We are very pleased to announce ACM Recommender Systems 2007.

This Conference builds on a wonderful legacy of research workshops and the Recommenders06 Summer School on the Present and Future of Recommender Systems, organized by MyStrands last year in Bilbao.

RecSys07 is organized by Prof. Joseph A. Konstan (General Chair). Prof. John Riedl and Prof. Barry Smyth are the Research Program co-Chairs and Dr. Francisco J Martin is the Practice/Industry Program Chair.

The Conference will bring together the leaders of this field–from both research and practice–to explore the latest innovations, discuss important challenges, and advance our understanding of recommender systems.

Please visit the RecSys07 website for further information.

The New York Times is announcing that Netflix plans to award $1 million to the first person who can improve the accuracy of movie recommendations by 10%. Netflix has been one of the companies leading research on recommender systems, and proving innovative excellence with its recommender system. But this seems not to be enough for Netflix, and they are now defying the research community with this prize announcement. And they are making 100 million movie ratings available to researchers to assist them in their quest.

What are your views?

1. Do you think it is possible to improve the accuracy of Netflix recommendations by 10%?

2. What are the key ingredients a new algorithm should include to improve Netflix’s current algorithm?

For those of you who missed Jim Bennet’s (VP Recommendation, Netflix) presentation in Recommenders06, here you will find the video and slides.

Day 2 came to a close with another roudtable discussion. It was moderated by Atakan Cetinsoy of MyStrands and the panelists included Jim Bennett of Netflix, Kaushal Kurapati of Ask.com and Paul Lamere of Sun Microsystems. Atakan started off the debate by contrasting search vs. recommendations posing the challenge that recommender systems must achieve a level of user adoption that can rival that of search. Search has quickly become a dominant use case deemed irreplacable whereas some users can classify recommendations as a “nice to have”. Atakan then went on to ask what it will take to overcome the barriers to that end.

Jim responded stating that he is in agreement that there needs to be more done in order to make recommender systems a more integral part of the user experience and he mentioned good old usability techniques as a key determinant of eventual success as it related to mass user adoption. Paul added to that the lack of interoperability of user profiles and preferences as a barrier but the panelists did not have a clear conviction on whether this interoperability challenge could be overcome any time soon.

As a second question panelists were asked their opinion on the sometimes conflicting dual role of recommender systems in trying to encourage content exploration (which is similar to browsing) while acting as a search type of function reducing the time and effort it takes for users to reach relevant content. Kaushal mentioned that search data clearly shows users have developed a bit of a solit personality when it comes to intent within a given session and it is not clear that they always have a good idea of where to start. Depending on the need and how well or ill-defined it is he added there may be a place for both roles provided expectations are set properly early on.

Panelists were also asked what impact internationalization would have on recommender systems thanks to rapid globalization. Jim mentioned that early on Netflix hired a lot of Indian engineers and their taste data resulted in a lot of Bollywood movies being recommended against Hollywood blockbusters the idea being we can not ignore international tastes and preferences when building such systems. It is a bit of an art in drawing the line and deciding to implement an entire new instance of a recommender system and that was the case when Netflix was planning a UK launch as well.

Finally, panelists were queried on how they see this discipline developing in the next 2 years and there was consensus that we need more commercial success stories and robust hybrid recommender system implementations that combine the best technology with great ease of use. So it is an execution challenge in that sense instead of a theoretical one.
Kudos to all participants and the organizers for a great conference!

Watch the round table discussion!

In this session, Dr Francisco J Martin illustrates why group recommendation is an interesting topic and illustrates the challenges that a real-world group recommendation system brings.

For 2 days, we’ve been seeing recommender systems that help us personalize our experience when trying to find things that we want. A recommender system is capable of suggesting content, products or services based on an individual’s preferences.

Lets not forget an important fact: humans are social by nature and there are many real-life situations where we, a group of people, want to find things that the whole group likes. Francisco J Martin addressed the topic and illustrated his point by explaining the process by which a group of people find something that all of them will like. He gave some real world, everyday examples to help explain how and when a set of preferences come into play  and illustrated that finding things a group might like is a little bit more complex. Consider making paella for a large group, there will be some who prefer fish, others meat and others still chicken. A simple solution would be to prepare a separate dish for each person, but this solution does not scale especially if you are making paella for 100 people given everyone’s unique preferences.

There are plenty of other relevant examples, ranging from sharing the TV in the family room, the tunes piped in at the gym while you work out to the music at your favorite store while you shop, who gets to choose the music for the group? We have seen examples for group recommendations at a gym as highlighted in John Riedl’s address yesterday. A use case that we are all familiar with involves bars and restaurants … frequently they play music that most people don’t like ((not sure where you get that fact from)) why don’t they adapt the playstream to reflect people’s tastes on the spot?

Francisco categorized how content has been targeted so far. Programming CNN, the New York Times, and the like is relatively easy: only 1 version has to be created, just the same for everybody. There’s no personalization at all. If we have a look at Youtube or Google we can retrieve content based on our own search criteria but without personalized results. We have seen this morning that Ask.com is moving toward personalization. Google too. They are starting to provide personalized results based on an individual’s previous search history. Pandora and MyStrands are examples of companies out there focusing on providing personalized experiences but nobody out there is focusing on providing group recommendations. partyStrands is MyStrands new social programming tool.

Francisco explained how partyStrands selects music for a group based on each individual’s preferences. partyStrands attempts to maximize the satisfaction of everybody over time thus uses a fair algorithm that generates a sequence of diverse songs with smooth transitions that maximizes the overall satisfaction of all the partygoers preferences. Fairness is understood in terms of equally rewarding each partygoer. Smoothness is understood in terms of artist similarity. The sequence does not transit from Iron Maiden to Luciano Pavarotti. The sequence is diverse which means that the same artists don’t repeat over for a long period of time.

partyStrands first selects a set of initial seeds (based on previously listened to tracks, favorites, tops, etc). Then it computes the artist distribution and extends it using the MyStrands recommender. Then partyStrands filters the extended distribution using the previous sequence of generated tracks to avoid repetitions. partyStrands then selects the most relevant artists (e.g., those with more votes). For each selected artist, partyStrands selects a number of relevant tracks based on a distance computed in terms of their similarity with the last track played, BPM, and popularity. partyStrands then randomly picks a track and adds it to the sequence. Later what partyStrands does is decay the influence of each user who supposedly is satisfied with the selection. partyStrands algorithm brings similar issues to those that appear when solving a social choice problem. You can cater to the most satisfied individuals, resulting in a small but very happy group of people, or you can cater to the least satisfied individuals, resulting in a handful of somewhat discontent people, or you can cater to the average satisfied individual, resulting in a large group of moderately happy people.  Dr Francisco J Martin mentioned that so far MyStrands is running the first trials of partyStrands but that they don’t yet have a mechanism in place to formally evaluate the satisfaction.

Finally, Francisco remarked that group recommendations are a very interesting topic. There are only few references in the literature and would be a very interesting topic for one of the Ph.D. students here.

Download this presentation!
Watch this presentation!

Marcus Stolze talked today about recommendation theory in the context of a “needs oriented system”. Needs oriented systems are characterized by providing recommendations based on “human” needs. For example, the person may want a a camera of a certain size or capable of fitting a certain lifestlye (perhaps it’s good for sports photography). The user may or may not know which feature set matches their needs, and may not be able to articulate their needs in terms of features.

Feature oriented recommender systems often focus on discrete features of an item. Many electronic gadget sites such as Best Buy are examples of feature oriented systems. These sites focus on describing every last technological feature of an item, and use these features to allow the user to navigate through the available inventory catalog.

Markus argues that we need to support a focused understanding of customer needs. He showed an example of how camera’s useable features could be used to “score” need based attributes. For instance, a camera’s flash quality, size, cost, and low light handling would have some impact on its daily use value.

He outlined some reasons why recommendation systems fail (inefficiency, lack of usefullness) or otherwise get discontinued (lack of funding). He claims that the knowledge engineering cost can be greatly reduced through a needs based focus, and underlined the importance for an authoritative history of the field.

Download this presentation!
Watch this presentation!

Prof Anton Nijholt’s research goal is to determine the current goals of the user: goal-based structuring.  Goals are important and make a difference.  If you can use a recognizable structure in the recommendations then you’re better off.  To learn what the user’s goal is, we can ask the user, predict it, or let users navigate, adjust, and refine their goals as desired.  Experiments were run to test a number of hypotheses.  One hypothesis claimed that using prediction can be more helpful in finding items than not using prediction.  Predicted items were interesting.  The other interesting hypothesis claimed that structuring the items based on a user’s goals makes it easier to find than without structure.  The Unified Theory of Acceptance and Use of Technology (UTAUT) questionnaire methodology was used to measure how easy it is for users to find interesting items.  The questions in the ATAUT covered performance, effort expectancy, and intent. 
   
 Anton then went on to the case where a user’s goals can be implicitly derived when they are immersed in computation rich environments.  Called ambient intelligence (ubiquitous computing + social intelligence interfaces), this scenario reflects computing everywhere; rooms with sensors, technologies that help us to identify a user’s psychological states, etc. that ultimately determine their goal.

Download this presentation!
Watch this presentation!

Now we get to our second session where promising students explain their current research topics. Definitely a highlight on day one, we’ve been looking forward to more.

First up, Justin Donaldson from Indiana University’s School of Informatics, USA, talking about Recommendation Mapping, with the subtitle, “Recommendation as ‘Visual Search’: Peeking Under the Hood of Recommendation Processes”. To help people discover new music, Justin proposes a change in interaction style by exposing the user to a visualization of the underlying network of associations, or as he puts it, move From “Oracle and Ordinal” to “Spatial and Relational”. After explaining some of the theory behind the work, Justin showed a nice demo that, apart from visualizing song affinities, showed off some nice UI techniques for exlporing tightly clustered data points. You can try it yourself by going to MyStrands Labs.

Download this presentation!
Watch this presentation!

Next came Holger Grossmann, Fraunhofer IDMT, Germany, speaking about “A pure content-based approach to music similarity using aspect profiles.” Holger began with a demonstration showing to what extent music retrieval techniques can be pursued today. Impressively, he instantly extracted the drum parts from an audio file of “I Will Survive”. Similarly, he extracted the bass parts and the main melody line. Playing all the extracted parts together was a real crowd pleaser! Holger then walked us through the supporting facts and arguments for his contention that pure content-based music similarity guessing is an appropritate means for browsing unknown archives. Hence, such techniques could strenghthen the position of independent content provideres.
Watch this presentation!

Third, we were very happy to have Mori Kurokawa, all the way from Keio University in Japan. Mori presented an Experimental Study of Probabilistic Models for Movie Recommendations. Using rating data obtained from a survey of 1600 people, Mori experimented with different Bayesian network models and presented his analysis and conclusions about the effectiveness of each.

Download this presentation!
Watch this presentation!

Download this presentation!
Watch this presentation!

Lastly, Jolie Martin, from Harvard University, with her talk “Ratings Variance in Recommender Systems”. For ratings varience, Jolie is talking about distribution in the relevant ratings. The higher the variance, the greater the uncertainty. So given the same average rating, would people prefer an item with higher or lower ratings variance? Put another way, would they want to choose sometime they might really love or really hate, or instead choose something that’s highly likely they’ll find mediocre? It turns out in many, though not in all cases, subjects somehwat surprisingly have a tendency to choose options with a higher variance. This goes against the usual economic analyses. In what she termed the more “hedonic” cases, like chocolate, people seemed to prefer higher varience, going for the long shot. In cases where there is a perceived high cost of a poor outcome, people were more likely to play it safe and prefer low variance.

Paul Lamere of Sun Microsystems gave a talk about content-based music
recommendation. First Paul put the problem into context: Music
collections are large and are growing rapidly.

According to Chris Anderson in the ‘Long Tail’ recommendation is going
to be the key to dealing with this ever increasing amount of music.

Paul explained that the tool most often used now for music
recommendation is the social recommender. The social recommender
gives good recommendations, but has a number of issues that limit or
reduce its effectiveness. Issues such as popularity bias, sparseness
of data, lack of transparency in recommendations, social inertia,
lack of symmetry, the cold-start problem all lead to less than
perfect recommendations.

Paul then went on to discuss how content based music recommendation
can deal with these problems. Content-based systems are immune to
popularity bias, feedback loops and many of the other problems
inherent in social recommenders. Paul continued by describing a
popular content-based recommender called Pandora that uses trained
musicians to build a model that is used to recommend music based on
content. Paul pointed out that the Pandora system does not scale very
well, with an estimated cost of about $10 to analyze each song. As
the amount of music generated every day continues to grow, Pandora
will find it more and more difficult to keep up. Paul suggested that
the solution was to automate the process, to use machines instead of
humans to analyze the music.

Paul pointed to the Music Information Retrieval field where automated
genre classification systems can outperform humans in classifying the
genre of music. Paul then went on to explain how a genre classification
system works and how a genre classifier can be extended to generate a
measure of music similarity. He indicated that the real research right now
was trying to figure out what features to extract from the music, and what
machine learning algorithms to use to build the music similarity models.

Finally, Paul gave a demo of the Search Inside the Music system. This
system used a music similarity system to generate music by finding the
music that ’sounds’ most similar to a seed song. Paul demonstrated
this with songs from a number of different genres, including punk,
classical and jazz. Finally, Paul showed a 3-dimensional visualization
of the music space. He was able to use the visualization to explore his
music collection, to get an indication about which songs were similar and
which songs were different. He was able to make interesting playlists
by selecting the first and the last song in the playlist and having
the system fill in the songs that will reduce the ‘iPod whiplash’.
Finally, Paul showed how the 3D visualization could be used to help
people interact with their album art, giving them new tools to explore
their music collections. The 3D visualization of the music space was
a crowd pleaser.

Download this presentation!
Watch this presentation!

John McPherson from MyStrands presented an example Flash radio program built on top of OpenStrands, the open API that allows access to a variety of MyStrands services, including recommendations. The app shows the currently playing song, as well as the previous and next songs. The app uses recommendations to populate the continuous playstream.

The problem he stated was: “Given the last few tracks that were played, what song should we give to you next?”

The answer is to incorporate explicit feed back (thumbs up, thumbs down), and implicit feedback (skipped tracks, e.g.) to formulate filters to use along with the previous tracks as seeds to feed the recommender.

John showed a nice example of starting the player off with a blank slate user profile. The radio played gernerically popular songs, most of which John skipped, but the ones he liked he rewarded with a thumbs up. Soon the radio started converging to his taste and playing almost all songs he liked.

For the geeks in the crowd, John went on to show some slides showing off some of the programming details of the OpenStrands API. Some HTTP calls for the REST api with the XML that gets returned, plus a little java code.

Watch this presentation!

Yahoo!’s Todd Beaupre presented music personalization within Yahoo!. Y! Music is striving to replace broadcast radio and TV in terms of content discovery. Y! sees personalization as a key pillar (alongside search and community) in its future and recommendations fit in that realm. Y! Music is the leading music site in terms of unique and total minutes of usage per month. Launchcast service heavily relies on billions of user ratings in creating a preferences profile. Although users can at times be confused by the meaning of rating levels creating inconsistencies. Originally users would rate music to be able to rehear it on Launchcast. Furthermore, ratings are pretty skewed with few people rating many items and many users not rating at all or rating few items.

From a user’s perspective convenience, quality, control and discovery are the key drivers for Yahoo! The inclusion of Y! Music in the IM client was a big success and resulted in a lift not easily achievable through algorithm modifications alone. So concentrate on the user experience instead of technicalities of the recommender system by email, IM and SMS.

Y! Music also attempts to create trust by providing recommendations explanations as appropriate. Music alerts are delivered and  personalized enjoying 70% open rates and 30% click thru. Going forward they will focus on further integration between systems to achieve mass adoption as well as embracing social networks further.

Download this presentation!
Watch this presentation!