Duke Listens!: Visit my main blog at MusicMachinery.com

Social recommender systems are currently the best way to get
recommendations.  They give pretty good recommendations, and are
pretty easy to implement.  However, they are not perfect.  They can
be susceptible to popularity bias, shilling, strange feedback loops and
all sorts of issues that can affect the quality of the
recommendations.

One such issue is known as the Donaldson-Bidet Effect.  The DBE is
often seen when a new recommender system is brought online.  It is a
feedback loop that results in an item that is no better than any other
item  becoming heavily favored.

Here's what happens:  The recommender has a number of new items in the
inventory. Since the items are new they are of equal popularity. When
the system presents the items in order of their popularity, no one
item is consistently on the top (since they all have the same
priority).

Now, if one of the items happens to appeal to some of the first users of
the system, that item will become slightly more popular than the other
items (since all the items are new, it only takes a few users showing
interest in an item to make it become the most popular).  This slightly
more interesting item now starts to appear at the top of the 'Most
Popular Items' list.  New visitors that come to the recommender will
now consistently see the slightly-favored item at the top of the list.
It is natural for users to show interest (i.e. click) on the most popular
item. This preference for the most popular item becomes the feedback loop.
Users will click on the the most popular item, increasing its popularity,
until it is entrenched at the top spot.

With the Donaldson-Bidet effect, what we see is that the early
visitors to a recommender have a strong influence over what becomes
popular.  This is evident in some of the current music recommenders.
Sites like Last.FM have a strong bias towards artists like 'The Postal
Service', reflecting their nerd/geek origins, whereas sites like
Rhapsody revolve around Jack Johnson, while MyStrands recommenders
tend to favor Coldplay and Green Day.  These preferences are
continually reinforced by users of these recommenders.


So how did the Donaldson-Bidet effect get its name? Well, it all
started at Recommenders06, the Recommendation Summer School held in
Bilbao Spain during September of 2006.  A number of attendees planned
on taking photos of the conference and posting them on Flickr. One
of the first to post images was a PhD student named Justin Donaldson.
Among the pictures that Justin posted was this picture:



along the caption 'Exciting new latrines in my hotel ... I'm kind of
frightened'.  This photo caught the eye of other Recommender06
attendees since it was slightly unusual, and clicked on it to see why
Justin was posting photos of his toilet and bidet.  Very quickly, this
photo rose to the top of the Flickr "Most Interesting" list for
Recommender06 photos.  Anyone visiting Flickr would find this photo as
one of the most interesting photos from Bilbao Spain.  Thus the
Donaldson-Bidet effect was born.


Update: To see the Donaldson-Bidet effect in action view the recommenders06 images sorted by 'interestingness'