IJCAI 2013 Tutorial on Recommender Systems
August 4, 2013, Bejing, China
Recommender systems help users navigating through large product assortments, making decisions in e-commerce scenarios and overcome information overload. Probably the most prominent example is the book recommendation service of Amazon.com. Their system takes the behavior, opinions and tastes of a large community of users into account and thus constitutes a social or collaborative recommendation approach. In contrast, content-based approaches rely on product features and textual item descriptions; knowledge-based algorithms, finally, generate item recommendations based on explicit knowledge models from the domain. Technically, recommender systems have their roots in different fields such as information retrieval, text classification, machine learning and decision support systems.
The tutorial will offer an introduction to the various basic strategies and methods for building recommender systems and will discuss the different known approaches and how the effectiveness of such systems can be determined.
Collaborative recommendation: Today, systems of this kind are in wide use and have also been extensively studied over the last fifteen years. We will cover the underlying techniques and open questions associated with collaborative filtering. Typical questions that arise in the context of collaborative approaches include the following:
Content-based filtering: Content-based approaches rely on the availability of (manually created or automatically extracted) item descriptions and a user profile that assigns relevance to these characteristics. In the context of content-based filtering, the following questions must be answered:
- How do we find users with similar tastes to the user for whom we need a recommendation?
- How do we measure similarity?
- What should we do with new users, for which a buying history is not yet available?
- How do we deal with new items that nobody has bought yet?
- What if we only have a few ratings that we can exploit?
- What other techniques besides looking for similar users can we use for making a prediction about whether a certain user will like an item or not?
Knowledge-based recommendation: In knowledge-based approaches, the recommender system makes use of additional, often manually provided information both about the current user as well as about the available items. Constraint-based and case-based recommenders are examples of such systems. The questions addressed include the following:
- How can systems automatically acquire and continuously improve user profiles?
- How do we determine which items match or are at least compatible with a user's interests?
- What techniques can be used to automatically extract or learn the item descriptions to reduce manual annotation?
Hybrid approaches: When combining different approaches within one recommender system, the following questions have to be answered and will be covered:
- What kinds of domain knowledge can be represented in a knowledge base?
- What mechanisms can be used to select and rank the items based on the user's characteristics?
- How do we acquire the user profile in domains where no purchase history is available and how can we take the customer's explicit preferences into account?
- Which interaction patterns are used in conversational recommender systems?
- Finally, in which dimensions can we personalize the dialog in order to maximize the precision of the preference elicitation process?
Evaluating recommender systems: Research in recommender systems is strongly driven by the goal of improving the quality of the recommendations. The question that immediately arises is of course how we can actually measure the quality of the proposals made by a recommender system. Thus, the questions addressed will include the following:
- Which techniques can be combined and what are the prerequisites for a given combination?
- Should proposals be calculated for two or more systems sequentially or do other hybridization designs exist?
- How should the results of different techniques be weighted and can they be determined dynamically?
Current topics in recommender systems: Finally, we will introduce active research topics such as application scenarios for Recommender Systems in the context of the Social Web and report on recent case studies.
- Which research designs are applicable for evaluating recommender systems?
- How can recommender systems be evaluated using experiments on historical datasets?
- What metrics are applicable for different evaluation goals?
- What are the limitations of existing evaluation techniques, in particular when it comes to conversational systems or the aspect of the business value of a recommender system?
The presentation will be based on the book "Recommender Systems - An Introduction" that is co-authored by the tutorial presentaers and was published by Cambridge University Press in 2010.
(Amazon.com, Cambridge University Press Online store).
Previous instances of the tutorial have been given at IJCAI 2011, ACM HT, 2011, and SAC 2012
Further information and resources can be found at http://recommenderbook.net .
Dietmar Jannach is a professor in Computer Science at TU Dortmund, Germany, and chair of the e Services Research Group. His main research interests lie in the application of Artificial Intelligence and knowledge-based systems technology to real-world problems in particular in e-business environments. He has authored numerous papers on intelligent sales support systems such as recommender systems or product configurators. Dietmar Jannach was also one of the co-founders of ConfigWorks GmbH, a company focusing on next-generation interactive recommendation and advisory systems. Dietmar Jannach was the program co-chair of the 5th ACM Conference on Recommender Systems held in Chicago, USA, in 2011.
Gerhard Friedrich is a chaired professor at Alpen-Adria Universität Klagenfurt, Austria, where he is head of the Institute of Applied Informatics and directs the Intelligent Systems and Business Informatics research group. From 1993 to 1997 he was the head of the Department for Configuration and Diagnosis Systems at Siemens Austria. Gerhard Friedrich received a PhD and an MS in computer science from Vienna University of Technology, Austria and was a guest researcher at the Stanford Research Institute and at Siemens Corporate Research. His research interests include knowledge acquisition, constraint satisfaction, configuration, planning and diagnosis. He is an editor of AI Communications, an associate editor of the International Journal of Mass Customisation a fellow of the European Coordinating Committee for Artificial Intelligence.
Dietmar Jannach, TU Dortmund (firstname.lastname@example.org)
Technische Universität Dortmund
Lehrstuhl Informatik 13
Prof. Dr. Dietmar Jannach
Phone: +49 (0)231 755-7272
Office hours: By appointment
E-mail: dietmar.jannach [ at ] tu-dortmund.de