat relevant, even though their top 10 could be better (e.g. by replacing google.blogspace.com with googlesystem.blogspot.com – I wonder what effect the blogspot URL has here, as it's shared by more other sites than the blospace.com URL; the google.blogspace.com URL is relevant historically, so it may make sense to have it appear say from results 9 – 20 or so). Of course, this query is by no far representative, and there may be many results where Google fares worse than Yahoo (and perhaps even Live.com).

>For example, an apple grower would find some fruit results in their >personalized search for "apple", increasing the diversity compared to >the computer-only results that would otherwise be returned for a >search on "apple".

I still don't quite follow you, Roger. A results list that contains results for fruit apples, apple computers, apple records, and johnny appleseed is, more diverse than a results list that only contains fruit apples.

That's the definition of diversity.

So if personalization is only showing fruit apples, then it is reducing that diversity. Not increasing it.

Mux, Roger in his example said that only some (not all) results would be replaced with apple-the-fruit in the more personalized results. In other words, in Roger's example the result was more diverse with personalization as it now contained two types (fruits and technology) as opposed to just one type (technology).

Ah, ok, so personalization would make the current results for "apple" more diverse than they are *now*, in Roger's example.

But I am reacting to your comment, Philipp, that: "Google results which are diverse – approaching your query from different angles – are good, and those which aren’t diverse can be a barrier to find what you’re after."

So according to that, Google results that showed results for fruit and trees, records (the Beatles' label), technology, and recipes (application of the fruit to cooking), etc. (more diversity) would be better than results that showed less diverse results (i.e. just fruit and technology) and even better than results showing even less diversity (technology only).

We have three levels of diversity here:
(A) Fruit/trees, recipes, computers, records
(B) Fruit/trees and computers
(C) Computers only

So Roger is basically saying that personalization moves you from (C) to (B), thus increasing diversity. My argument is that if diversity is so good to begin with, then Google should be showing (A). Therefore personalization would move you from (A) to (B), *reducing* diversity.

Yeah... if personalization would kick out all other result types, then it would reduce diversity, though it may also often be possible that there are enough of each kind in the 10 results so that no result type has to be exchanged completely. In general, I agree with you and also think that personalization has the big risk of reducing diversity. I do not want to be stuck in my old opinions, I am searching to find new stuff, and the apple I was searching for yesterday may well be the fruit, even when I want to look for Mac news today. But perhaps Google is equally aware of these problems and careful not to overdo their personalization? Guess time will tell...

>I do not want to be stuck in my old opinions, I am searching to find new stuff, and the apple I was searching for yesterday may well be the fruit, even when I want to look for Mac news today.

On that I am also in complete agreement with you :-)

It would actually be extremely interesting if Google could "categorize" the different meanings, by using for example some kind of colors.

Philipp, I wanted to write about this before but I didn't have the time. Sorry for the delay. I actually wrote a post commenting about how well Clusty (a search engine) does this job.

It essentially clusters the results giving the user the opportunity to narrow the results by class, a part from providing a top 10 search result just like any other search engine.

More details here: http://www.laedevolta.com.br/smallbits/2008/04/clusty-search-engine-clustering-results.html

Well, great post, as always.

Great post! Please check out Yahoo Glue (launched for Yahoo India). It goes in the direction of diversity in an innovative way:

Search for the Indian city 'Mumbai'
http://in.search.yahoo.com/search?p=mumbai&y=Search&fr=sfp&ei=UTF-8&glue=1&rd=r1

Great blog, keep up the awesome work.