Recommendation Engines

    Recommendation engines are systems that recommend content to you. Your friends are recommendation engines, as are the "top 100" lists of books and movies available at your local library and video store. In the context of personalization, the best recommendation engines are not the ones that offer you the authoritative top items in a collection but the ones that offer you things that you might actually like. Anyone can read Rolling Stone's top 500 rock albums of all time. But which ones will be right for YOU? That's what recommendation engines try to figure out.

    Recommendation engines use a combination of explicit and implicit information to provide people with suggestions. In some cases, users make explicit designations by making purchases, expressing preference via ratings or reviews, or choosing some things over others. But on the Web, users also continually generate data passively via the things they click on, items they spend a long time looking at or listening to, and the choices their friends are making.

    In the physical space of a museum, visitors make very few explicit data contributions. One might buy a ticket to a special exhibition or show or actively elect to take up an audio guide or exhibition brochure. Most of your preferences for one museum experience over another go unregistered and untracked. This means there's very little data on which museums can automatically offer recommendations for further experiences.

    Consider the case of Netflix, the dominant US online movie rental company. Netflix offers subscribers a steady stream of movies delivered directly to users' doorsteps (or computers). Netflix is in the business of selling subscriptions to its service. They do not want you to cancel your subscription because you've seen all the movies you want to see or can't find a movie of interest. They don't want to leave it to chance that your friends and family will be continually suggesting movies that you might like, or that you will be studiously scanning the AFI top 100 films list for ones you haven't seen. And so Netflix spends a lot of money and energy improving its recommendation system so it can keep suggesting movies that you might like to see. In October 2006, Netflix offered a million dollar prize to the team who could improve their recommendation system by ten percent. Netflix knows that good recommendations are key to their bottom line.  If Netflix suggests too many movies that you don't like, you will either start ignoring the recommendation system or cancel your subscription altogether.

    The Netflix recommendation engine is based on ratings. While you may think of rating as a frivolous activity, in many ways they are the ideal participatory data points for recommendations. Unlike comments or purchases, ratings are easy to generate and are appealing to the large percentage of people who fall into the “critic” and “collector” categories of social media use. Users can rate things whether they have experienced them or not, purchased them or not.

    Here's how the Netflix recommendation engine works. At any time, you can rate movies in the Netflix database from one to five stars. You can rate movies you haven't watched or rented. In fact, Netflix makes a game out of rating movies, encouraging you to do so upon initial account registration and on subsequent logins as part of the profile-building experience so the system can supply you with lists of "Movies You'll Love." They make rating easy and fun. They also ask you to rate broad genres (Action/Adventure, Comedy, Foreign, etc.) so they can provide you suggestions in particular categories. Finally, they allow you to indicate if you are not interested in a suggested title (a kind of reverse rating for things you haven't seen). The underlying message here is that Netflix can always help you find a movie you’ll love, so you should never stop subscribing.

    This implicit promise is also the key to why people willingly rate hundreds of movies on Netflix. Netflix promises to give you better recommendations if you rate more movies. Your user profile is functionally an aggregate of the movies you have rated, and the more finely tuned the profile, the more useful the recommendations. The more you use it, the better it gets--and that symbiotic relationship serves customer and vendor alike. This promise is what is missing from so many museum rating systems. When museums allow visitors to rate objects or express preferences, the visitors' expressions are rarely, if ever, fed back into a system that improves the museum experience. The presumption on the part of museums is that ratings systems are "fun" and that's why people use them on Netflix and other sites. But they aren't just fun ways to express yourself. They have direct personal impact. Whether you are panning a movie or gushing over a book, most online retailers track your explicit actions and use them to provide you with better subsequent experiences.

    But what's "better" in the museum context?  One of the biggest concerns about deploying recommendation systems in museums is that visitors will only be exposed to the narrow window of things they like and will not have "off path" experiences that are surprising, uncomfortable, and valuable. Fortunately, not all of us are in the business of selling movie rental subscriptions (or woks, or books, or whatever). While online retail recommendation engines are unsurprisingly optimized to present you with things you will like, there are other ways to filter information based on preference. For example, Librarything has a "books you'll hate" feature called the Unsuggester. When the BookSuggester was released in November of 2006, programmer Tim Spaulding wrote a blog post about the addition of the Unsuggester. While the Unsuggester was offered as a silly toy, Spaulding noted a change in his perception of what it means to be “well-read” based on its results. After noting the patterns of opposition between philosophy and chick lit, programming manuals and literature, Spaulding wrote: "These disconnects sadden me. Of course readers have tastes, and nearly everyone has books they'd never read. But, as serious readers, books make our world. A shared book is a sort of shared space between two people. As far as I'm concerned, the more of these the better. So, in the spirit of unity and understanding, why not enter your favorite book, then read its opposite?"

    The Unsuggester is based on different values than Netflix's Movies You'll Love and the BookSuggester. The value system for the Unsuggester is based on the idea that we can learn something from things that are foreign to our experience. The Unsuggester doesn't so much give you books you'll hate as books that you'd never otherwise encounter. The books on the list are the ones that are least likely to be found in your Librarything collection or the collections of other users who also have your books. In other words, the Librarything profile is based not on ratings but on books. And while the Unsuggester is silly, it's also a valuable set of responsive content to your profile. It's a window into a distant and somewhat unknowable world. And users have responded positively. When Spaulding suggested that few people were likely to actually read books on the Unsuggester list, an anonymous user responded, "You underestimate Thingamabrarians. Some of us are just looking for new ways to branch out from our old ruts... and something flagged as 'opposite' to our normal reading might just be what we're all looking for. (Besides, a lot of the 'niche' books are throwing up classics in the unowned lists, and many people like to improve their lit-cred.)"

    In other words, recommendation systems don't have to be optimized to give users something they’ll like. They just have to be responsive to users’ personal profiles in some understandable and meaningful way.

    Pandora is an example of a recommendation system with a different value set and strategy behind it, one that is closer to the traditional curatorial expert model than Netflix or Librarything. Pandora is an online music service that provides a personalized radio station based on a combination of user inputs and expert analysis. How is a museum like a radio station? Both are collections of discreet, loosely organized content pieces that are both familiar and new. Your overall enjoyment of the content experience is determined to a large extent by the balance of items you like and those you don’t, those you know and those that are new. The more time between the good stuff, the less likely you are to tune in again in the future. And your loyalty to the radio station (its stickiness) relies on the regular introduction of unfamiliar content in an enjoyable context.These criteria aren’t easy to meet, and the result is lots of people like me who never listen to non-talk radio. But Pandora has lured me back by successfully responding to my profile (a small set of personally-selected musical inputs) with a wealth of new musical experiences.

    Pandora uses collaborative filtering to create a real-time radio station for you based on your preferences. You enter a seed artist or song (or several) and Pandora starts playing music that it interprets as related in some way to your selections. This seed content is functionally your user profile. The extraordinary thing about Pandora is the complexity of its filtering. It doesn’t just group artists together and play music by similar musicians. Instead, it uses hundreds of tags, signifiers assigned to each song by a team of expert musicians, to find correlated songs that may be of interest. Pandora is a product of the Music Genome Project, in which musicians define the individual “genes” of a song via signifiers and use those to generate song “vectors” that can then be compared to create highly specific and complex musical narratives. Each song takes twenty to thirty minutes for experts to encode. This is a serious data project, not unlike the kinds of categorization and research projects performed on museum collections.

    For example, I created a radio station based on just one song: Diamonds on the Soles of Her Shoes by Paul Simon. That radio station then played:

    •          She’s a Yellow Reflector by Justin Roberts

    •          If Only the Moon Were Up by Field Music

    •          She’s Going by The English Beat

    •          You’re The One by Paul Simon

    •          Withered Hope by They Might Be Giants

    •          Big Dipper by Elton John

    •          Wait Until Tomorrow by New York Rock and Roll Ensemble

    •          The Tide is High by Blondie

    All but one of these songs and half the artists were new to me.  And I enjoyed seven out of nine. For each song, I could click a “Why?” button to see Pandora’s explanation for why it was played. For example, The Tide is High was included because it "features acoustic rock instrumentation, reggae influences, a subtle use of vocal harmony, repetitive melodic phrasing and excessive vamping."  There are over 400 different tags used to relate songs in the Music Genome Project, ranging from “brisk swing feel” to “lyrics that tell a story” to “sparse tenor sax solo.” From a single seed song, Pandora will generate a whole channel of music, and will shift and refine that channel based on your thumbs up/down rating of each song played. In this way, Pandora makes inferences about what you might like and introduces you to new music.

    It’s the introduction to new music that makes Pandora uniquely interesting as a recommendation system. Rather than user-based collaborative filtering, in which visitors receive recommendations based on what other “people like you” enjoyed, Pandora is an example of item-based collaborative filtering, in which visitors receive recommendations based on the similarity of previously selected items (seed songs) to potential members of the collection.

    Pandora and the Music Genome Project are controlled by experts, musicians who, like curators, are uniquely skilled at identifying and tagging songs to create musical genes that represent the full spectrum of musical expression. And their expertise makes for a better experience for me as a user/visitor. As an amateur listener, I could not tell you the particular elements of “Diamonds on the Soles of Her Shoes” that appeal to me. Listening and reacting to the Pandora-generated songs allowed me to understand the nuance of what I like and don’t like. Turns out I enjoy songs with “extensive vamping.” Could I have articulated that at the start? No. Not only does Pandora introduce me to new music, it expands my vocabulary for discussing music. I learned something! From experts!

    Users of Pandora are protective of the Music Genome Project experts. There have been discussions on the Pandora blog about the slow inclusion of user-based filtering, and listeners' related fears that it will taint the waters of the high-quality item-based process. The Music Genome Project involves visitors' ratings in a limited way. The core value is in the professional categorization of the songs.

    This brings us back to our fictitious museum and the multiple perspective labels. Remember, we've asked several staff members to write their own takes on exhibits exhaustively throughout the museum, and now we need a way to intelligently serve the "right" content to visitors. How will we design our recommendation engine? We could ask staff to map the "genetics" of every exhibit and every perspective, generating experiences that are both highly responsive to individual visitors’ preferences and which deepens visitors’ understanding and ability to articulate why they like what they like. In some cases, people might be surprised to learn that they prefer artists whose subject matter comes from childhood memories, or those who work in a specific medium. While the museum can’t be physically rearranged for each visitor or family, the content could be remixed conceptually to present a progressively engrossing, educational experience.

    Personalization doesn’t just give you what you want. It exposes you to new things, and it gives you a vocabulary for articulating and refining why you like what you like. Pandora’s collaborative filtering process contextualizes data from a very personal starting point. You get the analysis and the narrative, but you get the slice that will resonate most with you. The world opens a little wider and hopefully, you keep listening.

Continue to the next section, or return to the outline.