Steering Committee Workshop Organization

Recommender systems have been extremely successful in reaching relevant content to users. Rather than rely on a static notion of content relevance to a user's query or a profile, they incorporate endorsements of items by other users and/or ratings provided by the same user on other items considered similar. In this talk, I will make a case for developing recommendation strategies and systems not just for recommending content items but for users performing transactions. Consider a social network where users register items (e.g., toaster, lawn mower) they are willing to give away to other users in exchange for items in their wish list which they have registered with the system. The idea is users either swap items or more generally exchange items in cycles. I will discuss the algorithmic challenges in developing strategies for recommending exchange transactions to users and present approximation as well as heuristic algorithms we have developed for solving this problem. I will also discuss the results of a detailed set of experiments we ran to gauge the performance of the various algorithms and conclude with interesting directions for future work. How Far Should We Personalize? (Panel) Sihem Amer-Yahia (Y! Labs, USA) Yannis Ioannidis (University of Athens, Greece), Christian Jensen (Aalborg University, Denmark), Evi Pitoura (University of Ioannina, Greece), Elisa Quintarelli (Politecnico di Milano, Italy) Abstract: Given the proliferation of data and applications, different possibilities as well as different requirements for personalizing user experience emerge at various levels (e.g., content, UI, services, etc). For instance, people may like different services on their cell phone or in Facebook. They may be interested in different content depending on their location, task, preferences or group they belong to. Different presentation features serve different people better. For example, some users like lengthy explanations, others may want to see reviews. Endless personalization possibilities seem to exist in different applications, from personalized search and ads to personal mashups. Given the proliferation of data and applications, different possibilities as well as different requirements for personalizing user experience emerge at various levels (e.g., content, UI, services, etc). For instance, people may like different services on their cell phone or in Facebook. They may be interested in different content depending on their location, task, preferences or group they belong to. Different presentation features serve different people better. For example, some users like lengthy explanations, others may want to see reviews. Endless personalization possibilities seem to exist in different applications, from personalized search and ads to personal mashups. At the same time, there are several arguments against (over-) personalization. For example, if we do not have correct information about a user, personalization may hurt accuracy. In addition to gathering and maintaining a profile, onthe-fly personalization can be expensive. Over-personalization may lead to over-specialization. Making a recommendation of something a user would definitely like is valuable but what about serendipity and diversity? There is also a delicate balance in advertising between the accuracy of ads and their total irrelevance. From the publishers point of view, they would rather serve relevant content than not, from the users' perspective, more relevant ads may be annoying (e.g., a user announces the birth of a child in an email to friends, ads start suggesting where to buy baby stuff). With all these in mind, the panel's objective is to discuss when, what, how, and to what extent we should or should not personalize. CADS: A Collaborative Adaptive Data Sharing Platform Vagelis Hristidis Eduardo Ruiz School of Computing and Information Sciences Florida International University {vagelis,eruiz011}@cis.fiu.edu ABSTRACT Content management tools like Microsoft’s SharePoint allow users of an application domain to share documents and tag them in an ad-hoc way. Similarly, Google Base allows users to define attributes for their objects or choose from predefined templates. This ad-hoc or predefined annotation of the shared data incurs problems like schema explosion or inadequate data annotation, which in turn lead to poor search and analysis capabilities.Content management tools like Microsoft’s SharePoint allow users of an application domain to share documents and tag them in an ad-hoc way. Similarly, Google Base allows users to define attributes for their objects or choose from predefined templates. This ad-hoc or predefined annotation of the shared data incurs problems like schema explosion or inadequate data annotation, which in turn lead to poor search and analysis capabilities. We propose CADS, a Collaborative Adaptive Data Sharing platform, where the information demand of the community–e.g., query workload–is exploited to annotate the data at insertion-time. A key novelty of CADS is that it learns with time the most important data attributes of the application, and uses this knowledge to guide the data insertion and querying. In this position paper, we present the challenges and preliminary design ideas for building a CADS platform. We use the application of CADS on the Business Continuity Information Network (BCIN) of South Florida as a motivating example.