agreeablesocietyΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 4 χρόνια και 8 μήνες)

260 εμφανίσεις

Recommendation websites attempt to
recommend media(music, art, video),
websites, and other types of information
based on a users taste.

There are many methods with which the
recommender website tries to do this
known as recommender systems. There
also many different websites that use
these recommender systems for different

Among these websites is StumbleUpon and
Pandora Radio.

As said before there are a number of other
websites making use of recommender
systems but I choose these in particular so
we can get a good cross section of the
typical use of recommender systems for
specialized recommender websites.

I choose StumbleUpon because it is the
most popular one and Pandora because of
the creation of the “Music Genome
Project”, it covers the whole idea of
recommender systems that attempts to
analyze a users taste in something as
subjective as art.

Recommender Systems are different from
recommendation websites in that these
systems may be present on websites whose
sole purpose is not just to recommend. For
example the recommender system on informs the user of items
purchased by other users who also bought
said items. The “did you mean” feature of
Google can help you articulate your search

System collects information about the user in two ways:

Implicit data collection:

is a noninvasive method
where the computer analyzes the items bought
and the viewing times of specific items, Recording
what the user purchased online. Lastly analyzing
the user’s social network to see what similarities
they may have in terms of likes and dislikes.

Explicit data collection:

This is a form of collection
that requests the user to input information about
themselves. Requesting users to rank or rate items,
asking the user to create a list of items they like or
asking them to choose between two items.

The information collected about the user is compared it
information stored about other users. The system then suggests
items to the user that other users with similar tastes enjoyed.

Recommender System are a such a major part of the
internet today that there very few successful websites that
don’t have some sort of recommender system at use.

The search for the best recommender system is still on going
as revisions of the algorithms behind these systems still
appear today. The Netflix Prize awarded a million dollars to
the grand prize winner of their competition to find the best
collaborative filtering algorithm.

Recommender systems may seem controversial to some
especially when joined with social networking. There are
some who are uncomfortable with how much access the
system has to their personal information and how it shares
this information. However this is not a major problem
because the systems are discreet and don’t give away
important information about users.

“We created StumbleUpon so people can
“stumble upon” sites that have been
submitted and rated by like
people, rather than presented with the
most popular sites for a given keyword.”

Garret Camp in an interview with
Chris Sherman of

The website was founded in 2001 by:

Garret Camp Geoff Smith Justin

And Eric Boyd.

*the website was founded while
Garret Camp was doing his postgraduate degree at
the University of Calgary in Alberta Canada.


fast popularity attracted
the likes of Brad O’Neil, Ram Shriram, a
billionaire who was one of the founding
board members of Google, and Mitch

of Mozilla Firefox. Collectively they
fundraised $1.2 million US dollars.

In 2007 eBay bought StumbleUpon for
$75,000,000 US dollars

In 2009 Camp, and Smith among many
other investors where able to buy the
company back for what is rumored to
be less than what the people at EBay
bought it for.


StumbleUpon uses a recommender system
known as “collaborative filtering”.

StumbleUpon also has a social networking
aspect as the users, known as stumblers, are
able to comment on the websites
recommended to them. They also critic
each others blogs. People who seem to
have similar preferences based on the
information collected by StumbleUpon form
social networks.

The people at stumble Upon make profit in two ways
through sponsors in which stumblers pay a certain
“donation”( a suggestion of 20 dollars a year). This
donation gives them privileges other stumblers don’t
have. Some sites are sponsored, but around only 2% of
StumbleUpon sites fall under this category.

Ways to Stumble:

Classic stumble.
User ticks of their interests and
they are shown a website they click stumble to
move on and as they go on they can give
websites a thumbs up or down and this helps the
system narrow down their tastes to better the

Stumble Thru.
The user is able to stumble through
particular websites instead of through different
web pages from different sites.

Stumble Video and
Stumble video is a
website where users without the toolbar stumble
through videos submitted by people from the
StumbleUpon website who have
helps user shorten URLs to post on their website
Facebook statues.

StumbleUpon is a very addictive website.
Not only does it entail surfing the internet
where the content is specifically geared
towards your personal taste it also has a
social networking aspect.

Although it hasn’t become a major
concern to many. Stumbleupon among
other social networking websites is a
source of procrastination for many,
which may seem trivial but is dangerous.

“I’ve always believed the idea was a good

the basic idea of
profiling a
user’s tastes
, capturing it in the
Genome, and creating a technology that
helps people discover music

Tim Westergren founder Pandora
Media inc and Chief Strategy Officer of Pandora Inc. in
an interview with Sanpshot Music and Art Foundation.

Will Glaser Tim Westergren
Jon Kraft
Nolan Gasser


began playing the piano at the age of two and
composed at 8!He also studied for two years in Paris.

Will Glaser and Tim Westergren are the brains behind
Pandora Radio(1999) who joined with Jon Kraft to create
Pandora Media Inc. (2000).Currently ,the Music Genome
Project, a very complicated form of recommender system,
is owned by Pandora Media Inc.

In 2007 Sound Exchange's request that internet radio pay
double the price in royalties per song than satellite radio
was met. Continuous disagreements abut royalties led to all
American Internet Radio to be restricted to play in only
America. This also lead to users being restricted to 40 hours
of free listening per month and having to pay .99 cents
should they exceed their listening hours.

In 2008 Pandora Radio launched a mobile version of the
service through the iTunes App store and the app is usable
for the iPod iPhone and iPad. The app is also available for
Android phones and The BlackBerry among others.

How it works:

Pandora Radio was developed using the Music
Genome Project. To the user it may seem as
simple as the they input a song they enjoy and
get back similar song, but behind this is a very
complicated recommender system.

The Music Genome Project is as personalized as
“automated music recommendation” can be.
Specialized musicians working at the company
analyze each song categorizing them using up
to or more than 400 attributes! This Analysis take
up to 20 minutes depending on the song.
Pandora claims they have around 800,000 and
add thousands every month. The attributes
range from tonality and ostinato to categories
named “knack for Cathy tunes” or “Intelligent
Dance Music”

Although Pandora approached their
recommendation service in a way that makes it
seems more personalized it also the service has it’s
faults. The service relies on the opinion of well
qualified musicians. It is impossible for them not to be
subjective, however minor the extent of this
subjectveness. How educated are they on what real
gangster rap is?

In addition, there is also the question, am I really
discovering new music if I listen to music that is only
my taste? Some may argue that you can’t discover
what you already love and discovery should be left
to experience.

Can something as artistic and subjective as music be
put through a scientific procedure and categorized?

All in all we can say this about
recommender systems, they are a very
apparent part of the online world today.
For the question of how personalized they
can get we can conclude from the
numbers they draw they do work and are
as close as one can get to machine
learning in terms of machines learning
about taste. Continuous revision of current
algorithms used in these systems is sure to
lead to something beyond our imagination.