The Art, Science and Business of Recommendati...

In October last year,Netflix launched anunusual contest. The online movie rental company is offering 1 million dollars to anyone who can improve their recommendation engine by 10%. Netflix is known for its innovation andbold moves and in the grand scheme of things, $1M is not a lot of money for such a business.
The competition is still running (it "continues through at least October 2, 2011"), so is this a publicity trick or an attempt to do research on the cheap? Is better recommendations something that Netflix really needs or is it just nice to have? Today Netflix is facing a challenge from the awakened giant BlockBuster, so it is certainly looking for a competitive edge. A great recommendation system can retain and attract users to the service. For example when a user returns a movie, he/she is recommended another movie they might like - which increases the likelihood of return business.
Browsing and Recommendations
A good recommendation engine can make a difference not just for Netflix, but for any online business. This is because there are two fundamental activities online - Search and Browse. When a consumer knows exactly what she is looking for, she searches for it. But when she is not looking for anything specific, she browses. It is the browsing that holds the golden opportunity for a recommendation system, because the user is not focused on finding a specific thing - she is open to suggestions.
During browsing, the user‘s attention (and their money) is up for grabs. By showing the user something compelling, a web site maximizes the likelihood of a transaction. So if a web site can increase the chances of giving users good recommendations, it makes more money. Obviously this is a difficult problem, but the incentive to solve it is very big. The main approaches fall into the following categories:
Personalized recommendation - recommend things based on the individual‘s past behavior Social recommendation - recommend things based on the past behavior of similar users Item recommendation - recommend things based on the thing itself A combination of the three approaches above
We will now explore these different approaches by looking at old-timers like Amazon and newbies like Pandora and del.icio.us.
Amazon - The King of Recommendations
Amazon is considered a leader in online shopping and particularly recommendations. Over the last decade the company has invested a lot of money and brain power into building a set of smart recommendations that tap into your browsing history, past purchases and purchases of other shoppers - all to make sure that you buy things. Lets take a look at various pieces of Amazon‘s recommendation system to get an insight on how they work. Here are the sections that are shown in the main area of my Amazon account when I login:

The section above shows Social recommendations. Notice that it is very analytical, giving me a statistical reason for why I should buy this item. Also note that this recommendation is also a Personalized recommendation, since it is based on an item that I clicked recently.

The section above shows Item recommendation based on New Releases. Clicking on the Why is this recommended for you? link takes me to a view of my purchasing history. So this recommendation is also a Personalized recommendation, since it is based on my past behavior.
There are four more sections offered on the page and each of them leverages different combinations of the personalization mechanisms described above. We summarize them in the table below:

Not surprisingly, the system is symmetric and comprehensive. All recommendations are based on individual behavior, plus either the item itself or behavior of other people on Amazon. Whether you like to buy something because it is related to something that you purchased before, or because it is popular with other users, the system drives you to add the item to the shopping cart.
Beyond Amazon
The Amazon system is phenomenal. It is a genius of collaborative shopping and automation that might not be possible to replicate. This system took a decade for Amazon to build and perfect. It relies on a massive database of items and collective behavior that also "remembers" what you‘ve done years and minutes ago. How can new companies compete with that?
Surprisingly, there is a way. The answer is found in a subject that has little to do with online shopping - genetics. As you know, this science studies how pieces of DNA, called genes, encode human traits and behavior. For example, members of a family look and behave alike because they share a certain subset of genes. Genetics as a science has been around for over 150 years and has been a powerful tool for both medicine and history. But on January 6, 2000 things took an unexpected turn - Tim Westergren and his friends decided to apply the concepts of genetics to music.
Pandora - The Recommendation System Based on Genetics
TheMusic Genome Project was launched to decompose music into its basic genetic ingredients. The idea behind it is that we like music because of its attributes - and so why not design a music recommendation system that leverages the similarities between pieces of music. This kind of recommendation engine falls into the Item recommendation category. But what is new and profound here is that similarity of an item like a piece of music needs to be measured in terms of its "genetic" make up.

After years of struggle and processing massive amounts of music, the project accumulated enough data and launched the service calledPandora. Pandora became a hit because of its precision and low cost of entry. The user just needs to pick one artist, or a song, to create a station that instantly plays similar music.
This kind of instant gratification is difficult to resist. The fact that Pandora understands what makes music similar allows it to hook the user without having to learn what this user likes. Pandora does need the user‘s tastes or memory, it has its own - based on music DNA. Sure, sometimes it might not be perfect, as the user‘s taste might not be perfectly addressed. But it is rarely wrong.
The natural question is can this genes-based approach be applied to other areas - like books, movies, wines, restaurants or travel destinations? What constitutes genes for each category? For example, can we say that for wine, the genes might be things that describe how wine tastes: blackberry, earthy, fruity, complex, blend, etc. And for a book, can the genes be phrases that describe the plot? So if the genes are the attributes of the object that make it unique in our mind, we should have no problem coming up with genes for various things. In the past few years we have been doing this a lot online. It‘s called tagging!
Del.icio.us - Can Tags Become Genes?
Pandora had a big startup cost, because thousands of pieces of music had to be manually annotated. The social bookmarking phenomenondel.icio.us took a different approach - let people annotate things themselves. This self-organizing approach has worked really well, and del.icio.us quickly became popular among early adopters. Today, del.icio.us is considered to be more than bookmarking destination - it is also a news site and a search engine. But is del.icio.us a recommendation system?

The answer is yes. There is a basic recommendation system based on one gene - a single tag. For example, in the picture above we see popular links for the linux tag and we also see related tags like open source and ubuntu. But a much more exciting recommendation system is based on matching multiple tags. Unfortunately, the current heuristic does not always work, which is why it is not obvious. But luckily, it did work for the Read/WriteWeb page and generated a great list of similar blogs (see "related items" below):

So the del.icio.us approach holds intriguing possibilities of self-organizing classification and recommendation systems. With enough users and more tweaking, social tagging can result in a system that works equally well for books, wine and music. Provided, of course, that tags are so good that they become genes!
Conclusion
Recommendation engines are important pieces of online commerce systems and their user experience. Retailers have a big incentive to provide recommendations to those users who are "just browsing", to drive them towards a transaction. Amazon.com, the leader in the space, has a very compelling personalization offering. The problem that other retailers face is lack of user information and infrastructure.
Recent approaches to recommendation engines, like the genetics-inspired Pandora and social tagging pioneered by del.icio.us, are intriguing. These approaches hold the promise to provide instant gratification, without asking the user to reveal her preferences and past history. Regardless of how things unfold in the future, Amazon, Pandora and del.icio.us are examples of extraordinary recommendation technologies. We commend them and are watching in fascination for what is coming next.
TrackBack
»Browsing for workshops on Go’n’Shoot from Go‘n‘Media
Tracked on January 17, 2007 11:45 PM
»Weekly Linkage [01-19-07] from Experience Planner by Scott Weisbrod
Tracked on January 19, 2007 8:36 AM
»More about Netflix Prize and recommendation from leafar
Tracked on January 19, 2007 5:28 PM
»The New Face of Amazon - Tags, Ajax, Plogs & Wikis from Read/WriteWeb
Tracked on January 24, 2007 9:43 PM
Comments
Subscribe to comments for this post ORSubscribe to comments for all Read/WriteWeb posts
# 1
You might also include AggregateKnowledge‘s service in this discussion as they have found a way to enable Amazon-like capability for merchants and publishers by leveraging the network effect of information gathering. Definitely one to watch.
Posted by:P-Air |January 17, 2007 12:43 AM
# 2
one group to seriously factorhttp://seeqpod.com
Posted by: zaction |January 17, 2007 1:59 AM
# 3
Keep an eye out for MediaChoice, they have a patented recommendation systems that is very accurate. Been around for years and are working on the Netflix contest.
Posted by:Recommender |January 17, 2007 4:03 AM
# 4
Thanks for the thoughtful write-up, I especially liked how you differentiate the different flavors of recommendations available. Perhaps consider discussing the feedback loop to tune the quality of recommendations in each type?
You may want to check outhttp://www.loomia.com/, which is a turn-key service to provide recommendations to media and retail web sites. (Full disclosure: I‘m a founder)
Posted by:David |January 17, 2007 7:44 AM
# 5
I don‘t think it‘s necessarily the case that users who know what they want search while others browse. In my experience there‘s substantial spill in both directions and the choice of search v. browse has more to do with behavioral preferences and confidence in the underlying search engine than anything else.
Vistors to e-commerce sites are always open to suggestion/education, hence any up-sell/cross-sell strategies that leverage recommendation should address searchers and browsers, and a good UI should offer both avenues on every web page.
Posted by:Ranjit Padmanabhan |January 17, 2007 9:34 AM
# 6
Netflix, like Amazon, has figured out that helping people make choices is the key to market success. Netflix can see beyond Blockbuster and is probably now more worried about Google, now that they‘re both in the video download business. People won’t buy things if they can’t make fundamental choices, and most people won’t make as many choices if they don’t get a little help -- without being barraged with too many questions.
That’s the trick. We don’t want to answer too many questions about what we’re looking for. Half the time we don’t know what we’re looking for (“just browsing???) and the other half of the time we are incapable of describing it – or maybe that’s just me. Harvard professor Daniel Gilbert details people’s inability to understand or describe what makes them happy in his book, “Stumbling on Happiness???. We think that many of those same mental barriers exist for people shopping or searching online, and we believe a good recommendation platform is essential to overcome them.
Posted by:Rob Rustad |January 17, 2007 9:50 AM
# 7
"Better Recommendation" seems a great idea but not sure it will have the desired intentions or maybe I don‘t understand the objectives. I can only imagine that this process would involve gathering information on a particular usage pattern overtime. With that then the SE would be able to present more relevant search results. However this artificial narrowing to me takes away the impulsiveness.
Posted by:Adrian Keys |January 17, 2007 10:24 AM
# 8
@Rob Rustad
Thats why Amazon‘s People who looked at X bought Y steers them towards a choice of buying.
Alex
Posted by:Alex Iskold |January 17, 2007 11:19 AM
# 9
Thanks for the article, i am just in the process of redesign a local online bookstore - it is great to read how top websites use this powerful tool of recommendation.
Posted by:Alex |January 17, 2007 11:23 AM
# 10
I don‘t like most recommendation engines, though Amazon might be an exception. Most of them aren‘t very good.
Posted by:David Mackey |January 17, 2007 4:48 PM
# 11
@David,
Thats because they are off for you :)
What do people think about Pandora‘s approach? Do you think that pieces of music are highly similar objectively or is it all subjective?
Alex
Posted by:Alex Iskold |January 17, 2007 5:11 PM
# 12
I think Pandora is amazing.
I find that I‘ll try a few different band stations, find some I really like, and play them over and over.
It‘s great because it‘s both random and predictable-- I know that if I‘m playing X station, it‘ll eventually play Y song, and that excites me and motivates me to keep listening.
I find the pieces of music are similar sort of objectively- there‘s a definite "vibe" in my mind about each of my stations. If you pick a band like black eyed peas with many styles and remixed sounds, it doesn‘t work as well.
I tried last.fm multiple times and it‘s always failed. Either runs out of similar songs, plays weird non-songs, or plays horrible songs. Not sure what I‘m doing wrong there, as many people seem to like it, but I give up. I‘m Pandora for life.
Great interface too.
And wow I gotta say this-- I look forward to ADs on Pandora because sometimes they come with great stations linked to them. The Elantra Ads have the best.
Posted by:Pamela Fox |January 18, 2007 1:34 AM
# 13
For recommendation Search Engines, I would suggest looking at Gnod (www.gnod.net) and Music Map (www.music-map.com. Another excellent recommendation SE is LivePlasma(www.liveplasma.com). Finally, please check out What to Rent, a fun movie recommendation SE(www.whattorent.com). -Charles Knight SEO
Posted by: Charles Knight |January 18, 2007 5:43 AM
# 14
Great Article Alex,
However, I think you missed where the future lies for recommendation engines and personalization.
Recommendation and personalization technology that works across all three screens (your computer, mobile devices and perhaps most importantly your television) will become increasingly necessary as the long-tail of content expands beyond the web and into your living room.
Today’s consumer has access to a bewildering array of programming options when they turn on their television. With programming available on a seemingly endless amount of channels and on-demand services the need to simplify and streamline this superabundance of choice has never been greater. Without guidance and organization, superabundance is overwhelming and chaotic. Television viewers have always enjoyed this catered experience. Now, they need that special attention to insure that their viewing experience is truly relevant. Cable and satellite providers that are answering their call have realized that by providing personal touch they can retain customer loyalty.
One company leading the charge in this television revolution is Cambridge based ChoiceStream.
ChoiceStream’s premier personalization solution is used by the world’s largest entertainment, e-retail, TV and mobile brands, including AOL, Yahoo!, DirecTV, Overstock.com, Movielink, Akimbo and eMusic.
ChoiceStream’s RealRelevance is the only personalization system of its kind to profile both people and content. ChoiceStream automatically classifies all types of content—music, movies, TV shows, games—based on its various characteristics, otherwise known as attributes. ChoiceStream then learns from these attributes and each consumer’s interaction with the content to better understand the consumer’s unique tastes and preferences. ChoiceStream’s patent-pending personalization, Attributed Bayesian Choice Modeling, enables ChoiceStream to learn about a consumer quickly and accurately, avoiding out-in-left-field results that can immediately undermine consumers’ faith in the recommendations they’re receiving.
This personalization technology is currently in place on the Internet and has been changing the distribution models for today’s largest online content providers and retailers for some time. The technology only now is beginning to be implemented by the leading satellite and cable providers. DirecTV recently implemented an online personalized My TV Planner using ChoiceStream’s technology to recommend programming of interest to its 15.5 million viewers. Although these recommendations are currently only available online, this is likely the first step in the migration to the set-top box. Recommendation engines may very well be the answer to maximizing the television experience; just as search engines have been the answer for the Internet.
Search engines have been the most crucial element of consumer discovery on the Internet, but in this newly interactive television environment they lose some of their effectiveness. Search engine algorithms are based on text, information, and links. This philosophy doesn’t necessarily produce the best results in a television environment. Television viewers don’t want to search for programming in an empty box. They want relevant choices provided without having to search. Therefore the idea of pushing programming at viewers based on their previous choices and individual tastes is being adopted.
The key to pushing this relevant programming is the development of a rich classification system for the programming itself. To do that, cable and satellite leaders are enlisting the help of innovative technology vendors to build characteristic indexes around programming and movies available to users. This index contains more than the kind of information a search engine would utilize. For example, when implemented into an on-demand movie service, the index contains certain attributes of each movie. Is the movie thought provoking? Action packed? Satirical?
A viewer may not truly understand the reasons why they chose “The Breakup??? and were subsequently recommended “Thank Your for Smoking.??? The correlation between a romantic comedy and a satire isn’t necessarily obvious to the naked eye. However, the personalization technology can easily dig deeper into these movies attributes and determine that in fact there is a correlation between the two choices. In this case, both films take objective looks at dramatic conflict and although vastly different both leave viewers examining failed relationships. Some may argue that in this sense cable and satellite providers may know you, or at least your choices, better than yourself.
Personalization also offers the key to building more robust advertising models in a world of content everywhere. The mass adoption time-shifting television viewing, and the proliferation of video and programming available online, has given consumers more control than ever over what they watch, when they watch it and how. It’s safe to say the convergence of the three screens means that consumers will no longer sit passively while advertising is pushed their way. Advertisers are using personalization – preference-driven targeting – to help captivate these unleashed consumers, winning their attention with messages that are directly relevant to their needs and interests. Ensuring a more relevant, more rewarding advertising experience is a first step toward building loyalty between the consumer and advertiser.
(Full disclosure: I work with ChoiceStream)
Posted by:RealRelevance |January 18, 2007 6:49 AM
# 15
You might want to check out Loomia which anyone can use to do this:
http://www.loomia.com/
Posted by: sean |January 18, 2007 5:21 PM
# 16
Alex,
Good post. There is another type of recommendation that I like and that you didn‘t mention, at least not explicitly - People Recommendations.
That is "Here is Joe, I like his ‘stuff‘ (bookmarks, tags, photos....) who else is like Joe around here?" Or, more suitably, "Here I am, who else here is like me?"
Del.icio.us doesn‘t have it, butSimpy has it (and has had it for 1-2 years). Here isan example of people similar to me (look on the right hand side). If you click through some of those people, you will see what we have in common: we‘re all technical, java seems to be a predominant PL, and we are all interested in search, information retrieval, information gathering, etc.
Another interesting people recommendation at Simpy can be found integrated in bookmark search results. For example, here is a search foradaptiveblue. Note the "Related Users" on the right. You are searching for AdaptiveBlue? Well, then these users might be of interest to you.
Thus, Simpy offers these two recommendations, in addition to those offered by del.icio.us:
Person -> People
Search results -> People
Posted by:Otis Gospodnetic |January 19, 2007 8:02 AM
# 17
"This is because there are two fundamental activities online - Search and Browse."
I think the most interesting point is there are *currently* two fundamental activities. Recommendation integration is slowly happening in lots of places and will eventually be integrated into these options or become the 3rd fundamental activity.
Note: working on a podcast recommendation site which may mean I have a biased opinion.http://podcasts.grepr.com
Posted by:Eric C. |January 25, 2007 10:51 AM
# 18
Great article, though I disagree with the definition of "Social Recommendations". I‘d define social recommendations as recommendations and reviews from your friends or a network of trusted associates. At the risk of sounding spammy, check out the site TrustedOpinion. It does exactly that: provides recommendations based on what your friends (and their friends) thought.
It doesn‘t really fall into any of the categories described above.
Posted by:Todd |February 5, 2007 2:26 PM
# 19
A couple of commenters have mentioned the distinction known as item-to-item recommendation versus person-to-person recommendations. Item-to-item is the Amazon-style ‘people who bought X also bought Y‘. Item-to-item recommendations are often based (for popular products) on basket pairings (when I search for books on development methodologies, I often buy two or more of them at a time, so they are paired in my shopping basket and also in Amazon‘s database), which are more accurate than lifetime user pairings (I also do my Christmas shopping on Amazon, and buy pop-up books for my niece). Because of lowered shipping costs on joint orders, Amazon has a phenomenal set of basket pairings. Person-to-person recommendations (a la MovieCritic, circa 2000, and later NetPerceptions) come from a comparison of people‘s tastes (ratings). So each user rates (explicitly or implicitly) a large-ish set of things, like movies. Then recommendations are based on the tastes of one or more ‘recommenders‘ whose tastes overlap the person requesting the recommendation. The problem with person-to-person recommendations is that each individual had to rate a lot of things before a suitable ‘recommender‘ could be found. The upside to person-to-person recommendations is that, once a suitable recommender (or group of recommenders) is found, the recommendations are deep and span broad categories. For example, if I‘m 22 and my preferences for movies from 2000-2006 can be accurately measured, then a 32-year-old recommender whose tastes are also ‘current‘ can be used to find movies from 1990-2000 that appeal to the 22-year-old, but that were released before he would have been aware of them. Person-to-person recommendations are more amenable to time-shifting and medium-shifting. It turned out that an even bigger obstacle to the success of person-to-person recommendations was one of sparse data. To prepare to generate a set of recommendations you have to look at very sparse data from the set of {individuals X products}. For item-to-item recommendations you just have {products X products}. So unless the system has very many recommenders to start with, it never reaches critical mass.
I agree that Amazon knows as much about recommendations as anybody; but the accuracy of recommendations comes from their large dataset, not necessarily from the sophistication of their methods. Note that Amazon does not use a ‘genome-oriented‘ classification to find out why one product is bought alongside another--they could never scale that without an open source effort. They also have not rolled out person-to-person recommendations with any success. For a laugh (and testament to Amazon recommendation accuracy) check out this Onion headline:
http://www.theonion.com/content/node/57311
I think that Netflix is ahead of Blockbuster in many ways: they know that recommendations are the key to unlocking the Long Tail (or, put another way, they are trying to find a way to recommend movies to you that cost less for them to license). But their $1M prize is a very inexpensive way to generate buzz around their recommendation system. I think what they‘re really measuring is not the improvement of the algorithm, but the level of participation in their current recommendations program. As the number of ratings (and recommendation quality evaluations) goes up, so does the quality of their recommendations. Jump-started by a high-profile contest and a (cheap) million-dollar prize. Good move by Netflix.
Posted by: Dave Hunkins |February 9, 2007 8:06 AM
Comments are now closed.