Google is testing the functionality of allowing users to re-sort search engine rankings by simply moving a result up or down the page. If we assume this data to be used in future approaches to ranking pages based on their relevancy to a keyword, we find ourselves treading into the territory of active learning. This is where a machine learning algorithm offers a human a guessed output or an instance it cannot classify or rank and then internalizes the human response. The algorithm then holds on to the human response, applying these alterations to its future calculations.

How might this change Google’s ranking algorithm? By incorporating human responses into Google’s automated opinions on a set of ranking results for a keyword, we make the algorithm itself more multivariate. Ask a person to rank ten similar web pages for a keyword based on its relevance to the page, and then ask them to define what they considered relevant for the task, and you’ll come up with two sets of problems that no two individuals will likely come eye to eye on.

So depending on how important Google considers this supervised information (and, let’s face it, human supervision is a great data source and something they’d be extremely willing to exploit), SEO will be significantly more subsumed by the area of search engine marketing. Sites will need to have wide appeal and reach different personalities in order to stay on top, and not just decent anchor text and roughly relevant content peppered with consumer-targeted keywords.

Conversely, those donning the “black hat” may be more likely to recruit a league of minions to change rankings for their own ends—individuals risking getting their own IPs banned to make some extra money in a brand new “work from home” scheme. There are a number of ways that mixing things up with the incorporation of active learning can take place. For instance, while we know this new functionality is a type of active learning, we don’t know if messing with the rankings for one keyword will have any impact on the rankings for another; it depends on how much independence Google lends the features it uses for making its ranking decisions and whether these human-elicited changes become learned weights on features applied to all keywords as a whole or simple transformations included in the calculations for each keyword.

This entry was posted on Wednesday, August 27th, 2008 at 2:07 pm by relwell and is filed under Google, Natural Search (SEO). You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.