Hello everybody!
Welcome to the Recommenders06.com!
This was the official poster
and this was the official banner:
Open Issues in Recommender Systems
John Riedl, one of the pioneers of recommender systems and founder of Netperceptions, has given a “magistral” class on the open issues of recommender systems. He has remarked at least 3 times the following messages:
- Recommender Systems use data to help people find what they want
- Success come from understanding both data and people
- The social web will be powered by recommender systems
John came representing both the University of Minnesota and Community Lab a consortium formed by Carnegie Mellon, University of Minnesota, and University of Michigan to develop theory to predict contribution behavior in on-line communities.
Professor Riedl went on to emphasize that it is key to keep your community of contributors motivated at all times in order to maintain service quality and that can immensely benefit from ‘Intelligent Task Routing’. This is a research topic that deserved more attention and can benefit from a specialized recommender engine to avoid overwhelming active participants who tend to make up a smaller percentage of the entire user base as in the case of Wikipedia.org. Intelligent routing can also be used to balance the needs of the individual member against that of the entire community. As a motivational example, Amazon.com displays the names of their top product reviewers. Other areas deserving more attention are: Group Agreement and Review Mechanisms.
Finally Professor Riedl made the point that researhers and engineers alike should be alert to external attacks, which can be used to introduce bias to boost a given item or to diss it on purpose. Such efforts are being coined as “Shilling”, which oddly is a practice the casinos used by implanting fake winners on the floor in order to make real guests feel that they too can win if they keep betting.
Dynamic Personalization
Barry’s talk was a very interesting change of perspective on the nature and context of recommendation systems. His work focuses on mobile platforms, where recommendation systems play an essential role in improving usability, rather than filling the standard role of suggesting new content.
Mobile systems have additional challenges above and beyond that of conventional PC platforms. The main limitation of such systems are the lack of screen “real estate” which can limit the navigation options available at any one time. Furthermore, mobile device users are also often (at least) partially engaged in the “real world” while utilizing the device in question. So, the corresponding navigation systems must accomodate the lack of display real estate, as well as the lack of user attention… it’s almost as if the system would need to predict what the user wants to do, and simply “make it happen”.
Barry’s work seems to be a very large step in this direction. He has succesfully leveraged recommendation technology to “short circuit” the standard “click-distance” for common tasks. The click distance represents the sequence of interactions (clicks in this case) necessary to reach a defined state in the program. By providing an online analysis of user behavior, the system is able to predictintends to do, shortening the number of clicks necessary to reach the user’s common task. Since this system is trained on an individual user, and capable of responding to changes in user behavior over time, it is an example of “dynamic personalization”, and a compelling feature for mobile based systems.
Barry went on to show a measured improvement in the average click distance of dynamic personalization based menus versus conventional hierarchical based menus.
Knowledge-based Recommender Technologies
Alexander Felfernig’s talk on knowledge based recommendation technologies was another highlight of the “real world” section of the recommendation talks. Felfernig focuses on a number of application areas (such as financial services) that include “deep domain knowledge”. These areas regularly rely on trained individuals with several years (or decades) of experience. Passing on the requisite knowledge of “what to do in situation x” is rarely a seamless process. A knowledge based recommendation system improves the training, maintenance, and “error handling” capacities of general customer/client management systems. These systems can be employed in sales, financial services, e-government, tourism, and computer support centers, virtually any field where expert advice and decision making is employed on a medium to large scale.
Felfernig also provided supporting evidence of improved customer and user experience in the knowledge based recommendation systems. These results echoed earlier results in this field, underlining the potential for such systems across a broad range of fields.
Anatomy of a Social Recommendation Engine
Jim Shur, Chief Software Architect at MyStrands, has given a presentation explaining how the MyStrands recommender system works. The talk covered some of the practical realities of a real-world recommender system where speed and scalability are absolute requirements and silly problems like not recommending the same song but from a different album have to be dealt with to give a good user experience. Some of the attendees were surprised at how open MyStrands was about the workings of their recommender technology. “Kind of shocked, really”, said one. The first part of the talk went into how they do as much processing as possible upfront to minimize any heavy lifting at runtime. Part II talked about how requests are handled and provide for both personalized and customized recommendations. The talk concluded with some notes on the architecture.
Playlist Evaluation
Very nice presentation by Jim Shur. Yes, he has presented twice today. Jim gave this presentation on behalf of Rick Hangartner for whom circumstances made it impossible to attend. Rick is a dynamic speaker who always has interesting things to say, so hopefully we’ll have an opportunity to hear him present at an event in the near future.
MyStrands uses gazillions of playlists coming from gazillions of users so it makes sense to have an evaluation mechanism in place that filters out bad playlists and/or individual associations. The goal is to avoid bad, or “clinker”, recommendations, whether caused by a malicious attack or just silly or random playlists. One method could be to throw out all associations that have less than some threshold of supporting data, but Jim explained that this is too drastic and throws away lots of perfectly good associations in the process. The idea is to apply various metrics along with a theory of human tastes derived from observable behavior to decide which associations are helpful and which are not. And only add the helpful ones to matrix of associations that underlie the MyStrands recommendation engine.
Overview of the first day and Round Table
After one day I know what a recommender is, how they are differentiated across applications, algorithms and architecture and what their future looks like. Not bad for a day. Maybe I could now kick off and go see the amazing Guggenheim Museum. But I know I am being optimistic. It became evident to me from today´s proceedings that in fact I´ve only scratched the surface of recommenders, not only from the technology perspective but also their social and economic impacts.
The day started off with a thorough overview by Prof. Juntae Kim from S.Korea, who gave an overview of different recommender technology, their methological differences, their benefits and costs and the history of the subject. It served well for framing the rest of the day events. Next came John Riedl, prof from
Minnesota, whose depth and experience in this field became evident through a series of insights into fundamental issues related to recommenders. He used the community-oriented extension of grouplens project (one of the first recommenders for movies) to develop and communicate the underlying point that data should be considered along, not separate, with users.
Next came prof. Barry Smith, who showed us what the future of recommenders would look like where users would be given recommendation on their mobiles. Together with his company Barry is using recommenders to dynamically change the menu options so as to help “deep” content hidden in the bowls of the data warehouse can be found quickly (the 151 code for speed dialing). A somewhat complementary approach was taken by Prof Felfernig, who has successfully applied a generic knowledge-based recommenders to a wide range of domains.
The next batch of presentations were the real world applications of recommenders. Jim Shur opened the hosting company (MyStrands) kimono on the technology and architecture of their recommendation engines. The design tradeoffs they’ve made, why they made them (scalability and speed) were fascinating to listen to. The next kimono opening was Mike Mull of Yahoo! music and radio recommendation engines. On the grand scheme of things MyStrands stood with its kimono wider open than Yahoo! but still the number of requests (1000/sec) Yahoo! has to deal with drove home what the real world issues of implementing these monsters involved. Engineers always have a high place in the order of things my book.
Next came the renowned Basque food. There was a problem to solve though; how to eat as many of those amazing dishes without a) ignoring attendees and b) not appearing glutinous. People talked to me afterwards which I guess meant I did ok.
After lunch was there were 3 student presentations all presenting their ongoing PhD work. The next generation of recommender technology researchers were impressive. PhD student these days seem so much more ahead of the game than my time.
Guess what? The more your idea scales the more likely that you’ll be the focus of blackhat attacks. The final session was about evaluating social recommendations and security. Dr Bamshad Mobasher presented an interesting overview and taxonomy of the types of attacks recommendors are susceptible to. Algorithm designers are well advised to read a few of Bamshad’s papers to learn about the signature of these attacks and the kind of algorithm you need. Jim Shur then presented MyStarnd’s solution to the security problem from a more engineering perspective showing us what we would need to think about when designing.
Last event was the roundtable. The moderator, Peyman Faratin, chose a panel of speakers (John Reidl, Bamshad Mobasher and Barry Smith) and presented a menu of potential topics which the panel would be required to talk about. The menu covered all aspects from theory to technology/engineering to social to economic/business models. The audience were actively involved in helping the moderator select the topics of discussion - the personalization of roundtable if you will. The beauty of it was that the search for a topic of discussion for the roundtable became the discussion engaging not only the audience but also the panelists. Another demonstration that people solve problems through social mechanisms underpinning the centrality of an architecture of participation.
Student Presentations
Today’s program also contained 4 student presentations. Claudio Baccigalupo of IIIA-CSIC talked about the necessity to treat playlist recommendations differently as it contains the added intelligence of song sequence in addition to track selection. Indeed, professional DJs place as much importance on sequencing and transitions as to track selections further justifying a separate treatment. His research conducted on MyStrands track similarity data has delivered a system that is capable of creating a best choice sequence given a seed track. So far the playlist quality measurement has been conducted mostly on an empirical basis requiring additional work to follow up.
The 2nd presentation was by Ralf Jung of Saarland University. His system offered enviromental cues such that it detects individuals entering a given space through the use of RFID tags and is then able to create a unique soundscape based on this intelligence.
Download this presentation!
Watch these two presentations!
Maria Nunez of Universite Montpellier II made the point that the human psycological aspect of recommender systems are neglected yet play a vital role in understanding the actual impact recommender systems create. This is an area that can combine personality traits and mood settings to further personalize and contextualize recommendations.
Finally, Nava Tintarev of University of Aberdeen declared her intent to further her thesis studies on the subject of explaining recommendations putting special emphasis towards human centric issues driving recommendation explanations.
Characteristics of High-Volume Recommender Systems
Mike Mull of Yahoo! gave a talk on the “characteristics of high volume recommender systems” involving millions of users, thousands to millions of items, billions of ratings, millions of play events, hundreds of thousands of requests/day. He focused on the rating data first, underlining two main points; a) how the semantics of ratings were still ill-defined and how the value of ratings fell into an almost binomial distribution.
Yahoo’s recommender system takes anything between hours to day to compute and for this reason they chose an offline strategy, storing affinity data in the DB which is updated weekly. At run time there were diurnal patterns with hits to the system occuring mainly at 12 am and pm for the Launchcast (personalized radio) and Recommenders. The interesting point he mentioned was that the baseline of serving a request is 1 sec and anything above that will be unacceptable. Peak number of requests/sec to Launchcast and YMJ Recommenders were 1200 and 17 respectively. He then finished with some of Yahoo’s system design choices.
What is a recommender system?
Professor Juntae Kim has kicked off the day’s program with a comprehensive backgrouder on recommender systems starting with this Wikipedia definition:
“Recommendation systems are programs which attempt to predict items (movies, books, music etc.) that a user may be interested in, given some information about t he user’s profile.”
Many types of recommender systems exist such as non-personalized, demographic, content based, content based, collaborative (user based), collaborative (item based) and model based. Item based collaborative models have been applied successfully in commercial settings thanks to scalability and quality advantages as compared to others. Model based approaches differ from the rest which rely on memory of events. Instead they involve the creation of a probabilistic, decision tree or neural net model that attempts to identify the underlying logic to users’ choices.
Professor Kim went on to provide examples of former recommendation engine start ups such as FireFly and NetPerceptions as well as commercial usage examples by MyStrands, AOL, Amazon and Samsung Mall. Professor Kim colncluded his presentation by touching on the following open issues and challenges in this field: Sparsity, Scalability, Cold-start, Implicit Ratings, Dealing with Multiple Criteria, Context-dependant Recommendations.
