Preference Learning (PL–10) Proposal for a Joint Tutorial and Workshop at ECML/PKDD–2010 Johannes F¨ urnkranz and Eyke H¨ ullermeier

1

Prelude

The joint tutorial and workshop “Preference Learning” (PL–10) is planned as a follow-up event of the workshops on the same topic that have been held at ECML/PKDD–2008 and ECML/PKDD–2009 [23, 24]. Both workshops have been quite successful in several respects, in particular in terms of the number of attendees and the quality and quantity of submitted papers. An edited volume based on the presentations and discussions at the workshop is currently in press [11]. The participants of last year’s workshop expressed considerable interest in continuing this workshop series. We would not only like to satisfy this wish, but also use the opportunity to popularize the research in this area by complementing the workshop with a half-day tutorial on this topic.

2

Workshop topic and research issues

The topic of preferences has recently attracted considerable attention in Artificial Intelligence (AI) research, notably in fields such as agents, nonmonotonic reasoning, constraint satisfaction, planning, and qualitative decision theory [13, 30, 6]. Preferences provide a means for specifying desires in a declarative way, which is a point of critical importance for AI. Drawing on past research on knowledge representation and reasoning, AI offers qualitative and symbolic methods for treating preferences that can reasonably complement hitherto existing approaches from other fields, such as decision theory and economic utility theory. Needless to say, however, the acquisition of preferences is not always an easy task. Therefore, not only are modeling languages and representation formalisms needed, but also methods for the automatic learning, discovery, and adaptation of preferences. 1

It is hence hardly surprising that methods for learning and predicting preference models from explicit or implicit preference information and feedback are among the very recent research trends in disciplines such as machine learning and knowledge discovery [9]. Approaches relevant to this area range from learning special types of preference models such as lexicographic orders [5, 29, 7], over collaborative filtering techniques for recommender systems [2, 14] and ranking techniques for information retrieval [27, 28, 26], to generalizations of classification problems such as label ranking [15, 8, 4]. Like other types of complex learning tasks that have recently entered the stage in the field of machine learning, preference learning deviates strongly from the standard machine learning problems of classification and regression. It is particularly challenging as it involves the prediction of complex structures, such as weak or partial order relations, rather than single values [1]. Moreover, training input will not, as it is usually the case, be offered in the form of complete examples but may comprise more general types of information, such as relative preferences or different kinds of indirect feedback [25, 26]. Topics of interest to the workshop include, but not limited to • quantitative and qualitative approaches to modeling preferences and different forms of feedback and training data; • learning utility functions and related regression problems; • preference mining, preference elicitation, and active learning; • learning relational preference models; • generalizations or special forms of classification problems, such as label ranking, ordinal classification, and hierarchical classification; • comparison of different preference learning paradigms (e.g., “big bang” approaches that use a single model vs. modular approaches that decompose the learning of preference models into subproblems); • ranking problems, such as learning to rank objects or to aggregate rankings; • methods for special application fields, such as web search, information retrieval, electronic commerce, games, personalization, or recommender systems.

3

Outline of Tutorial

In the tutorial, we will try to present the field in all its aspects. We will clarify several different types of preference learning problems, discuss algorithms for solving them, point out their commonalities and differences, and present a few selected applications. 2

1. Introduction 2. Preference Learning Tasks • Object Ranking • Instance Ranking • Label Ranking 3. Loss Functions for Ranking and Preference Learning • ranking errors (Spearman, Kendall’s tau, . . . ) • multipartite ranking measures (AUC, C-index, . . . ) • information retrieval measures (precision@k, NCDG, . . . ) 4. Preference Learning Techniques • learning utility functions • learning preference relations • model-based preference learning • local aggregation of preferences 5. Applications • information retrieval • recommender systems We attach a survey paper [12], and sample slides of a 45-min presentation, which only touches on a few of the above subjects. This set of slides will be considerably extended for the tutorial.

4

Motivation for a workshop and duration

As mentioned earlier, preferences in general and preference learning in particular are emerging research topics of high interest and potential. For example, a very successful workshop series has been started four years ago under the name “Multidisciplinary Workshop on Advances in Preference Handling”. It has taken place at IJCAI–05, ECAI–06, VLDB–07, and AAAI-08, and clearly documents the great interest in the topic of preferences. Besides, it proves the interdisciplinary character of this topic, as it regularly attracts participants from different fields, such as AI, databases, operations research,

3

and decision theory. In the same spirit, a special issue of the Artificial Intelligence Journal, co-edited by one of the proposers (Eyke H¨ ullermeier), has 1 been launched. It received an extremely positive feedback within the AI community, and more than 60 people expressed their interest to submit a paper. While the scope of the aforementioned workshop and special issue is relatively broad, our workshop is especially focused on issues of machine learning and data mining, i.e., on problems of preference elicitation and learning preferences in an automatic, data-driven manner. Apart from our previous workshops (PL–08, PL–09), there are a small number of related workshops in the machine learning field, such as “Learning in Structured Output Spaces” at ICML–06 or “Learning to Rank for Information Retrieval”, at ACM SIGIR–07. These workshops are even more specialized that ours, which not only focuses on learning and data mining methods but, at the same time, is open to all sorts of problems related to preference acquisition. Continuing our previous efforts in this direction, the main goal of the workshop is to make another step toward establishing the emerging topic of preference learning as a proper subfield of machine learning. Moreover, the workshop aims at providing a forum for the discussion of recent advances in the use of machine learning and data mining methods for problems related to the learning and discovery of preferences, and to offer an opportunity for researchers and practitioners to identify new promising research directions. The workshop should be half-day (preceded by a half-day tutorial). As we address a quite recent research topic which is still developing in a very dynamic way, we would particularly like to encourage the presentation of somewhat preliminary results. Therefore, we shall solicit two types of contributions, namely short communications (short talks) and full papers (long talks).

5

Format of the workshop

The final format of the workshop will of course depend on the number of participants, but ideally we plan to have around 10 full presentations, with ample time for discussion. In order to facilitate discussions in-between (which we believe are of uttermost importance for a workshop), we are considering special measures, such 1

See http://www.uni-marburg.de/fb12/kebi/si-ai

4

as nominating “discussants” for each paper, i.e., persons different from the authors who have carefully read the paper and prepared a few questions.

6

Workshop chairs and tutorial presenters

Name: Institution: Postal address Phone : Fax: Email: Internet: Name: Institution: Postal address: Phone: Fax: Email: Internet:

7

Johannes F¨ urnkranz TU Darmstadt, Fachbereich Informatik Hochschulstraße 10 64289 Darmstadt 06151 16-6238 06151 16-5482 [email protected] http://www.ke.tu-darmstadt.de/ Eyke H¨ ullermeier Universit¨ at Marburg, FB Mathematik und Informatik Hans-Meerwein-Straße 35032 Marburg 06421 28–21569 06421 28-21573 [email protected] http://www.uni-marburg.de/fb12/kebi

Organizers’ experience

Both organizers have a longstanding experience in machine learning research in general and in preference learning in particular. In recent years, they have made several joint contributions to preference learning and related fields [22, 8, 17, 16, 9, 18, 3, 19, 10, 20, 21], and they are currently working on a joint project “Learning by Pairwise Comparison for Problems with Structured Output Spaces” (funded by the German Science Foundation). Both organizers are regularly serving in the PC of major machine learning conferences and are members of the editorial board of several journals. Previous and ongoing organization of conferences/workshops: Johannes F¨ urnkranz: Co-chair of the ICML 2010 and the ECML/PKDD 2007 conferences; workshop or tutorial chair at various conferences (e.g., ECML/PKDD 2001, ECML/PKDD 2002, ICML-04, DS-05), organizer of workshop on “Advances in Inductive Rule Learning” (ECML/PKDD 2004), 5

co-organizer of workshops on “Preference Learning” (ECML/PKDD 2008 and 2009), “From Local Patterns to Global Models” (ECML/PKDD 2008), “ILP for KDD” (ICML-96), “Machine Learning in Games” (ICML-99) and “Preference Learning: Models, Methods, Applications” (KI-03). Eyke H¨ ullermeier: General chair of IPMU–2010, 13th Int. Conf. on Information Processing and Management of Uncertainty in Knowledge-Based systems, co-organizer of workshops “Preference Learning” (ECML/PKDD 2008 and 2009), “Uncertainty and Fuzziness in Case-Based Reasoning” (part of 7th International Conference on Case-Based Reasoning, Belfast, Northern Ireland, August 2007), “Uncertainty and Fuzziness in Case-Based Reason¨ udeniz, ing” (part of 8th European Conference on Case-Based Reasoning, Ol¨ Turkey, September 2006), “Symposium on Fuzzy Systems in Computer Science 2006” (Magdeburg, September 2006), “Soft Computing for Information Mining” (part of 27th German Conference on Artificial Intelligence, Ulm, September 2004), “Preference Learning: Models, Methods, Applications” (part of 26th German Conference on Artificial Intelligence, Hamburg, September 2003), “Entscheiden bei unvollst¨andiger Information: Neuere Methoden und Anwendungen” (part of GI-Jahrestagung, Conference of the `PC–2000, German Informatics Society, Dortmund, September 2002), Ra French Workshop on Case-Based Reasoning (Toulouse, May 2000).

8

People

As mentioned previously, preference handling is a quite interdisciplinary research topic. Therefore, we expect to attract attendees not only from the core machine learning and data mining field, but also from adjacent research areas, such as databases, statistics and decision theory, logic and AI. The following list contains a number of people actively working on topics related to the workshop and, hence, potential contributors as well as reviewers: Fabio Aiolli, Shotaro Akaho, Bob Arens, Craig Boutilier, Jerzy W. Bala, Bernard De Baets, Alejandro Bellogin, Paul Bennett, Darius Braziunas, Chris Burges, Ivan Cantador, Pablo Castells, Cassio de Campos, Weiwei Cheng, Yann Chevaleyre, Wei Chu, William Cohen, Fabio Cozman, Koby Crammer, K. Dembczynski, Carmel Domshlak, Alan Eckhardt, Eibe Frank, Thomas Gaertner, Joachim Giesen, Zoubin Ghahramani, Thore Graepel, Ali Hadjarian, Brent Han, Ralf Herbrich, Samuel Hiard, Thomas Hofmann,

6

Thorsten Joachims, Toshiro Kamishima, Alexandros Karatzoglou, Daniel Kikuti, Frederic Koriche, W. Kotlowski, Madhu Kurup, Jerome Lang, Paul H. Lee, Wei Li, Eneldo Loza Menc´ıa, Jerome Mengin, Wai Ming, Klaus Mueller, Alvaro Ortigosa, Klaus Obermayer, Sang-Hyeun Park, Patrice Perny, Filip Radlinksi, Jason Rennie, Giovanni Semeraro, Yoram Singer, R. Slowinski, Alessandro Sperduti, Nati Srebro, Bilyana Taneva, Vicenc Torra, Volker Tresp, Evgeni Tsivtsivadze, Antti Ukkonen, Nicolas Usunier, Stijn Vanderlooy, Shankar Vembu, Peter Vojtas, Willem Waegeman, Toby Walsh, W.M. Wan, Kiri Wagstaff, Markus Weimer, Kiri Wagstaff, Fusun Yaman, Shipeng Yu, Philip L.H. Yu, Bruno Zanuttini, Jianping Zhang, Peter Zolliker. The Programme committee of the previous editions of the workshop included Fabio Aiolli, Bernard De Baets, Paul Bennett, Weiwei Cheng, Wei Chu, Eibe Frank, Toshihiro Kamishima, Eneldo Loza Menc´ıa, Sang-Hyeun Park, Antti Ukkonen, Stijn Vanderlooy.

9

Publication of workshop results

We will edit on-line proceedings of all accepted papers so that the results are widely accessible. If there is sufficient interest and quality of the papers, as we do expect, we will also consider a post-workshop publication in the form of a special issue of a journal.

References [1] G. Bakir, T. Hofmann, B. Sch¨olkopf, A. Smola, B. Taskar, and S. Vishwanathan, editors. Predicting structured data. MIT Press, 2007. [2] J.S. Breese, D. Heckerman, and C. Kadie. Empirical analysis of predictive algorithms for collarborative filtering. In Proceedings UAI–98, Madison, WI, 1998. [3] K. Brinker, J. F¨ urnkranz, and E. H¨ ullermeier. A unified model for multilabel classification and ranking. In Proceedings ECAI–2006, 17th European Conference on Artificial Intelligence, pages 489–493, Riva del Garda, Italy, 2006. [4] O. Dekel, CD. Manning, and Y. Singer. Log-linear models for label ranking. In S. Thrun, LK. Saul, and B. Sch¨olkopf, editors, Advances in Neural Information Processing Systems (NIPS-2003). MIT Press, 2004.

7

[5] J. Dombi, C. Imreh, and N. Vincze. Learning lexicographic orders. European Journal of Operational Research, 183:748–756, 2007. [6] J. Doyle. Prospects for preferences. 20(2):111–136, 2004.

Computational Intelligence,

[7] P. Flach and T. Matsubari. A simple lexicographic ranker and probability estimator. In Proceedings ECML–07, 17th European Conference on Machine Learning, Warsaw, Poland, 2007. Springer-Verlag. [8] J. F¨ urnkranz and E. H¨ ullermeier. Pairwise preference learning and ranking. In Proc. ECML–03, 13th European Conference on Machine Learning, Cavtat-Dubrovnik, Croatia, September 2003. [9] J. F¨ urnkranz and E. H¨ ullermeier. Preference learning. K¨ unstliche Intelligenz, 1/05:60–61, 2005. [10] J. F¨ urnkranz, E. H¨ ullermeier, E. Mencia, and K. Brinker. Multilabel classification via calibrated label ranking. Machine Learning, 73(2):133–153, 2008. [11] J. F¨ urnkranz and E. H¨ ullermeier, editors. Springer-Verlag, 2010.

Preference Learning.

[12] J. F¨ urnkranz and E. H¨ ullermeier. Preference learning: An introduction. In Preference Learning [11]. To appear. [13] J. Goldsmith and U. Junker. Special issue on preference handling for artificial intelligence. AI Magazine, 29(4), 2008. [14] Vu Ha and P. Haddawy. Similarity of personal preferences: theoretical foundations and empirical analysis. Artificial Intelligence, 146:149–173, 2003. [15] S. Har-Peled, D. Roth, and D. Zimak. Constraint classification: a new approach to multiclass classification. In Proceedings 13th Int. Conf. on Algorithmic Learning Theory, pages 365–379, L¨ ubeck, Germany, 2002. Springer. [16] E. H¨ ullermeier and J. F¨ urnkranz. Comparison of ranking procedures in pairwise preference learning. In IPMU–04, 10th International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, Perugia, Italy, 2004. [17] E. H¨ ullermeier and J. F¨ urnkranz. Ranking by pairwise comparison: A note on risk minimization. In FUZZ-IEEE–04, IEEE International Conference on Fuzzy Systems, Budapest, Hungary, 2004. 8

[18] E. H¨ ullermeier and J. F¨ urnkranz. Learning label preferences: Ranking error versus position error. In Proceedings IDA–05, 6th International Symposium on Intelligent Data Analysis, pages 180–191, Madrid, 2005. [19] E. H¨ ullermeier and J. F¨ urnkranz. On minimizing the position error in label ranking. In Proceedings ECML–07, 17th European Conference on Machine Learning, Warsaw, Poland, 2007. Springer-Verlag. [20] E. H¨ ullermeier and J. F¨ urnkranz. Learning preference models from data: On the problem of label ranking and its variants. In G. Della Riccia, D. Dubois, R. Kruse, and H.J. Lenz, editors, Preferences and Similarities, pages 283–304. Springer-Verlag, Wien, New York, 2008. [21] E. H¨ ullermeier and J. F¨ urnkranz. On loss functions in label ranking and risk minimization by pairwise learning. Journal of Computer and System Sciences, 76(1):49–62, 2010. [22] E. H¨ ullermeier, J. F¨ urnkranz, W. Cheng, and K. Brinker. Label ranking by learning pairwise preferences. Artificial Intelligence, 172:1897–1917, 2008. [23] E. H¨ ullermeier and J. F¨ urnkranz, editors. Proceedings of the ECML/PKDD-08 Workshop on Preference Learning, Antwerp, Belgium, 2008. [24] E. H¨ ullermeier and J. F¨ urnkranz, editors. Proceedings of the ECML/PKDD-09 Workshop on Preference Learning, Bled, Slovenia, 2009. [25] T. Joachims. Optimizing search engines using clickthrough data. In KDD–2002, Proceedings of the ACM Conference on Knowledge Discovery and Data Mining, 2002. [26] T. Joachims, L. Granka, B. Pan, H. Hembrooke, F. Radlinski, and G. Gay. Evaluating the accuracy of implicit feedback from clicks and query reformulations in web search. ACM Transactions on Information Systems, 25(2), 2007. [27] T. Joachims, L. Granka, B. Pang, H. Hembrooke, and G. Gay. Accurately interpreting clickthrough data as implicit feedback. In Proc. SIGIR–05, Conference on Research and Development in Information Retrieval, Salvador, Brazil, 2005. [28] F. Radlinski and T. Joachims. Active exploration for learning rankings from clickthrough data. In Proceedings International Conference on

9

Knowledge Discovery and Data Mining, pages 570–579, San Jose, CA, USA, 2007. [29] M. Schmitt and L. Martignon. On the complexity of learning lexicographic strategies. Journal of Machine Learning Research, 7:55–83, 2006. [30] T. Walsh. Representing and reasoning with preferences. AI Magazine, 28(4):58–68, 2007.

10