This project addresses the practical need to develop more advanced personalization and recommendation technologies. Although recommender systems represent the most researched and developed personalization technologies applicable in a variety of applications, most current-generation recommender systems focus on recommending items to users and represent user preferences for an item with a single rating, and therefore are not sufficient to capture the intricacies of some of the more complex settings. This project develops several enhancements necessary for the next generation of recommender systems, such as context awareness, multi-criteria ratings, rating aggregation, flexibility of recommendations, and non-intrusiveness. In particular, the proposed approach explores the synergies between the recommendation process and the multidimensional/OLAP data model and extends the traditional recommendation framework to incorporate the advanced capabilities in a systematic manner. Overall, this research project will make contributions to both theory and practice by developing new frameworks, models, algorithms, and implementations that provide effective ways to deal with information overload and promote access to relevant information. Technologies resulting from this research can bring a broad range of benefits in many areas, including business and electronic commerce, social settings, and education. In addition, research results will be incorporated in the undergraduate and graduate courses on Business Intelligence and Information Technologies at the University of Minnesota. The results will also be made available to scientific community through publications in refereed journals and conferences. In addition, the project website (http://ids.csom.umn.edu/faculty/gedas/NSFcareer/) will be used to disseminate the publications, datasets, software, and course materials that result from this project.
This project addresses the practical need to develop more advanced personalization and recommendation technologies. Recommender systems represent the most researched and developed personalization technologies applicable in a variety of settings. While the current generation of recommender systems has been successfully used in several applications (as exemplified by companies like Amazon and Netflix), most current systems address the problem of providing recommendations of a single type (i.e., recommending relevant items to individual users), and represent user preferences for an item with just a single rating, and therefore are not sufficient to capture the intricacies of some of the more complex application domains. In addition, much of recommender systems research focuses purely on algorithmic/computational aspects of recommender systems, while behavioral and economic implications of recommender systems have not been extensively explored. This project explores a variety of issues and limitations related to the current generation of recommender systems; and it proposes a number of enhancements necessary for the next generation of recommender systems. In particular, the major outcomes of this project include: (a) studying the impact of recommender systems on individual users and demonstrating that user preferences and economic behavior can be significantly manipulated by recommendations, which has significant implications for the design, application, and evaluation of recommender systems as well as for e-commerce practices; (b) developing the general framework for context-aware recommender systems and introducing three different types of techniques for incorporating contextual information into the recommendation process, which allows to model more complex recommendation applications; (c) developing a recommendation query language that allows the users to express complex recommendations in a concise and intuitive manner, thus adding substantially more flexibility and expressive power into the recommendation process; (d) going beyond the traditional accuracy-based evaluation of recommender systems – developing several important additional measures of recommendation quality, such as recommendation diversity and recommendation stability, which allows to have more nuanced and more practical definitions of recommendation quality; (e) demonstrating the limitations of some of the state-of-the-art recommendation algorithms with respect to the diversity and stability measures, and then designing several efficient computational techniques that allow to optimize recommendation performance according to these measures; (f) demonstrating the benefits of using richer (i.e., multi-criteria) user preference data in recommender systems and developing new algorithms that are able to leverage the multi-criteria information to provide better recommendations; (g) providing a comprehensive analysis of the impact that various data characteristics may have on the performance of different recommendation algorithms; and (h) developing data analysis and visualization techniques for identifying and representing trends in complex multidimensional data. It is important to note that many of the proposed techniques and approaches do not depend on any specific application domain. In other words, the solutions resulting from this research could be applied to a wide variety of personalization applications and can bring a broad range of benefits in many areas, including the areas of business and electronic commerce (personalization services, personalized products, personalized/dynamic prices, product recommendations), society (personalized interactions, personalized emails and Web sites, personalized information searches, personalized TV and radio programming), education (personalized learning, personalized archives and digital libraries), etc. In summary, the research contributes not only to science but also to practice by developing new algorithms, models, frameworks, measures, and implementations that provide more effective ways to deal with information overload and to promote access to relevant information.
