Latent Dirichlet Allocation Surprisingly Well Correlated w/ Google Rankings
seomoz.org
seomoz.org
To a technologically savvy but naive outsider, this might seem obvious: shouldn't pages that rank highly in Google have strong topic-based correlation to pages that the user wants to see? But from the SEO perspective, I guess the conclusion would be that your page is more likely to be ranked highly if it includes all the trappings of other high ranked pages, with, you know, like synonyms and stuff. At a certain point, one has to start thinking, wouldn't it be simpler to make a page that people actually want to find?
Are there good examples of actually useful pages that Google doesn't do a good job of ranking? I occasionally find myself lately getting frustrated with Google about ignoring my rarer search terms, but generally I find the good pages are at the top if they exist at all.
[1] LDA is Latent Dirichlet Allocation, which is very similar to Latent Semantic Analysis, which in turn is very similar to Principle Component Analysis and Singular Value Decomposition. So it's possible you've already heard of the concept, but coming from another angle in another field.
But I've never really looked at LDA, and Wikipedia says: Compared to standard latent semantic analysis which stems from linear algebra and downsizes the occurrence tables (usually via a singular value decomposition), probabilistic latent semantic analysis is based on a mixture decomposition derived from a latent class model. This results in a more principled approach which has a solid foundation in statistics. so maybe they made the right choice. (Not that I see what "solid foundation in statistics" really means in this context)
My understanding of LDA is that it gives you document scores against queries based on the topics extracted using the LDA algorithm on the text in the page.
Pagerank, on the other hand scores based on external pointers (ie, references) but doesn't have anything to do with the text on the page.
I thought you were talking about some functional similarities, not the mathematical similarities.
Yes. Reviews by actual people.
A long, long time ago (last year I think) I used to use google blogsearch to get at the "other side" of hardware reviews: what people who are not being paid to review hardware think of a specific piece of hardware.
Doesn't work any more. There's almost nothing but spam. It's hopeless.
They have found a correlation between a set of words related to a topic you are searching for and how highly a search engine ranks that page?
Well duh! Did anyone really think search engines did a keyword search and then applied Pagerank/HITS (http://en.wikipedia.org/wiki/HITS_algorithm) or whatever? That would give dreadful results.
If you really want to understand this, I recommend Building a Vector Space Search Engine in Perl http://perl.about.com/b/2007/05/24/building-a-vector-space-s...
I build the vector space classifier in http://classifier4j.sf.net based almost entirely on that article, even though I don't know Perl. It's very readable, and gives you a great understanding.
They also use a lot of other ranking factors beyond just the words on the page so seeing such a high correlation from a "bag of words" model is pretty interesting (to me at least).
Did you test non-LDA methods? Because to me it looked like a correlation between a set of related words and ranking for a topic related to those words.
Without testing non LDA methods I can't see what you've proved.
If I've read the graph right, it's about 0.33. For a Pearson (product-moment) correlation coefficient, that would mean that about 10% of the variance in Google rankings is explained by a linear regression on LDA scores. They've actually used the Spearman (ranking-based) correlation coefficient, which is equivalent to ranking all the values of each variable from 1..N and then computing the Pearson correlation coefficient for the ranks. So, kinda-sorta with lots of handwaving, that means that about 10% of the ordering of the Google rankings is explained by the LDA scores.
Clearly that's a lot better than for the other scoring methods they mentioned, and that probably indicates that Google are doing something a bit like LDA (but this will be true for any approach that takes note of synonyms, and it's hardly news that Google do that). But it doesn't, e.g., suggest that PageRank and other things based on link structure aren't extremely important to Google's rankings.
It is also fascinating to see this much correlation with an on-page factor which is entirely in the webmaster's control. Previously the highest correlations had been with link metrics.
YMMV but this research is interesting to me as someone who works in this field.