A New Recommender System for the Interactive Radio Network FMhost

Abstract

We describe a new recommender system for the Russian interactive radio network FMhost. The new recommender model combines collaborative and user-based approaches. The system extracts information from tags of listened tracks for matching user and radio station pro les and follows an adaptive online learning strategy based on user history. We also provide some basic examples and describe the quality of service evaluation methodology.

Keywords

• music recommender systems
• interactive radio network
• ecommerce
• quality of service

Introduction and related work

Music recommendation is an important topic in the eld of recommender systems; see, e.g. the proceedings of International Society for Music Information Retrieval Conference (ISMIR ) [ 1 ], Workshop on Music Recommendation and Discovery (WOMRAD) [ 2,3 ], and Recommender Systems conference (RecSys) [ 4 ]. Although such services as LastFm, Yahoo!LaunchCast and Pandora are well known, they work on a commercial basis and, moreover, the latter two of them do not broadcast for Russia. Despite many high-quality papers on di erent aspects of music recommendation, there are only few studies on radio station recommender systems for online services.

This work is devoted to the Russian online radio hosting service FMhost and, in particular, its new hybrid recommender subsystem. Recently, the focus of computer science research for the music industry has shifted from music information retrieval and exploration [ 5,6,7 ] to music recommender services [ 8,9 ]. The topic is not new (see, e.g., [ 10 ]); however, it is now inspired by new capabilities of large online services to provide not only millions of tracks for listening to, but even radio station hosting. Social tagging is also one of the important factors which allows to apply new tag-similarity based recommender algorithms to the domain [ 11,12 ].

Recently, a widely acclaimed public contest on music recommender algorithms, KDD Cup, was held by Yahoo! (http://kddcup.yahoo.com/). In KDD Cup, track 1 was devoted to learning to predict users' ratings of musical items (tracks, albums, artists and genres) in which items formed a taxonomy. Each track belonged to an album, albums belonged to artists, and together they were tagged by genres. Track 2 aimed at the developing learning algorithms for separating tracks scored highly by speci c users from tracks not scored by them. It attracted a lot attention from the community to problems which are both typical for recommender systems and speci c for music recommendation: scalability issues, capturing the dynamics and taxonomical properties of items [ 13 ]. The current trends of music recommender systems re ect advantages of hybrid approaches and show the need for user-centric quality measures [ 14 ]. For instance, in [ 15 ] an interesting approach based on \forgetting curve" to evaluate \freshness" of predictions was proposed. In [ 16 ], the authors posed an important question, namely how much metadata do we need in music recommendation, and after a subjective evaluation of 19 users the authors concluded that pure content-based methods can be drastically improved by using genres.

In [ 17 ], the authors proposed the music recommender system Starnet for social networking. It generates recommendations based either on positive ratings of friends (social recommendations), positive ratings of others in the network (nonsocial recommendations), and it also makes random recommendations. Another interesting online music recommendation system we can mention is Hotttabs [ 18 ], dedicated to guitar learning. Some authors aim at improving music recommender systems by using semantic extraction techniques [ 19,20 ]. Paper [ 21 ] describes a system of genre recommendation for music and TV programs, which can be considered as an alternative channel selector. The authors of paper [ 22 ] proposed a recommender system GroupFan which is able to aggregate preferences of group users to their mutual satisfaction.

Many online services (e.g., Last.fm or LaunchCast) call their audio streams \radio stations", but in reality they produce a playlist from a database of tracks based on a recommender system rather than actually recommend a radio channel. FMhost, on the other hand, provides users with online radio stations in the classical meaning of this term: there are human DJs who perform live, a radio station actually represents a strategy or mood of a certain person (DJ), they play their own tracks, perform contests etc. Thus, the problem we are solving di ers from most of the work done in music recommendation, and some of the challenges are unique.

The paper is organized as follows. In Section 2, we describe our online service FMhost. In section 3, we propose our new recommender model, two basic recommender algorithms, and describe the recommender system architecture. Quality of Service (QoS) measurement for the system and some insights on FMhost user behaviour are discussed in Section 4. Section 5 concludes the paper.

Online service FMhost.me

2.1

A concise online broadcasting dictionary

Before we proceed, we need to shortly explain some basic domain terminology.

A chart is a radio station track rating; for example, the rock chart shows a certain number (say, 10) of most popular rock tracks, ranked from the most popular (rank 1) to the least popular (rank 10) according to the survey. A live performance (or just live for short) is a performance to which one or several DJs (disk jockeys ) are assigned. They do it from their own PCs, and the audio stream is being redirected from them to the Icecast server and then everywhere. Also they may have their own blog for each live, where people may interact with DJs who perform live. LiquidSoap is a sound generator that broadcasts audio les (*.mp3, *.aac etc.) into an audio stream. Icecast is a retranslation server that redirects audio stream from one source, for example LiquidSoap, to many receivers. 2.2

The FMhost project

FMhost is an interactive radio network. This portal allows users to listen and broadcast their own radio stations. There are four user categories in the portal: (1) unauthorized user; (2) listener; (3) Disk Jockey (DJ); (4) radio station owner.

User capabilities vary upon their status. Unauthorized listeners can listen to any station, but they cannot vote or become DJs. They also cannot use the recommender system and the rating system.

Listeners, unlike unauthorized users, can vote for tracks, lives, and radio stations. They can use a recommender system or rating system. They can subscribe to lives, radio stations, or DJs. They also can be appointed to a live and become a DJ.

There are three types of broadcasting: (1) stream redirection from another server; (2) AutoDJ translation; (3) live performance.

Stream redirection applies when a radio station owner has its own server and wants to use FMhost as a broadcasting platform, but also wants to broadcast using his own sound generator, e.g., SamBroadcaster (http://spacial.com/ sam-broadcaster), LiquidSoap (http://savonet.sourceforge.net/) etc. AutoDj is a special option that allows the users to play music directly from the FMhost server. Every radio owner gets some space where he can download as much tracks as he can, and then LiquidSoap will generate the audio stream and the Icecast (http://www.icecast.org/) server will redirect it to the listeners. Usually the owner sets a radio schedule which is being played.

Live performances are done by DJs. Everyone who has performed live at least once can be called a DJ. He can also be added to a radio station crew. Moreover, a DJ can perform lives at any station, not only on his own station where he is in a crew.

FMhost was the rst project of its kind in Russia, starting in 2009. Nowadays, following FMhost's success, there exist several radio broadcasting portals, such as http://frodio.com/, http://myradio24.com/, http://myradio24.com/, http://www.radio-hoster.ru/, http://www.taghosting.ru/, http://www. economhost.com/, and even http://fmhosting.ru/. In late 2011, FMhost was taken down for a serious rewrite of the codebase and rethinking of the recommender system's architecture. In this paper, we describe the results of this upgrade.

The previous version of the recommender system experienced several problems, such as tag discrepancy or personal tracks without tags at all. A survey by FMhost with about a hundred respondents showed that more than half of them appreciated the previous version of our recommender system and more than 80% of the answers were positive or neutral (see Table 1); nevertheless, we hope that the new recommender model and algorithms provide even more accurate recommendations and make even less prediction mistakes.

User opinion I like it very much, all recommendations were relevant Good, I like most of the radio stations Sometimes there are interesting stations I like only few recommended radio stations None of the recommended stations was satisfactory The new version features a more complex system of user interaction. Every radio station has an owner who is not just a name but also has the ability to assign DJs for lives, prepare radio schedule, and assign lives and programs. There will be a new broadcasting panel for DJs that will allow them to play tracks with additional features that were not available before, such as an equalizer or fading between tracks. A new algorithm for the recommender system, a new rating system, and a new chart system will be launched.

The rating system has been developed to rank radio stations and DJs according to their popularity and quality of work. A new core is being implemented and a new concept of LiquidSoap and Icecast is being designed. The system is designed so that to eliminate all problems that have surfaced in the previous version.

Models, algorithms and recommender architecture

3.1

Input data and general structure

Our model is based on three data matrices. The rst matrix A = (aut) tracks the number of times user u visits radio stations with a certain tag t. Each radio station r broadcasts audio tracks with a certain set of tags Tr. The sets of all users, radio stations, and tags are denoted by U , R, and T respectively. The second matrix B = (brt) contains how many tracks with a tag t a radio station r has played. Finally, the third matrix C = (cur) contains the number of times a user u visits a radio station r. For each of these three matrices, we denote by vA, vB, and vC the respective vectors containing sums of elements: vA = P aut, vB = P brt, and vC = P aur. We also denote for each matrix t2T t2T r2R A, B, C the corresponding frequency of visits matrix by Af , Bf , and Cf ; the frequency matrix is obtained by normalizing the matrix with the respective visits vector, e.g., Af = (aut (vuA) 1). Our model is not purely static; the matrices A, B, and C change after a user u visits a radio station r with a tag t, i.e., each value aut, brt, and cur is incremented by 1 after this visit.

The model consists of three main blocks: the Individual-Based Recommender System (IBRS) model, the Collaborative-Based Recommender System (CBRS) model, and the End Recommender Systems (ERS) that aggregates the results of the former two.

Each model has its own algorithmic implementation. Since both our previous works [ 23,24 ] and this work implicitly use biclustering ideas, we continue to name our general algorithms with the RecBi acronym; this time it is the RecBi3 family. We call the resulting algorithms for the three proposed models RecBi3.1, RecBi3.2, and RecBi3.3, respectively. Here we do not use notation from formal concept analysis, but refer to [ 25 ] for the basic notation used in our previous algorithms RecBi2.1 and RecBi2.2. 3.2

IBRS

The IBRS model uses matrices Af and Bf and aims to provide a particular user u0 2 U with top N recommendations represented mathematically by a special structure T opN (u). Formally, T opN (u0) is a triple (Ru0 ; u0 ; rank), where Ru0 is the set of at most N radio stations recommended to a particular user u0, u0 is a well-de ned quasiordering (re exive, transitive, and complete) on the set Ru0 , and rank is a function which maps each radio station r from Ruo to [0; 1].

The RecBi3.1 algorithm computes the 1-norm distance between a user u0 and a radio station r, i.e.m d(u0; r) = P t 2 T jau0t brtj. Then all distances between the user u0 and the radio stations r 2 R are calculated. Further the algorithm constructs the relation u0 according to the following rule: ri rj i d(u0; ri) d(u0; ri). The function rank operates on Ru0 according to the following rule: rank(ri) = 1 d(u0; ri)= max d(u0; rj ): rj2R Finally, after selecting N radio stations for N greatest rank values in the set Ru0 , we have the structure T opN (u0) which represents a ranked list of radio stations recommended to the user u0.

As shown in Fig. 1, our model takes into account not only listened tracks but also liked tracks, liked radio stations, and favorite radio stations. To re ne the IBRS submodel we tune it with the SMARTS algorithm known from decision making theory [ 26 ]. According to the method and expert decisions, we should count each track tag of a listened radio station, liked radio station, liked track, and favorite radio station with a di erent weight. The SMARTS procedure provided us with the four weights for listened radio station, liked radio station, favorite radio station, and liked track according to our experts' assessment of mutual criterion importance, namely 0:07, 0:16, 0:3, and 0:47. In the SMARTS method, we consider each tag type as a criterion with two terminal values 0 and 100% on a real number scale. Some tag t may have some or even four of these types simultaneously; in this case, the algorithm adds to aut the total weight of the tag (i.e., the sum of weights) after a user u visits some radio station with this tag. In case there are several elements with the same rank so that T opN (u) is not uniquely de ned, we simply choose the rst elements according to some arbitrary ordering (e.g., the lexicographic ordering of station names). The CBRS model is based on the Cf matrix. The matrix also yields a vector nC which stores the total number of listened stations for each user u 2 U . This vector also changes over time, and this value is used as a threshold to transform matrix Cf to distance matrix D as follows: dij = jcfir cfjrj; if cfir ni 1and cfjr nj 1 jcfir + cfjrj; if cfir > ni 1and cfjr < nj 1or vice versa (1)

This distance takes into account the frequency nuC of all radio station visits for user u and considers its inverse value as a threshold to decide whether a particular station r should be considered as popular for this user. Thus, users with di erent signs of cfir ni 1 and cfjr nj 1 become more distant than for the conventional absolute distance. This distance dij actually serves as a sort of polarizing lter, and in Section 4 we compare it with common approaches.

After computing D, the algorithm RecBi3.2 constructs the list T opk(u0) = (Uu0 ; u0 ; sim) of k users similar to our target user u0 who awaits recommendations, where sim(u) = 1 duu0 = um02aUx du0u0 . We de ne the set of all radio stations user u0 listened to as L(u0) = frjcfur = 0g. In a similar way, we de ne T opN (u0) = (Ru0 ; u0 ; rank); where rank(r) = sim(u ) cfu r and u = arg

max u2Uu0 ;r2U=L(u0) sim(u) cfur: It is worth mentioning that rank : r 7! [0; 1]. The problem of choosing exactly N topmost stations is solved in the same way as in the IBRS submodel. 3.4

ERS After IBRS and CBRS have nished, we are left with two ranked lists of recommended stations T opIN (u0) and T opCN (u0) for our target user u0 from IBRS and CBRS respectively. The ERS submodel proposes a simple solution for aggregating these lists into the nal recommendation structure T opEN (u0) = (RuE0 ; uE0 ; rankE ). For every r 2 RuC0 [ RuI0 , the function rankE (r) maps r to the weighted sum rankC (r) + (1 ) rankI (r); where we let 2 [ 0,1 ], rankC (r) = 0 for all r 62 RC and rankI (r) = 0 for all r 62 RI . The algorithm RecBi3.3 adds the best N radio stations according to this criterion to the set RuC0 . 4

Quality of service assessment

To evaluate the quality of the developed system, we propose a variant of the cross-validation technique [ 27 ]. Before we proceed to the detailed description of the procedure, we discuss some important analyses that we conducted on the FMhost data for the period from 2009 till 2011.

Basic statistics

It is a well-known fact that social networking data often follows the so called power law distribution [ 28 ]. To decide which amount of active users or radio stations we have to take into account for making recommendations, we performed a simple statistical analysis of user and radio station activity. Around 20% of the users (only registered ones) were analysed.

W = P ( 2)=( 1); which means that the fraction W of the wealth is in the hands of the richest P of the population. In our case, 50% of users make 80% of all radio station visits, and 50% of radio stations have 83% of all visits. Thus, if the service tends to take into account only active stations and users, it can cover 80% of all visits by considering only 50% of their active audience. However, new radio stations still deserve to be recommended, so this rule can only be applied to the user database. 4.2

Quality assessment

To evaluate QoS for the IBRS subsystem (RecBi3.1 algorithm), we count average precision and recall on the set RN R, where N is a number of randomly \hidden" radiostations. We suppose that for all r in RN and every user u 2 U the algorithm does not know whether the radio stations were liked, added to favorites, or even visited, and we change Af and R accordingly. Then RecBi3.1 attempts to recommend Top-N radio stations for this modi ed matrix Af .

Top-N average precision and recall are computed as follows: ; : ; :

To deal with CBRS, we use a modi cation of the leave-one-out technique. At each step of the procedure for a particular user u, we \hide" all radio stations r 2 RN by setting cfur = 0. Then we perform RecBi3.2 assuming that cfu0r is unchanged for u0 2 U=u. After that we compute

To tune the ERS system, we can use a combination of these two procedures trying to nd the optimal as = arg max 2 Precision Recall (Precision + Recall)

We suppose that in one month of active operation we will have enough statistics to tune and choose appropriate similarity and distance measures as well as Precision = u2U

Recall = u2U

P jRuI\Lu\RN j

jLu\RuIj jU j P jRuI\Lu\RN j jLu\RN j jU j Precision = u2U Recall = u2U

P jRuC\Lu\RN j

jLu\RuCj jU j P jRuC\Lu\RN j jLu\RN j jU j thresholds. We suppose that the resulting system will provide reasonably accurate recommendations using only a single (last) month of user history and only 50% of the most active users. For quality assessment during the actual operation, we will compute Top-3, Top-5, and Top-10 Precision and Recall measures as well as whether the system provides a user only with Top-10 items with a highest rank. In addition, online surveys can be launched to assess user satisfaction with the new RS system. 5

Conclusion and further work

In this work, we have described the underlying models, algorithms, and the system architecture of the new improved FMhost service. We hope that the developed algorithms will help a user to nd relevant radio stations to listen to. In future optimization and tuning, special attention should be paid to scalability issues and user-centric quality assessment. We consider matrix factorization techniques as a reasonable tool to increase scalability, but it has to be carefully adapted and assessed taking into account the folksonomic nature of tracks tags. Another attractive feature of the developed system is that it can serve as a kind of World of Music map built on track-to-track similarity matrices with tags [ 7 ]. Another important issue is dealing with the triadic relational nature of data (users, radio stations (tracks), and tags), which constitutes the so called folksonomy [ 29 ], a primary data structure in tagging resource-sharing systems. As shown in [ 30 ], this data can be successfully mined by means of triclustering, so we also plan to build a tag-based recommender system by means of triclustering. Acknowledgments. We would like to thank Rustam Tagiew, Jonas Poelmans and Mykola Pechenizkiy for their comments, remarks and explicit and implicit help during paper preparations. Work of Sergey Nikolenko has been supported by the Russian Foundation for Basic Research grant 12-01-00450-a, the Russian Presidential Grant Programme for Young Ph.D.'s, grant no. MK-6628.2012.1, for Leading Scienti c Schools, grant no. NSh-3229.2012.1, and RFBR grants 1101-12135-o -m-2011 and 11-01-00760-a.

References

1. Klapuri , A. , Leider , C., eds. : Proceedings of the 12th International Society for Music Information Retrieval Conference , ISMIR 2011 , Miami, Florida, USA, October 24 - 28 , 2011 . In Klapuri, A. , Leider , C., eds.: ISMIR, University of Miami ( 2011 )

2. Anglade , A. , Baccigalupo , C. , Casagrande , N. , Celma , O. , Lamere , P. : Workshop report: Womrad 2010 . In Amatriain, X. , Torrens , M. , Resnick , P. , Zanker , M., eds.: RecSys, ACM ( 2010 ) 381 { 382

3. Anglade , A. , Celma , O. , Fields , B. , Lamere , P. , McFee , B. : Womrad: 2nd workshop on music recommendation and discovery . In: Proceedings of the fth ACM conference on Recommender systems. RecSys '11 , New York, NY, USA, ACM ( 2011 ) 381 { 382

4. RecSys '11: Proceedings of the fth ACM conference on Recommender systems , New York, NY, USA, ACM ( 2011 ) 609116 .

5. Hilliges , O. , Holzer , P. , Kluber, R., Butz , A. : Audioradar: A metaphorical visualization for the navigation of large music collections . In Butz, A., Fisher , B.D. , Kruger, A. , Olivier , P., eds.: Smart Graphics . Volume 4073 of Lecture Notes in Computer Science., Springer ( 2006 ) 82 { 92

6. Gleich , D.F. , Rasmussen , M. , Lang , K. , Zhukov , L. : The world of music: User ratings; spectral and spherical embeddings; map projections . Online report ( 2006 )

7. Gleich , D.F. , Zhukov , L. , Rasmussen , M. , Lang , K. : The World of Music: SDP Embedding of High Dimensional data . In: Information Visualization 2005 . (2005) Interactive Poster .

8. Brandenburg , K. , Dittmar , C. , Gruhne , M. , Abeer , J. , Lukashevich , H. , Dunker , P. , Grtner , D. , Wolter , K. , Grossmann , H.: Music search and recommendation . In Furht, B., ed.: Handbook of Multimedia for Digital Entertainment and Arts. Springer US ( 2009 ) 349 { 384

9. Celma , O. : Music Recommendation and Discovery - The Long Tail, Long Fail, and Long Play in the Digital Music Space . Springer ( 2010 )

10. Avesani , P. , Massa , P. , Nori , M. , Susi , A. : Collaborative radio community . In: Proceedings of the Second International Conference on Adaptive Hypermedia and Adaptive Web-Based Systems. AH '02 , London, UK, UK, Springer-Verlag ( 2002 ) 462 { 465

11. Symeonidis , P. , Ruxanda , M.M. , Nanopoulos , A. , Manolopoulos , Y. : Ternary semantic analysis of social tags for personalized music recommendation . In Bello, J.P. , Chew , E. , Turnbull , D., eds.: ISMIR. ( 2008 ) 219 { 224

12. Nanopoulos , A. , Rafailidis , D. , Symeonidis , P. , Manolopoulos , Y. : Musicbox: Personalized music recommendation based on cubic analysis of social tags . IEEE Transactions on Audio, Speech & Language Processing 18 ( 2 ) ( 2010 ) 407 { 412

13. Koenigstein , N. , Dror , G. , Koren , Y. : Yahoo! music recommendations: modeling music ratings with temporal dynamics and item taxonomy . In: Proceedings of the fth ACM conference on Recommender systems. RecSys '11 , New York, NY, USA, ACM ( 2011 ) 165 { 172

14. Celma , O. , Lamere , P. : Music recommendation and discovery revisited . In: Proceedings of the fth ACM conference on Recommender systems. RecSys '11 , New York, NY, USA, ACM ( 2011 ) 7{ 8

15. Hu , Y. , Ogihara , M. : Nextone player: A music recommendation system based on user behavior . [1] 103{108

16. Bogdanov , D. , Herrera , P. : How much metadata do we need in music recommendation? a subjective evaluation using preference sets . [ 1 ] 97 { 102

17. Mesnage , C.S. , Ra q , A., Dixon , S. , Brixtel , R.P. : Music discovery with social networks . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 1 { 6

18. Barthet , M. , Anglade , A. , Fazekas , G. , Kolozali , Sefki Macrae, R.: Music recommendation for music learning: Hotttabs, a multimedia guitar tutor . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 7 { 13

19. Tatl , I. , Birturk , A. : Using semantic relations in context-based music recommendations . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 14 { 17

20. Knees , P. , Schedl , M. : Towards semantic music information extraction from the web using rule patterns and supervised learning . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 18 { 25

21. Knopke , I. : The importance of service and genre in recommendations for online radio and television programmes . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 26 { 29

22. Popescu , G. , Pu , P. : Probabilistic game theoretic algorithms for group recommender systems . In: Workshop on Music Recommendation and Discovery 2011 , Workshop on Music Recommendation and Discovery ( October 2011 ) 7 { 12

23. Ignatov , D. , Poelmans , J. , Zaharchuk , V. : Recommender System Based on Algorithm of Bicluster Analysis RecBi . In Ignatov, D., Poelmans , J. , Kuznetsov , S., eds. : CEUR Workshop proceedings Vol- 757 , CDUD' 11 - Concept Discovery in Unstructured Data. ( 2011 ) pp. 122 { 126

24. Ignatov , D.I. , Kuznetsov , S.O. : Concept-based Recommendations for Internet Advertisement . In Belohlavek, R., Kuznetsov , S.O., eds. : Proc. CLA 2008 . Volume Vol. 433 of CEUR WS., Palack University, Olomouc, 2008 ( 2008 ) 157 { 166

25. Ganter , B. , Wille , R.: Formal Concept Analysis: Mathematical Foundations. 1st edn . Springer-Verlag New York, Inc., Secaucus, NJ, USA ( 1999 )

26. Edwards , W. , Barron , F. : SMARTS and SMARTER: Improved Simple Methods for Multiattribute Utility Measurement . Organizational Behavior and Human Decision Processes 60 ( 3 ) ( 1994 ) 306 { 325

27. Ignatov , D.I. , Poelmans , J. , Dedene , G. , Viaene , S.: A New Cross-Validation Technique to Evaluate Quality of Recommender Systems . In Kundu, M.K. , Mitra , S. , Mazumdar , D. , Pal , S.K., eds.: PerMIn . Volume 7143 of LNCS., Springer ( 2012 ) 195 { 202

28. Clauset , A. , Shalizi , C.R. , Newman , M.E.J. : Power-law distributions in empirical data . SIAM Rev . 51 ( 4 ) ( November 2009 ) 661 { 703

29. Vander Wal , T. : Folksonomy Coinage and De nition ( 2007 ) http://vanderwal.net/folksonomy. html (accessed on 12.03 . 2012 ).

30. Ignatov , D.I. , Kuznetsov , S.O. , Magizov , R.A. , Zhukov , L.E. : From Triconcepts to Triclusters. In: Proceedings of the 13th international conference on Rough sets, fuzzy sets, data mining and granular computing . RSFDGrC'11 , Berlin, Heidelberg, Springer-Verlag ( 2011 ) 257 { 264