Content uploaded by Marc Torrens
Author content
All content in this area was uploaded by Marc Torrens on Jul 31, 2018
Content may be subject to copyright.
Arecommender system is a software application capable of
suggesting interesting things to its users after learning
their preferences over time (Jannach et al. 2010, Ricci et
al. 2011). Recommender systems were envisioned in the 1970s
(Negroponte 1970), conceptualized and prototyped in the early
1990s (Goldberg et al. 1992), and implemented and rst com-
mercialized in the mid-1990s (Resnick and Varian 1997). They
have two (sometimes diametrically opposed) value propositions
that contribute to their popularity. On one hand, they help
their users cope with the problem of information overload (that
is, they take the users’ side). On the other hand, they provide
companies that operate them with an effective way to drive
more sales or increase the level of engagement of the services
that they offer (that is, they take the business’s side).
Multiple vendors have taken recommender systems to market
in a series of waves over the last two decades. The rst two waves
can respectively be matched with the web 1.0 and 2.0 revolu-
tions. Both waves introduced dozens of companies selling rec-
ommendation technologies (see table 1) and thousands of com-
panies using recommender systems with the nal goal of
increasing their revenues. Not all of the recommender vendors
that started during these waves are still active. Nevertheless, rec-
Articles
FALL 2011 19
Copyright © 2011, Association for the Advancement of Articial Intelligence. All rights reserved. ISSN 0738-4602
The Big Promise of
Recommender Systems
Francisco J. Martin, Justin Donaldson, Adam Ashenfelter,
Marc Torrens, and Rick Hangartner
nRecommender systems have been part of the
Internet for almost two decades. Dozens of ven-
dors have built recommendation technologies
and taken them to market in two waves, rough-
ly aligning with the web 1.0 and 2.0 revolu-
tions. Today recommender systems are found in
a multitude of online services. They have been
developed using a variety of techniques and user
interfaces. They have been nurtured with mil-
lions of users’ explicit and implicit preferences
(most often with their permission). Frequently
they provide relevant recommendations that
increase the revenue or user engagement of the
online services that operate them. However,
when we evaluate the current generation of rec-
ommender systems from the point of view of the
“recommendee,” we nd that most recom-
mender systems serve the goals of the business
instead of their users’ interests. Thus we believe
that the big promise of recommender systems
has yet to be fullled. We foresee a third wave
of recommender systems that act directly on
behalf of their users across a range of domains
instead of acting as a sales assistant. We also
predict that such new recommender systems
will better deal with information overload, take
advantage of contextual clues from mobile
devices, and utilize the vast information and
computation stores available through cloud-
computing services to maximize users’ long-
term goals.
ommender services are now a ubiquitous part of
many online services. Dozens of applications cap-
ture our explicit and implicit preferences on a dai-
ly basis (normally with our permission) with the
goal of recommending something interesting in
the near future (Kirn 2010). Hundreds of
researchers around the world have contributed to
improve the accuracy, scalability, security, and
many other aspects of recommendation systems
(Konstan, Riedl, and Smyth 2007; Pu et al. 2008;
Bergman et al. 2009; Amatriain et al. 2010).
However, when we look at the current recom-
mender systems generation from the point of view
of the “recommendee” (users’ side) we can see that
recommender systems are more inclined toward
achieving short-term sales and business goals.
Instead of helping their users to cope with the
problem of information overload they can actual-
ly contribute to information overload by propos-
ing recommendations that do not meet the users’
current needs or interests. Consider the following
questions: Could you imagine the Netflix recom-
mender suggesting that you watch a TV show that
is broadcasted tonight instead of prompting you to
stream another movie from the Netflix repository?
Could you imagine the Amazon recommender sug-
gesting that you borrow a novel from your friend,
who already bought it a few months ago, instead of
recommending to buy it now? Could you imagine
a bank recommender suggesting that you transfer
your savings to a certied deposit with higher
interest in another bank? Could you imagine the
iTunes recommender suggesting that you stop buy-
ing because you are exceeding your monthly budg-
et for songs?
The recommendations we just mentioned are
hard to imagine because recommender systems
have mostly proliferated on the business’s side.
There is no technical reason why a recommender
system could not make such suggestions by con-
sidering a broader view of users’ goals and sources
of information. In fact recommender systems were
originally envisioned to operate this way (Gold-
berg et al. 1992; Maes 1994; Resnick et al. 1994;
Negroponte 1995) but most vendors favored
licensing their recommendation technology to
retailers over direct-to-consumer initiatives. How-
ever, the current confluence of easily programma-
ble mobile devices, open APIs, distribution chan-
nels (such as the Apple App store), and the low cost
of cloud-based computing and storage creates an
excellent breeding ground for a new wave of
direct-to-consumer recommender systems.
One aim of this article is to encourage both
researchers and practitioners to embark in new
commercial adventures. The window of opportu-
nity is now open to innovate in a third generation
of recommender systems that act directly on
behalf of their users and help them cope with
information overload. Next we review the rst two
waves of recommender systems, share lessons
learned, and describe what we can expect from the
next wave.
The First Wave of
Recommender Systems
In the 1980s the increasing use of electronic mail
and newsgroups caused the rst common infor-
mation overload problems (Palme 1984).
Researchers started to investigate new methods of
handling these issues, drawing on theories and
techniques from general information ltering sys-
tems (Goldberg et al. 1992; Resnick et al. 1994) and
word-of-mouth automation (Shardanand and
Maes 1995). At this early stage, researchers also
investigated more sophisticated systems that
worked proactively on behalf of the user in per-
sonal computing contexts (Kozierok and Maes
1993). Pattie Maes used the metaphor of a personal
assistant to exemplify the collaboration between a
user and an interface agent capable of learning the
user’s interests, habits, and preferences to cope
with the problem of information and work over-
load (Maes 1994). As the agent gradually learns
how better to assist its user, the range of tasks that
can be delegated increases. Common early tasks
included scheduling meetings and ltering email.
Researchers reasoned that harnessing the collec-
tive assessment of many individuals could quickly
establish abstract notions of priority and classica-
tion. This basic idea was the basis for Tapestry, the
rst experimental system supporting collaborative
ltering. Tapestry was built at the Xerox Palo Alto
Research Center to handle the (relatively) large vol-
ume of email received by their researchers at the
time (Goldberg et al. 1992).
In addition to email, collaborative ltering quick-
ly found success in a number of domains. The Grou-
pLens project used collaborative ltering in order to
help people nd interesting articles in online news-
groups (Resnick et al. 1994). In 1995, the GroupLens
team proposed movie recommendations through
the Movielens project (see gure 1). The Ringo sys-
tem at MIT used social information ltering in order
to generate personalized music recommendations
(Shardanand and Maes 1995).
It was not long before collaborative ltering
techniques found their way into commercial prod-
ucts. In 1995 Firefly Networks (originally Agents,
Inc.) became the rst company to focus on offering
a recommendation service. Firey Networks used a
successor of Ringo that was capable of generating
user proles to recommend music. It expanded its
service to recommend movies and websites as well.
As the 1990s ended, more and more commercial
applications included collaborative ltering, and
efforts were made to better monetize the new tech-
Articles
20 AI MAGAZINE
nology. Resnick and Varian (1997) proposed three
business models for independent, direct-to-con-
sumer recommender systems. These included sub-
scriptions or pay-per-use, advertiser support, and
content-owner support. However, recommender
systems were costly to run and maintain. There-
fore, many recommender system startups focused
primarily on licensing the technology to the bur-
geoning e-commerce domain. Net Perceptions,
founded by a group of researchers at University of
Minnesota in 1996, quickly became the leading
vendor of business-to-business recommender sys-
tems, powering the recommendations of many of
the key web companies. Likeminds, created in
1997 by O’Reilly and AOL, also specialized in com-
mercializing collaborative ltering tools to help
other websites offer personalized recommenda-
tions. In 1998, Firefly Network was acquired by
Microsoft and its technology seeded Microsoft’s
Passport (now known as LiveID).
As the millennium ended, new web portals and
e-commerce sites were in erce competition to
attract users, differentiating their services by
actively developing their recommendations. Very
soon a number of software manufacturers of appli-
cation and content management servers (such as
E.piphany, Blue Martini, Vignette, and others)
complemented their product suites with recom-
mendation or personalization modules. Table 1
gives a simple overview of the relevant businesses.
By the end of 1999 many popular websites were
already buying recommender system services or
were planning to roll out their own service. How-
ever, in March 2000 the dot-com bubble burst and
the e-commerce domain collapsed, with compa-
nies losing more than 5 trillion dollars of market
value from March 2000 to October 2002.
Although the views and experiences of the rele-
vant vendors may vary to some extent, there are a
number of prominent lessons learned during the
rst wave that shaped the commercial ventures of
the second wave of recommender systems (see
table 2).
The Second Wave of
Recommender Systems
The number of Internet users and the number of
websites never stopped growing despite the col-
lapse of the Internet sector. At the end of 2000
there were more than 250 million Internet sub-
Articles
FALL 2011 21
First Wave of Recommender Systems Vendors
Companies founded between 1991 and 2000
ATG Dynamo (acquired by Oracle), Blue Martini (acquired by Escalate Retail), Braisins, Broadvision, E.piphany
(acquired by SSA Global Technologies acquired by Infor), Firefly Network (acquired by Microsoft), IBM WebSphere
Personalization, Like Minds (acquired by Andromedia), Manna, Open Sesame (acquired by Bowne Inc.), Net
Perceptions, Vignette (acquired by Open Text).
Second Wave of Recommender Systems Vendors
Companies founded between 2001 and 2010
Adobe Recommendations (before Omniture recommendations), Aggregate Knowledge, Amadesa, Avail Intelligence,
Barilliance, Bay Note, Blue Know, Certona, Changing Worlds (acquired by Amdocs), Choice Stream, Clever Set
(acquired by ATG Dynamo), Commendo, Criteo, Directed Edge, Easyrec, 4 Tell, Gravity, Ericsson SDP, I Go Digital,
Iletken, Istobe, July Systems, Loomia, Match Mine, Media Unbound (acquired by Rovi), My Buys, Olista (acquired
by Connectiva), Pontis, Predictive Intent, Prediggo, Rich Relevance, 7 Billion People, Strands, We Like, Xiam
(acquired by Qualcomm), Xtract.
Table 1. Most of the Key Vendors of Recommender Systems Founded in the Last Two Decades.
Figure 1. A Screenshot from Movielens.
This screenshot is an example of a rst-wave recommender system for movies
and the rst high quality source of data for recommender system research.
(Used with permission.)
scribers (see gure 2). These users became more
eager to share their own content through personal
web pages, blogs, and online communities. The
amount of user-generated content soon surpassed
professional media content, and the social or web
2.0 revolution started. While the market for vari-
ous digital media content exploded, other prob-
lems soon surfaced as user-generated content
exhibited some of the same issues encountered in
collaborative-ltering-based recommender systems
(Su and Khoshgoftaa 2009). In particular, problems
related to scalability, sparsity, and shilling became
much more challenging to solve.
In 2002, 2003, and 2004, Friendster, MySpace,
and Facebook were launched (respectively). Many
people started to amass online friends and broad-
cast each detail of their personal lives. Social net-
working quickly became a vital personal commu-
nication channel as it substituted for email, Short
Message Service (SMS), or even face-to-face com-
munication for many people.
As the usage of their systems grew, web compa-
nies started to analyze recommendation algorithms
in order to make them more scalable, robust, and
less prone to bias or sabotage. Amazon.com,
Netflix, and Yahoo! Music popularized their recom-
mender systems and solved many of the scalability
issues of the rst recommender systems (Linden,
Smith, and York 2003). Five-star rating schemes
combined with purchase and rental behavior pat-
terns became a common source of data for most big
e-commerce websites. However, despite the growth
and improvements of the underlying algorithms
and architecture, recommendation technologies
were still too complex or expensive to be affordable
by smaller online retailers.
Rating-based collaborative ltering systems
relied on their users to rate dozens of items before
the service could provide appropriate suggestions.
The systems also tended to be static, with no abil-
ity to adapt to a given user’s changing tastes. This
caused problems for domains where many items
were unrated, and a user’s taste could be expected
to change quickly. Music was one such domain,
and an important component of the burgeoning
digital media market. This opportunity led to the
development and adoption of techniques that
minimized the number of explicit interactions to
train a recommender, similar to the approach
introduced by the experimental Ringo system sev-
eral years prior. Sites like Audio Scrobbler (later
merged with Last.fm), Music Mobs, MOG, and
MyStrands started tracking what people were lis-
tening to on their computers to tacitly indicate
interest, and to automatically generate social-based
recommendations with little to no prior informa-
tion required from the user.
Pandora (formerly known as the Music Genome
Project) is noteworthy in the way that it integrat-
ed a recommender system into its music-streaming
service. Instead of using listening behavior or user-
generated playlists to relate music, Pandora used
dozens of music experts that rated songs following
hundreds of preestablished attributes. Then it used
a basic rating system (thumbs up, thumbs down)
to tailor future recommendations for each user.
However, despite operating for nearly 10 years,
Pandora still can only offer recommendations for
around one million songs. It is currently an impor-
tant example in the long debate on the trade-off
between scalability and the quality of expert-based
recommendations, which goes beyond the scope
of this article.
In addition to the differences in music recom-
Articles
22 AI MAGAZINE
Integration Is Much Harder Than It Looks
Schedule enough time upfront for all the last-mile problems you will nd. Much more if you try to integrate in a
constantly changing environment such as the one many web companies had when initial web technologies were
still under development.
Customer’s In-House Competition Is Fiercer Than Market Competition
Integrating a recommender system implies parallel battles with the IT and marketing departments in midsize
companies. Don’t waste your time with big web companies as they see recommendations as too strategic to be
delegated.
Cold Start Is An Issue If Your Algorithm Is Exclusively Fed by the System’s Own Data
If the recommender system is purely based on your own data, make sure it is started with a substantial initial
amount, or wait to unveil the service until enough data is collected. Otherwise the initial experience for the users
will be very poor.
Don’t Forget About Scalability
Deciding what processes a recommender system must run online and which ofine is crucial to be able to scale.
The rst generation of recommender systems tried to solve everything online or with memory-based approaches
that did not scale when the number of users or number of items to recommend increased.
Table 2. Some Lessons Learned During the First Wave of Recommender Systems.
mendation approaches, online news also popular-
ized different forms of recommender systems. Ini-
tially, Findory offered a personalized news service
utilizing content-based collaborative ltering tech-
niques for news story recommendation (gure 3).
However, online news sites such as Digg, Reddit,
and Hacker News soon became the more popular
means for nding relevant news. Instead of more
personalized recommender algorithms, stories
were promoted and recommended primarily by
global popularity trends.
In 2006 Strands organized a summer school on
the “Present and Future of Recommender Systems”
that brought together researchers, practitioners,
and students. Building on its success, the Associa-
tion for Computing Machinery (ACM) recom-
mender systems conference “RecSys” was estab-
lished, held in Minneapolis (2007), Lausanne
(2008), New York (2009), Barcelona (2010), and
Chicago (2011) underway. Collaborative ltering
has been the method underlying most recom-
mender systems studies, as can be seen in the
weighted tag cloud visualization in gure 4.
While research on algorithms has captured most
of the attention for recommender system studies,
the rise of social networking indicated that simply
keeping tabs on your friends’ listening, watching,
and reading habits is a more natural way to keep
informed of the latest trends. Undirected word-of-
mouth suggestions were automatically transmitted
through systems like Facebook’s News Feed with-
out algorithmic processing but embedded in a con-
venient interface.
At the end of 2006, Netflix challenged the world
to improve the accuracy of its movie recommen-
dation system by 10 percent. It created a contest
that brought lots of attention to recommender sys-
tems, not just for the $1,000,000 prize, but for the
fact that a popular service like Netflix was willing
to invest in such an exotic scheme just to improve
the quality of its recommender service. In 2009 the
BellKor “Pragmatic Chaos” team won the contest
(Töscher, Jahrer, and Bell 2009). Their efforts may
have set the bar for a best-in-class recommenda-
tion algorithm for movies, but practitioners have
noted that algorithmic precision is just one of
many factors that affect a user’s adoption of a rec-
ommendation, and other issues such as interface
design, long-term performance evaluation, or con-
text-awareness are also prominent parts of a rec-
ommender system outcome.
A number of companies saw the opportunity
Articles
FALL 2011 23
Dozens of vendors
~ 100 vendors
~ 1000 vendors
Recommendation Technology Cost
Internet Users
Data
Mobile Subscribers
1991 2000 2001 2010 20111st Wave 2nd Wave 3rd Wave
2B
500M
5B
Hundreds of thousands $
Most content generated by experts —
Managing Gigabytes
Most content generated by users —
Managing Terabytes
Most content generated by machines —
Managing Exabytes
A few thousand $
250M
Figure 2. Internet Trends 1991–2010.
for offering recommendations as an affordable
software service (Clever Set, Loomia, Rich Rele-
vance, Strands, and others) to smaller web com-
panies. See table 1 for a more complete list. The
common approach was for the vendor to offer a
remote application programming interface (API)
that collected user activity on web pages (click
stream, ratings, purchases, and others), and gen-
erated on-the-fly recommendations relevant to
the users’ behavior. Despite strong evidence that
different product and service domains (such as
movies, music, news, and others) have different
needs for recommender services, many software-
as-service recommender system vendors general-
ized their solutions with only minor tailoring for
specific domains. An additional problem for these
vendors arose when web usage analytics compa-
nies started to incorporate the same recom-
mender techniques as part of their service offer-
ings.
Recommender systems have now become well
studied both theoretically and practically. Table 3
shows a number of lessons learned deploying rec-
ommenders during the second wave. However,
despite all the progress made, most commercial
recommender systems have taken the business’s
side and their recommendations are rmly linked
to short-term sales of products or services instead
of considering real users’ contexts and needs. Thus,
we predict a new wave of recommender systems on
the user’s side that will better deal with informa-
tion overload.
Articles
24 AI MAGAZINE
Figure 3. A Screenshot from Findory.
This gure is an example of a second-wave recommender system for news articles — a realization of Negroponte’s (1995) “Daily Me”
and a precursor of current personalized newspapers for the iPad. (Used with permission.)
The Third Wave of
Recommender Systems
With more than 5 billion mobile subscribers, and
more than 2 billion Internet users (see figure 2),
we live in a era where information has gone from
scarce to superabundant (Cukier 2010).
Technically savvy individuals are already utiliz-
ing a broad range of services, but must manage and
prioritize the flows of information and suggestions
on their own. On a typical day a given person may
already be receiving a multitude of recommenda-
tions from mobile and nonmobile computing plat-
forms. Personal sleep aid systems monitor sleep
Articles
FALL 2011 25
Figure 4. Tag Cloud Created from the Titles of All the Articles Published at the ACM RecSys Conferences Using IBM Wordle.
Hybridization Solves Many of the Issues with Recommender Systems and Wins Contests Too
Using exclusively one recommender systems technique produces less robust solutions. An ensemble of techniques
has shown to be successful and easily improve accuracy.
Easy to Integrate Also Means Easy to Remove
Offering software-as-service recommender systems reduces the burden of integration and makes it easier to get the
customers but also makes it easier for the customers to switch to another vendor.
Integration with Customer Care Services Is a Pending Subject
You can buy hundreds of products from Amazon or Apple and get better and better recommendations but
unbelievably their customer care services do not know if you are a good or bad customer. Integration of CRM and
recommender systems is a challenging issue.
Right Balance of Attention Needs to Be Given to Algorithms and User Experience
Although most research focuses on algorithms, user interfaces and user experience have a heavier weight on the
nal performance of recommender systems. Algorithms are mature, but an engineering approach to know how and
when to recommend is still immature.
Cold Start Is Not a Problem If You Think Out of the Box
With the web 2.0 revolution, user-generated content made it possible to overcome the cold-start problem by just
considering relatedness among items and people on the social web.
The Best Nearest Neighbor Algorithm Is the Social Graph
Do not spend time calculating nearest neighbors and just let people tell you!
Rating-systems-explicit rating systems are very often not well dened semantically and lead to poor
recommendations. Implicit feedback on actual preferences have resulted in much better recommendations than
explicit ratings.
Recommendation Technologies Have Been Commoditized
Technologies that were licensable for hundreds or even millions of dollars only 10 years ago are now freely
available (See gure 2). This, together with the high number of players and the difculties to evaluate or
differentiate their products, has been one of the main causes that pushed many vendors out of business as the price
for recommendation technologies dropped to levels that made operating their businesses unsustainable.
Table 3. Some Lessons Learned During the Second Wave of Recommender Systems.
such as map-reduce, provide super-
computing power at a fraction of its
previous cost. At the same time, large
varieties of information are being made
available online through public APIs.
In addition, new computing libraries,
frameworks, and services such as SciPy,
Apache Mahout, and Google Predic-
tion API are making it easier for devel-
opers to utilize sophisticated algo-
rithms on massive amounts of data.
This will enable the third-wave recom-
mender system ventures to focus more
on the user experience and less on the
technical side of algorithms.
User’s Long-Term Goals
Finally, third-generation recommender
systems will have a greater emphasis
on handling long-term user goals and
constraints, such as monthly budget
limits, or scheduling constraints on the
user’s personal calendar. Most current
recommendation systems focus on
short-term interactions: cross-selling,
up-selling, or suggesting media for
immediate consumption. Removing
the emphasis on short-term prot will
allow recommendation systems to
encourage discovery and, we hope,
maximize a user’s long-term satisfac-
tion.
Conclusion
We characterize two basic types of rec-
ommender systems that have been
deployed in practice. The rst type
includes recommenders that take the
user’s side (such as Siri) and are operat-
ed on the user’s computer (smartphone
or tablet) by a completely independent
service provider. The second type cor-
responds to recommenders that take
the business side (for example, Ama-
zon’s recommender). These recom-
menders are put in place and operated
by the retailer of the product being rec-
ommended or operated on behalf of
the retailer.
As technology advances and recom-
mender systems become more com-
monplace for the public, we may now
be on the verge of a third wave of rec-
ommender systems that considers the
goals and activities of users in addition
to their preferences and needs. The
third wave of recommender systems
will be capable of learning about a
cycles, making suggestions for healthy
patterns of resting (such as Zeo, Fitbit).
Personal workout recommenders may
display jogging route recommenda-
tions (MapMyRun, Nike+, Strands, and
others).
Personal news sites recommend arti-
cles from major news outlets or inde-
pendent bloggers (such as Hacker
News, Reddit, Digg, and others), while
email systems help prioritize and lter
our email (for example, Google gmail
spam ltering and priority inbox). Per-
sonal assistant software schedules
meetings or nds reservations for a
restaurant (for example, Siri; see gure
5). Entertainment providers recom-
mend movies and television programs
to watch (such as Netflix).
Instead of providing yet another
domain-specic recommendation serv-
ice, we foresee a third wave of recom-
mender systems that will return to the
metaphor of personal assistant (Negro-
ponte 1995, Maes 1994). In this
arrangement, recommender systems
once again persist with and act on
behalf of individual users to help them
cope with the problem of information
overload across a range of domains. As
the pace of data growth and consump-
tion increases, the added dimensions
of cloud and mobile computing enable
better personalized recommendations
that can be provided in a larger variety
of relevant locations. Out of all of the
recently started services, we see Siri
(acquired by Apple) as coming closest
to what the third-generation recom-
mender systems will look like. In gen-
eral, the third wave of recommender
systems will have these prominent
dimensions: contextual awareness,
cloud-based stores and computation,
and emphasis on users’ long-term
goals.
Contextual Awareness
New recommenders will use clues from
mobile devices to generate contextually
appropriate recommendations. Smart-
phones provide a more convenient and
persistent basis for recommendations,
as well as having access to new infor-
mation, such as calendar, email, noti-
cations, and user GPS coordinates, that
helps drive timely and contextually
appropriate suggestions. Additional
technologies such as eye tracking, ges-
ture detection, or skin-tension meas-
urement are expected to play a role in
the construction of such personal assis-
tants, utilizing minimally invasive bio-
metric feedback to characterize user
affect or emotional state, according to
relevant research by Rosalind Picard
(2000). The advanced features and pro-
grammability of modern smartphones
make them the perfect channel to drive
the third generation of recommender
systems.
Cloud-Based Stores
and Computation
The amount and diversity of data that
new recommender systems will process
is going to increase dramatically while
the cost to process or store will decrease
signicantly. During the rst two
waves, the IT investment required for
large-scale recommendation systems
was prohibitive for many vendors to
pursue direct-to-consumer initiatives.
However cloud computing has
reshaped the landscape. On-demand
computational resources, such as Ama-
zon EC2 or Rackspace, combined with
distributed-computing paradigms,
Articles
26 AI MAGAZINE
Figure 5. A Screenshot from Siri.
This is an example of a third-wave recom-
mender system and a precursor of the
mobile personal assistant revolution.
user’s contextual behaviors and prefer-
ences to anticipate a user’s actions at
any point in time and act upon them.
It is an exciting time for recom-
mender systems. The current
confluence of ever faster networks,
massive adoption of mobile devices,
the rise of cloud computing and social
technologies open a new window of
opportunity. Companies like Facebook,
Google, and Apple are well positioned
given their dominance in social media,
search, and mobile technology to capi-
talize on some of this opportunity. But
the current industry scenario could not
be better for both researchers and prac-
titioners who want to address new
challenges for recommender systems.
There are ample opportunities for nim-
ble entrepreneurs to connect data,
algorithms, devices, and user needs to
create solutions that carve out signi-
cant new market space.
References
Amatriain, X.; Torrens, M.; Resnick, P.; and
Zanker, M., eds. 2010. Proceedings of the
Fourth ACM Conference on Recommender Sys-
tems (RecSys ’10). New York: Association for
Computing Machinery.
Bergman, L. D.; Tuzhilin, A.; Burke, R. D.;
Felfernig, A.; and Schmidt-Thieme, L., eds.
2009. Proceedings of the Third ACM Confer-
ence on Recommender Systems (RecSys ’09).
New York: Association for Computing
Machinery.
Cukier, K. 2010. Data, Data Everywhere: A
Special Report on Managing Information.
The Economist. 25 February.
Goldberg, D.; Nichols, D.; Oki, B. M.; and
Terry, D. 1992. Using Collaborative Filtering
to Weave an Information Tapestry. Commu-
nications of the ACM 35(12): 61–70.
Jannach, D.; Zanker, M.; Felfernig, A.; and
Friedrich, G. 2010. Recommender Systems: An
Introduction. New York: Cambridge Universi-
ty Press.
Kirn, W. 2010. My Cart, My Self. The New
York Times Magazine. 30 December.
Konstan, J. A.; Riedl, J.; and Smyth, B., eds.
2007. Proceedings of the 2007 ACM Conference
on Recommender Systems (RecSys ’07). New
York: Association for Computing Machin-
ery.
Kozierok, R., and Maes, P. 1993. A Learning
Interface Agent for Scheduling Meetings. In
Proceedings of the 1993 International Work-
shop on Intelligent User Interfaces. New York:
Association for Computing Machinery.
Linden, G.; Smith, B.; and York, J. 2003.
Amazon.com Recommendations: Item-to-
Articles
FALL 2011 27
Item Collaborative Filtering. IEEE Internet
Computing 7(1): 76–80.
Maes, P. 1994. Agents That Reduce Work
and Information Overload. Communications
of the ACM 37(7): 30–40.
Negroponte, N. 1970. The Architecture
Machine: Toward a More Human Environment.
Cambridge, MA: The MIT Press.
Negroponte, N. 1995. Being Digital. New
York: Vintage Books.
Palme, J. 1984. You Have 134 Unread Mail!
Do You Want to Read Them Now? In Pro-
ceedings of the IFIP WG 6.5 Working Confer-
ence on Computer-Based Document Sevices,
175–184. Amsterdam, The Netherlands:
Elsevier Noth-Holland, Inc.
Picard, R. 2000. Affective Computing. Cam-
bridge, MA: The MIT Press.
Pu, P.; Bridge, D. G.; Mobasher, B.; and Ric-
ci, F., eds. 2008. Proceedings of the 2008 ACM
Conference on Recommender Systems (RecSys
’08). New York: Association for Computing
Machinery.
Resnick, P., and Varian, H. R. 1997. Recom-
mender Systems. Communications of the
ACM 40(3): 56–58.
Resnick, P.; Iacovou, N.; Suchak, M.;
Bergstrom, P.; and Riedl, J. 1994. Grou-
pLens: an Open Architecture for Collabora-
tive Filtering of Netnews. In Proceedings of
the 1994 ACM Conference on Computer Sup-
ported Cooperative Work, CSCW ’94, 175–
186. New York: Association for Computing
Machinery.
Ricci, F.; Rokach, L.; Shapira, B.; and Kantor,
P. B., eds. 2011. Recommender Systems Hand-
book. Berlin: Springer.
Shardanand, U., and Maes, P. 1995. Social
Information Filtering: Algorithms for
Automating “Word of Mouth.” In Human
Factors in Computing Systems CHI 95 Confer-
ence Proceedings, 210–217. Reading, MA:
Addison-Wesley.
Su, X., and Khoshgoftaa, T. M. 2009. A Sur-
vey of Collaborative Filtering Techniques.
Advances in Articial Intelligence 2009: 1–20.
Töscher, A.; Jahrer, M.; and Bell, R. M. 2009.
The BigChaos Solution to the Netflix Grand
Prize. Technical Report, Commendo
Research & Consulting GmbH, Köflach,
Austria.
Francisco J. Martin is a scientist entrepre-
neur passionate about machine learning
and distributed systems. He has cofounded
three AI-related startups (Intelligent Soft-
ware Components, SA, in 1999, Strands,
Inc., in 2004, and BigML, Inc., in 2011) and
directly raised more than $75 million in
venture capital. He holds a ve-year degree
in computer science from Technical Univer-
sity of Valencia and received his Ph.D. (sum-
ma cum laude) from Technical University of
Catalonia.
Justin Donaldson is cofounder and presi-
dent of BigML, Inc., a web service for cloud-
based machine learning of large sets of data.
He has worked for six years on recom-
mender systems at Strands.com, focusing on
related analysis, interface, and visualization
methods. He holds a B.S. in computer sci-
ence from DePauw University, an M.S. in
HCI, and a Ph.D. in informatics from Indi-
ana University.
Adam Ashenfelter is a cofounder at BigML,
Inc. He holds a B.S. and M.S. in computer
science from Oregon State University. He
spent seven years working on recommenda-
tion systems, machine learning, and data
fusion projects while at CleverSet, Inc., and
Strands, Inc.
Marc Torrens received a MSc degree in com-
puter science from the Universitat Politèc-
nica de Catalunya (UPC), Barcelona, Catalo-
nia, in 1997, and a PhD from the École
Polytechnique Fédérale de Lausanne (EPFL),
Lausanne, Switzerland, in 2002. He
cofounded Iconomic Systems SA in 1999
and served the company as technology of-
cer. Iconomic Systems developed software
for planning travels on the Internet and was
successfully sold to i:FAO AG, Germany. In
2004, Torrens cofounded Strands, Inc., in
Oregon, USA. Strands develops recommen-
dation and personalization technologies to
help people discover new things. He is cur-
rently serving Strands as chief innovation
ofcer. Torrens has published more than 20
referred papers on his work in international
conferences and journals on articial intel-
ligence and usability.
Rick Hangartner is chief scientist at Strands
Labs, Inc., and an adjunct faculty member
at Oregon State University. He received his
BA and MS from the University of South
Florida and his PhD from Oregon State Uni-
versity. His research focuses on statistical
inference, nonparametric data modeling,
and adaptive systems.