322 M. Furini and M. Montangero two other policies, the revenue of a customer depends on the number of customers that receive the song directly and indirectly from him/her. Thus, the way in which the song spreads among customers greatly influences their revenue; e.g., a customer might have a big income with little effort if the customers to whom the song is delivered put a lot of effort in reselling the song; customers might get an unexpected reward also after a long period of time from the moment he/she sold the song. The proportional reward policy produces better results than the equal distribution, as a customer needs a smaller number of customers in its subtree, even if to get a full refund of C I , the minimum number of customers is reasonable in both cases. Moreover, although we analyzed the worst-case scenario, this is unlikely to hap- pen in reality, especially if we think that music is distributed according to social relationships. Hence, with a multi-channel distribution strategy, in average, even a smaller number of customers has to be reached and it is likely that many customers (the higher they are in the tree, the better) might have a revenue grater than C I . From the store point of view, the choice of the policy depends on which, among the customers, the store wants to favor: the selfish policy favors the ones that buy from the store; the equal favors the customers that join the music distribution earlier; the proportional favors the customers that actually distribute the song. Related Work Content distribution in a mobile environment is a subject investigated in recent liter- ature: some are experiencing the development of ad-hoc P2P networks in a mobile environment [13, 14], others are proposing to disseminate contents in Wi-Fi based ad-hoc networks through epidemic algorithms [18, 29]. The multi-channel distribution outlined in this paper does not require a real P2P network, as the song delivery simply requires the cooperation of two customers, making the operation more similar to what happens when a friend text-messages or sends an MMS to another friend. In this case, the message content is the song and the network used is other than the cellphone network. Many of the approaches present in literature are designed to stimulate users cooperation in a P2P networks [2, 3, 25, 28]; for example, peers are asked to route queries or are limited in the use of bandwidth according to the amount of bandwidth they provided to the system. For scalability reasons, most of these mechanisms are distributed, requiring only local information available at each peer. This might lead to malicious modification of peer local information by the peer itself, and hence tamper resistant software should be employed at the user side. Our mechanism is simple to implement because it comes with a centralized control mechanism at no cost: the store can always keep track of the spreading of the song and is always able to correctly assign revenues. Whenever a user wants to play a song he/she just bought, he/she needs to buy the license from the store. At this stage, the store can easily record who sold and who bought the song, updating the information about song spreading. 14 Incentive Mechanisms for Mobile Music Distribution 323 Weare not aware of any distribution strategy that couples cellphone networks and free-of-charge technologies to distribute contents, and that makes use of incentive mechanisms to stimulate the customer cooperation in content distribution. Conclusions In this paper we analyzed the characteristics of the current mobile music scenario, investigating the communication infrastructure, the pricing strategy and the copy- right protection scheme currently used. The analysis highlighted that a replication of the strategy used to distribute contents in the Internet-based music market is not worth applying in the mobile scenario, as it presents critical problems (excessive download time and high cost). To mitigate such problems, we show that a multi-channel distribution strategy can be successful. In such a strategy, customers can re-distribute the song acquired by using the free-of-charge communication technologies provided in cellphones. We showed that by using a smart protection scheme, music sharing could avoid piracy. We also present an incentive mechanism, coupled with three different reward poli- cies, which stimulates customers cooperation by providing a financial compensation to those customers who help distributing music files. The evaluation of the multi-channel distribution strategy equipped with the pro- posed incentive mechanism showed that considerable benefits may be received by all the entities involved in the mobile music distribution, from music stores to cus- tomers, to cellphone network providers. References 1. K. C. Almeroth and A. Garyfalos. Coupons: Wide scale information distribution for wireless ad hoc networks. In Proc. of IEEE Globecom, December 2004. 2. K. G. Anagnostakis and M. B. Greenwald, Exchange-based incentive mechanisms for peer-to- peer file sharing. In International Conference on Distributed Computing Systems, 2004. 3. K. G. Anagnostakis, M. B. G. Yang, T. Condie, S. Kamvar, and H. Garcia-Molina, Non- cooperation in competitive p2p networks. In International Conference on Distributed Com- puting Systems, 2004. 4. P. Antoniadis, C. Courcoubetis and B. Strulo, Incentives for content availability in memory- less peer-to- peer file sharing systems. SIGecom Exch., Vol.5, No. 4, pp. 11–20, 2005, ACM press. 5. K. Biddle, P. England, M. Peinado, and B. Willman. The darknet and the future of content distribution. In Proc. of the ACM Workshop on DRM, 2002. 6. B. Brown, A. J. Sellen, and E. Geelhoed. Music sharing as a computer supported collabora- tive application. In Proceedings of ECSCW 2001, Bohn, Germany, 2001. Kluwer academic publishers. 7. L. Buttyan and J. P. Hubaux. Stimulating cooperation in selforganizing mobile ad hoc net- works. ACM/Kluwer Mobile Networks and Applications (MONET), 8(5):579–592, October 2003. 324 M. Furini and M. Montangero 8. C. J. Cobb and W. D. Hoyer. Planned versus impulse purchase behavior. Journal of Retailing, 62, April 1986. 9. R. Dingledine and P. Syverson. Reliable mix cascade networks through reputation. In Pro- ceedings of the Sixth International Financial Cryptography Conference (FC02), March 2002. 10. M. Feldman, K. Lai, I. Stoica and J. Chuang, Robust incentive techniques for peer-to-peer net- works, Proceedings of the 5th ACM conference on Electronic commerce, pp. 102–111, 2004. 11. M. Furini, M. Montangero, “The Impact of Incentive Mechanisms in Multi-Channel Mobile Music Distribution”, Multimedia Tools and Applications, Vol. 37. No. 3, pp. 365–382, March 2008. Springer Netherlands Editor. 12. M. Furini, M. Montangero, “The Use of Incentive Mechanisms in Multi-Channel Mobile Music Distribution”, Proceedings of 2nd IEEE International Conference on Automated Pro- duction of Cross Media Content for Multi-Channel Distribution (AXMEDIS 2006), Leeds, UK, December 12–15. IEEE Computer Press 2006. 13. E. Harjula, M. Ylianttila, J. Ala-Kurikka, J. Riekki, J. Sauvola, Plug-and-play application platform: towards mobile peer-to-peer, Proceedings of the 3rd international conference on Mobile and ubiquitous multimedia MUM, October 2004. 14. N. Hatt BlueFramework - Application Framework for Bluetooth Enabled Mobile Phones,TIK- MA-2005-16, ETH Z ˜ A 1 = 4 rich, Switzerland, 2005. 15. T. H T. Hu, K. Wongrujira, and A. Seneviratne. Reputation in peer-to-peer networks. In Proceedings of the IEEE International Conference on Communications (ICC 2004), pages 1411 ˆ A–1415, Paris, France, June 2004. 16. D. Hughes, G. Coulson, J. Walkerdine, Free Riding on Gnutella Revisited: The Bell Tolls?, IEEE Distributed Systems online, Vol. 6, No. 6, June 2005. 17. IFIP. Digital music report 2008 – Summary. Research report, International Federation of the Phonographic Industry, 2008. [on-line] Available at http://www.ifpi.org/content/ library/DMR2008-summary.pdf 18. A. Khelil, C. Becker, J. Tian, K. Rothermel, An epidemic model for information diffusion in MANETs, Proceedings of the 5th ACM International Workshop on Modeling, Analysis and Simulation of Wireless and Mobile Systems, September 2002. 19. J. Liang, R. Kumar, Y. Xi, and K. Ross, Pollution in P2P File Sharing Systems. In IEEE Infocom, Miami, FL, USA, March 2005. 20. T.S. Messerges and E.A. Dabbish, Digital Rights Management in a 3G Mobile Phone and Beyond Proceedings of Digital Right Management (DRM), Washington, October 2003. ACM Press. 21. Microsoft Corporation, Microsoft PlayReady powers next-generation media experiences on mobile networks, http://www.microsoft.com/presspass/press/2007/feb07/02-123GSMNew TechnologyPR.mspx 22. M. J. O’Grady and G. M. P. O ˆ A’Hare Just-In-Time Multimedia Distribution in a Mobile Com- puting Environment. IEEE Multimedia, 62–74, 2004. 23. G. P. Premkumar. Alternative distribution strategies for digital music. Communication of the ACM, 9(9):89–95, September 2003. 24. P. Resnick, K. Kuwabara, R. Zeckhauser, and E. Friedman. Reputation systems. Communica- tions of the ACM, 43(12):45–48, 2000. 25. A. Roczniak and A. El Saddik, Impact of incentive mechanisms on quality of experience. Proceedings of the 13th annual ACM international conference on Multimedia, pp. 311–314, 2005. 26. D. Rook. The buying impulse. Journal of Consumer research, 14(2):189–199, September 1987. 27. N. B. Salem, M. J. L. Buttyan, and J. P. Hubaux. A charging and rewarding scheme for packet forwarding. In Proceedings of MobiHoc, June 2003. 28. Q. Sun and H. Garcia-Molina, Slic: A selfish link based incentive mechanism for unstructured peer-to-peer networks. In International Conference on Distributed Computing Systems, 2004. 14 Incentive Mechanisms for Mobile Music Distribution 325 29. Paul Tennent, Malcolm Hall, Barry Brown, Matthew Chalmers, Scott Sherwood, Toward social mobility: Three applications for mobile epidemic algorithms Proceedings of the 7th international conference on Human computer interaction with mobile devices & services MobileHCI ’05. 30. Wireless-Intelligence. World cellular connection. Research report, Wireless Intelligence, 2005. [on-line] Available at https://www.wirelessintelligence.com Chapter 15 Pattern Discovery and Change Detection of Online Music Query Streams Hua-Fu Li Introduction In recent years, many applications generate large amount of data streams in real time. For example, sensor data generated from sensor networks, online transaction flows in retail chains, stream Web click-sequences and records in Web services and applications, performance measurement in network monitoring and traffic manage- ment, call records in telecommunications. Mining data streams differs from mining traditional static data sets in two main aspects [2]:  The volume of a continuous data stream over its lifetime could be huge and fast changing.  The queries require timely answers, and the response time is short. Hence, it is not possible to store all the streaming data in main memory or even in secondary storage. This motivates the design for in-memory summary data struc- ture with small memory footprints that can support both one-time and continuous queries. Furthermore, online approach of mining such data has to sacrifice the cor- rectness of their analysis results by allowing some counting errors, i.e., it generates approximate results, and only has single pass over the data [8]. Recently, online music downloading is a hot Web service. Many companies, such as Apple’s iTunes [18], Napster [16], Loudeye, Yahoo’s MusicMatch [17], Kuro [20], KKBox [19], and EasyMusic.com, provide this Web service. Accord- ing to the reports of IFPI (International Federation of the Phonographic Industry; IFPI: http://www.ifpi.org/), there are more than 60 hundred millions of online mu- sic downloads at 2006. For example, the amount of online music downloading of Apple’s iTune is about 50 hundred millions from 2004 to 2007. Hence, knowledge discovery of such online music downloading behaviors of customers is an important research and a practical issue for data mining. H F. Li (  ) Department of Computer Science, Kainan University, Taoyuan, Taiwan e-mail: hfli@mail.knu.edu.tw B. Furht (ed.), Handbook of Multimedia for Digital Entertainment and Arts, DOI 10.1007/978-0-387-89024-1 15, c Springer Science+Business Media, LLC 2009 327 328 H F. Li Fig. 1 Computation model for music query streams The issue comes from the context of online music-downloading services (such as Apple’s iTunes, Napster, Loudeye, Yahoo’s MusicMatch, Kuro, KKBox, and EasyMusic. com), where the stream in question are streams of queries, i.e., music- downloading requests, sent to the server, and we are interested in finding the useful music melody structures requested by most customers during some period of time. The discovered patterns can be used to predict the future trend of online music styles and to personalize the Web services of online music downloading. With the processing model of music query streams presented in Fig. 1 [11, 12], the melody stream processor and the summary data structure are two major components in such a streaming environment. The user query processor receives user queries in the form of <Timestamp, Customer-ID, Music-ID>, and then transforms the queries into music data (i.e., melody sequences) in the form of <Timestamp, Customer-ID, Music-ID, Melody-Sequence> by querying the music database. Note that the com- ponent Buffer can be optionally set for temporary storage of recent music melody sequences from the music query streams. Among various techniques of data mining, frequent pattern mining is one of the most popular data mining approaches used to discover the customers’ behaviors from large data sets. However, traditional data mining techniques for discovering frequent patterns are not feasible for mining frequent patterns from such application of online music downloading. Because such data characteristic is streaming, the pro- posed methods for mining such streaming data need new capabilities such as using limited memory to maintain the essential information embedded in an unbound data stream, one-pass data scan and real time processing of each incoming data element. Mining music data is one of the most important research issues in multimedia data mining. Although several techniques have been developed for discovering and analyzing the content of static music data [3, 9, 10, 14, 15], new techniques are needed to analyze and discover the content of streaming music data. Recently, two 15 Pattern Discovery and Change Detection of Online Music Query Streams 329 efficient one-pass mining algorithms, FCS-stream [11] and MMS LMS [12], were pro- posed by Li et al. for discovering the closed frequent melody structures and the maximal frequent melody structures over the entire history of a continuous music query stream. Both algorithms are stream mining methods of a landmark window. However, the knowledge embedded in streaming data is likely to be changed as time goes by. Identifying the recent changes of data streams quickly can provide valuable information for the analysis of the streaming data [5]. Hence, we need new single-pass approaches for mining frequent patterns from the streaming online mu- sic downloading requests within a sliding window. Several one-pass mining methods [5, 6] were proposed for finding frequent patterns over data streams within a sliding window. The baseline method, called SWFI-stream, for mining frequent patterns over transaction data streams within a sliding window was proposed by Chang and Lee [5]. In the framework of SWFI- stream algorithm, there are two phases for mining frequent patterns over stream sliding windows. One is a window initialization phase. The phase is activated while the number of incoming data transactions generated so far is less than or equal to a predefined window size. The other is a window sliding phase. The second phase is activated after the window becomes full. The SWFI-stream algorithm is composed of four steps. First, all sub-patterns of a transaction are extracted. Second, these sub- patterns are inserted into a prefix-tree lattice structure. Third, all in-frequent patterns are pruned from the lattice structure. Finally, all frequent patterns are generated form the lattice structure. The first two steps are performed in the window initialization phase and the last two steps are performed in the window sliding phase. There are several performance bottlenecks of the typical solution. First, SWFI- stream needs the extra memory to maintain the original window in a temporal list and a prefix-tree lattice structure for storing the frequent patterns and semi-frequent patterns. Second, the processing complexity of enumerating each incoming trans- action is exponential, i.e., O.k 2 /, where k is the length of transaction. Third, the cost of maintaining the prefix-tree lattice structure of this typical solution is also exponential. Chi et al. [6] proposed a sliding window based algorithm, called Moment, which might be the first method to find frequent closed itemsets from transaction data streams. A summary data structure, called CET (Closed Enumeration Tree), is used in the Moment algorithm to maintain a dynamically selected set of itemsets over a sliding window. These selected itemsets consist of closed frequent itemsets and a boundary between the closed frequent itemsets and the rest of the itemsets. CET covers all necessary information because any status changes of itemsets (e.g. from infrequent to frequent or from frequent to infrequent) must be through the boundary in CET. Whenever a sliding occurs, it updates the counts of the related nodes in CET and modifies CET. Experiments of Moment algorithm show that the boundary in CET is stable so the update cost is little. However, Moment must maintain huge CET nodes for a closed frequent itemset. The ratio of CET nodes and closed frequent itemsets is about 30:1. If there are a large number of closed frequent itemsets, the memory requirement of Moment algorithm will be inefficient. 330 H F. Li In this paper, an efficient stream mining algorithm, called FTP-stream (Frequent Temporal Pattern mining of streams), is proposed to find the frequent temporal pat- terns over melody sequence streams. In the framework of our proposed algorithm, an effective bit-sequence representation is used to reduce the time and memory needed to slide the windows. The FTP-stream algorithm can calculate the support thresh- old in only a single pass based on the concept of bit-sequence representation. It takes the advantage of “left” and “and” operations of the representation. Experi- ments show that the proposed algorithm only scans the music query stream once, and runs significant faster and consumes less memory than existing algorithms, such as SWFI-stream and Moment. The proposed FTP-stream algorithm is an exact stream mining method. That means the FTP-stream algorithm can generate the set of frequent patterns over music query streams without any information loss. It is because that the proposed algo- rithm uses bit-sequence representation of chord-sets to record the exact frequency of each chord-set. Then, the algorithm constructs the set of frequent patterns by us- ing these bit-sequence representations of chord-sets. Consequently, generating exact results is one of the benefits of our proposed algorithm. After mining frequent temporal patterns from online music query streams, the next issue of this work is that how to use these frequent patterns to predict the future trend of online music styles and to personalize the Web services of online music downloading. Hence, we need new information, i.e., changes of patterns, to as- sist the domain experts to predict the Web user behaviors and personalize the Web services. The second research issue of this paper is change detection of frequent patterns across data streams. With data streams, people are more often interested in mining queries such like “Compared to the history, what are the distinct features of the cur- rent status?”, “What are the most popular melody structures in the last four hours?” and “What are the relatively stable factors over time?” To answer such queries, we have to examine the changes of streaming data to assist the domain experts to predict the future trend of popular online music styles [7, 13]. Therefore, a sim- ple single-pass algorithm, called MQS-change (changes of Music Query Streams), is proposed to detect the changes of frequent patterns across music query streams. Experiments show that the proposed MQS-change algorithm is an effective method to detect the changes of data streams efficiently. Based on our best knowledge, the proposed MSQ-change algorithm is the first stream mining algorithm for discover- ing the changes of frequent patterns over music query data streams. Furthermore, for answering such above example query “What are the most popular melody struc- tures in the last four hours?”, the definitions of MFI (maximal frequent itemset), MFS (maximal frequent item-string), ICI (increasing changed itemset) and ICS (in- creasing changed item-string) can be used as popular melody structures in this paper although there are many other definitions of most popular melody structures depend on domain knowledge of experts. Note that MFI, MFS, ICI, ICS are defined in con- cluding Section. Hence, if the sliding window can be modified to contain the melody sequences generated from last four hours, we can use the proposed MQS-change to mine the most popular melody structures from last four hours. 15 Pattern Discovery and Change Detection of Online Music Query Streams 331 Problem Definition of Pattern Discovery of Music Query Streams In this section, several features of music data are described and the problem definition of pattern discovery of music query streams is described. The basic ter- minologies on music used in this paper are referred to [9, 10, 14]. A chord is the sounding combination of three or more notes at the same time. A note is a single symbol on a musical score, indicating the pitch and duration of what is to be sung and played. A chord-set is a set of chords. Let « Dfi 1 ;i 2 ;:::; i n g be a set of chord-sets, called items for simplicity, where n is the total number of chord-sets used for pattern mining. An itemset is a subset of items, i.e., a set of chord-sets. A k-itemset is an itemset with k items, denoted as .x 1 ;x 2 ;:::; x k /, where k is the length of that itemset. For brevity, the commas are omitted. For example, a 3-itemset (a, b, c) is written as (abc), where a, b, c are chord-sets, and the length of (a, b, c)is3.Amelody sequence stream (MSS) is a sequence of incoming melody sequences, Œm 1 ;m 2 ;:::; m N /, where a melody sequence m i is an itemset and N is an unknown large number of melody sequences that will arrive. Note that, in the representation of Œm 1 ;m 2 ;:::; m N /, the symbol “[” is the starting point of incoming melody sequence of the data stream and the symbol “)” is the current point of the data stream. Hence, it means that m 1 is the first melody sequence and m N is latest incoming melody sequence of the data stream. The sequence of w recent melody sequences of MSS is called the sliding window (SW) of MSS, where w is the size of the SW. The support of an itemset X, denoted as sup.X/, is the number of melody sequences in SW containing X as a subset. An itemset X is a frequent temporal pattern (FTP), if and only if sup.X/ s  w, where s is a user-defined minimum support threshold in the range of [0, 1]. An itemset X is called infrequent temporal pattern (ITP), if and only if sup.X/ < sw. A frequent temporal pattern is called maximal frequent temporal pattern (MFTP) if and only if it is not a subset of any other frequent temporal patterns. Definition of Problem 1 Given a melody sequence stream MSS and the size of slid- ing window w, the problem of online mining of user-centered music query streams is to discover the set of frequent temporal patterns by one scan of the w recent melody sequences of MSS with an adjustable user-defined minimum support threshold s in the range of [0, 1]. Example 1. Let the first four melody sequences of a stream of melody sequences be <m 1 ;.acd/>;<m 2 ; .bce/>; <m 3 ;.abce/>, and <m 4 ;.be/>, where m 1 ; m 2 ;m 3 , and m 4 are the identifiers of melody sequences and a, b, c, d and e are the identifiers of chord-sets, i.e., item identifiers. Let the size w of sliding window be 3 and the user-specified minimum support threshold s be 0.5. The stream of first four melody sequences is composed of two sliding windows, i.e., SW 1 DŒm 1 ;m 2 ;m 3  and SW 2 DŒm 2 ;m 3 ;m 4 , where first window SW 1 contains the sequences m 1 ;m 2 and m 3 , and the second window SW 2 contains the sequences m 2 ;m 3 and m 4 . Thus, 332 H F. Li A Melody Sequence Stream FTPs of SW 1 FTPs of SW 2 <m 1 , (acd) > <m 2 , (bce) > <m 3 , (abce) > <m 4 , (be) > (a), (b), (c), (e) (ac), (bc), (be), (ce) (bce) (b), (c), (e) (bc), (be), (ce) (bce) A melody sequence stream is formed by melody sequences arriving in series Fig. 2 An example melody sequence stream and the frequent temporal patterns in two consecutive sliding windows SW 1 and SW 2 the number of melody sequences of FTPs must have at least two sequences (0.5 of w D3 is 1.5). The result of example 1 is shown in the right side of Fig. 2, and the explanation on how to get the FTPs in Fig. 2 is given in Section “The Proposed Algorithm FTP-stream”. In Fig. 2, the discovered FTPs in SW 1 are four 1-itemsets, f.a/; .b/; .c/; .e/g, four 2-itemsets, f.ac/; .bc/; .be/; .ce/g, and one 3-itemset, f(bce)g. The dis- covered FTPs in SW 2 are three 1-itemsets, f.b/; .c/; .e/g, three 2-itemsets, f.bc/; .be/; .ce/g, and one 3-itemset, f(bce)g. In this example, we can find that f.a/; .ac/g areFTPsinSW 1 , but are not FTPs in SW 2 . Mining of Frequent Temporal Patterns in Music Query Streams Data Processing: Bit-sequence Representation In the proposed algorithm, for each item X in the current sliding window, a bit- sequence with w bits, denoted as Bit.X/, is constructed. If an item X is in the i-th music sequence of current sliding window, the i-th bit of Bit.X/ is set to be 1; otherwise, it is set to be 0. The process is called bit-sequence transform. Example 2. Consider an example melody sequence stream in Fig. 2 and assume that sliding window is composed of three melody sequences. Five items (chord-sets), a, b, c, d, and e, are used in this example. The first window SW 1 consists of three con- secutive melody sequences: <m 1 ;.acd/>;<m 2 ;.bce/>, and <m 3 ;.abce/>. Because the item a appears in the 1st and 3rd melody sequences of SW 1 , the bit- sequence of a is 101, i.e., Bit.a/ D101. Finally, a set of bit-sequences of 1-itemsets, i.e., Bit.b/ D011; Bit.c/ D111; Bit.d/ D100 and Bit.e/ D011, is generated by using bit-sequence transform for each new item of SW 1 . [...]... candidates from 3-candidates to l-candidates, where l is the size of largest itemset It is because the set of frequent 2-itemsets can be determined by the 2C-lists and its bit-sequences Consequently, it overcomes the performance bottleneck of 2-candidate generation of the level-wise frequent pattern mining algorithms Example 5 Consider the bit-sequences and 2C-lists of items of SW2 in Fig 4, and let the minimum... value of “1” indicates, that the spectrum is completely noise-like and a value of “0” indicates, that the spectrum is sinusoidal-like Linear Predictive Coding One of the fundamentals for digital speech communication is the compression of the speech signal in order to reduce the amount of transmitted data and hence, to save bandwidth The digital speech compression often bases on the principle of Linear... dmr@idmt.fraunhofer.de B Furht (ed.), Handbook of Multimedia for Digital Entertainment and Arts, DOI 10.1007/978-0-387-89024-1 16, c Springer Science+Business Media, LLC 2009 349 350 K Brandenburg et al degree of complexity has been proposed [110], [43], [85] Many publications have addressed suitable modelling methods that represent the musical gist whilst keeping the description blurry enough to account for. .. performing bit-sequence transform of each incoming item Therefore, each item has its unique 2C-list Each entry in the 2C-list consists of two fields: X and sup.X /, where X is the item identifier of the item being inserted and sup.X / registers the number of sequences containing the item X The process of stream mining is composed of three phases: window initialization phase, window sliding phase, and. .. (KL)-Transform (see 16), which results in a linear combination of the orige inal feature vector elements, sorted by decreasing significance Thus, the number of elements in the feature vector can be limited, according to the requirements of the system Logan [70] compared the usage of DCT instead of the KL Transform in the area of music information retrieval and came to the conclusion, that a DCT is adequate for. .. 12), MQS-change-PNSB (MQS-change for PNS and PNB, as shown in Fig 13), MQS-change-ICIS (MQS-change for ICI and ICS, as shown in Fig 14), and MQS-change-DCIS (MQS-change for DCI and DCS, as shown in 15 Pattern Discovery and Change Detection of Online Music Query Streams 343 − − Fig 13 Procedure MQS-change-PNSB of MQS-change algorithm − − − − Fig 14 Procedure MQS-change-ICIS of MQS-change algorithm Fig 15)... Generation Phase of FTP-stream The frequent temporal pattern phase is performed only when the up-to-date set of FTPs is requested In this phase, the FTP-stream algorithm uses a level-wise Fig 4 Bit-sequences and 2C-lists of items after sliding SW1 to SW2 15 Pattern Discovery and Change Detection of Online Music Query Streams 335 method to generate the set of candidate temporal patterns CTPk (candidate temporal... consists of two temporal lists, MFI-list and MFS-list, where MFI-list is a list of entries which contains current maximal frequent itemsets, and MFS-list is a list of entries which maintains maximal frequent item-strings so far Each entry of MFI-list consists of two fields: pattern-id Y and support-list Y.support-list, where pattern-id is a unique identifier of this maximal frequent itemset, and support-list... CTP-Gen-W2C (Candidate Temporal Pattern Generation Without 2-Candidates) Then, FTP-stream uses the bitwise AND operation to compute the supports of these candidates in order to find the frequent ones FTPk The generation-then-test process is stopped until no new candidates with k C 1 items CTPkC1 / are generated One of the benefits of the proposed algorithm is that the FTP-stream algorithm generates candidates... composed of a list of sup.Y /; i /, where i is the window identifier of window wi containing the itemset e 15 Pattern Discovery and Change Detection of Online Music Query Streams 341 For example, an entry of MFIlist indicates that the itemset abcd is a maximal frequent itemset and its estimated support is 30% in window w1 , 37% in w2 , 46% in w3 , and 70% . (ed.), Handbook of Multimedia for Digital Entertainment and Arts, DOI 10.1007/978-0-387-89024-1 15, c Springer Science+Business Media, LLC 2009 327 328 H F. Li Fig. 1 Computation model for music. m 1 ;m 2 , and m 3 . The bit-sequences of items and 2C-lists of sliding window SW 1 in the initialization phase of FTP-stream algorithm are shown in Fig. 3. Fig. 3 Bit-sequences and 2C-lists of items of. (a list of 2-C andidates), is constructed after performing bit-sequence transform of each incoming item. Therefore, each item has its unique 2C-list. Each entry in the 2C-list consists of two