jump to navigation

Recommendation Engines Model Outcomes, Not Processes June 1, 2010

Posted by Peter Varhol in Strategy.

My original path to computers and computer science was circuitous.  As an undergraduate student in psychology, I became interested in how chemicals in our brains influenced thought, memory, and behavior, and picked up strong minors in biology and chemistry.

However, the school where I did graduate work didn’t have research facilities to continue such learning, so instead I worked with a proxy for what was happening inside the brain – mathematics.  My masters thesis in psychology modeled the behaviors of test subjects playing consecutive rounds of a variation of the Prisoner’s Dilemma game.

My interest was piqued now not by behavior, but rather the tool I was using to model it.  I continued on for a degree in mathematics, and from there into artificial intelligence.

Artificial intelligence was the stumbling block in my goal of understanding human thought and behavior.  While there were a number of techniques used, the whole goal was to simulate chains of thought used by humans in reaching toward a goal.  And the fact was that none of these techniques worked very well in that regard.

There is a purpose to this tale.  Reading this article from Time.com about recommendation engines made me realize that the requirement that we simulate a human chain of thought was the wrong approach.

You understand the concept of recommendation engines; they are automated tools that recommend items for purchase or rent, based on other items you have chosen.  They work through the power of correlation.  When we purchase an item or rent a movie, we are asked to rate that product.  If we have a record of who has rented what, and how they rated them, software can draw correlations between what others rent and what they might be interested in.

Note that this doesn’t attempt to replicate thought or logic, or even build any sort of underlying rationale.  Recommendation engines can use a number of different statistical techniques that relate multiple independent variables to one or more dependent variables.  It’s all correlation; while there may be underlying causation somewhere in the results, recommendation engines, and the math behind them, don’t really care.

I tend to take recommendation engines with more than a touch of skepticism, but I will occasionally pick an item from them.  The fact that I do occasionally find something means that they are successful.  Such an engine isn’t looking for perfection, which is another point in favor of this type of technology.  Whereas the point of artificial intelligence was to model human thought processes, there are acceptable results under some circumstances by just modeling the outcomes.

The article points out that we give different ratings depending on the circumstances under which we rate (time of day, mood, time passed, and so on).  That’s okay, because we’re aggregating hundreds of thousands or millions of ratings, which would tend to even out any outlying ratings.

More systems like this would be useful.  They aren’t as difficult to develop (although you have to understand statistical techniques and design), and they are useful in cases where it’s not necessary to model the underlying logic to get reasonable information.



1. Tweets that mention Recommendation Engines Model Outcomes, Not Processes « Cutting Edge Computing -- Topsy.com - June 3, 2010

[…] This post was mentioned on Twitter by Peter Varhol, Robert P Reibold. Robert P Reibold said: Recommendation Engines Model Outcomes, Not Processes « Cutting …: Artificial intelligence was the stumbling bloc… http://bit.ly/cLaWPh […]

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: