Sequence Classification and Melody Tracks Selection

This dissertation, "Sequence Classification and Melody Tracks Selection" by Fung, Michael, Tang, 鄧峰, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled Sequence Classification and Melody Tracks Selection submitted by Tang Fung Michael for the degree of Master of Philosophy at the University of Hong Kong in August 2001 One major approach to music retrieval is to model music as a sequence of features, after which traditional information retrieval techniques are applied on thesequence. Becauseofthetemporalnatureofmusicandtheinexactnessofuser queries, most efforts on music retrieval systems focus on issues such as indexing and approximation match. In contrast, the processing of music before feature extraction, such as the identification of melody track, were often considered easy or done. This may be the case in a controlled environment, such as one for musicology research, where the pieces are carefully analyzed by human beings before being submitted to the database. However, in an environment where large volumes of music are obtained from the Web, manual music analysis is impractical. Sincemanywell-knownmusicalfeaturesoftenpertaintothemelody ofmusicalpieces, andusersoftenrememberthemelodyofasong, algorithmsthat select the melody tracks of a piece are important for Web-based content-based retrieval systems. Such algorithms merely answer "yes" or "no" upon whether an input music track is melody. Hence melody tracks selection can be seen as a classification problem. Classification is a very important topic in the field of machine learning and data mining. Most of the existing classifiers, however, works on itemsets only. Sequences can also be transformed into itemsets, but information loss is inevitable. Hence classification algorithms which operate on sequences directly are important.Asaresult, anumberofalgorithmsformelodytracksselectionareproposed. Some of them exploit music knowledge on selecting melody tracks. In particular, "Range" works with the property of melodies and the interval range of human voices. "SilenceRatio" is based on an hypothesis about the amount of silence within a melody. "PMRatio" deals with the limit of monophony of human voice. "AvgVel" and "TrackName" make use of the features provided by MIDI file format. Other melody tracks selection algorithms presented in this thesis are se- quence classification algorithms. The music is first converted into feature se- quences. These sequences are then classified for identifying melody tracks. The sequence classification involved makes uses of emerging substrings. Emerging substrings in sequence database are similar to the emerging patterns in itemset database, which capture significant difference between two databases. Emerg- ing substrings mining algorithms are covered in this thesis. They make use of a datastructurecalledmerged suffix tree tostoreinformationofallpossibleemerg- ing substrings. The sequence classification algorithms developed shall work with other sequence data, including DNA, stock prices and text. Experiments were conducted on the classification power on the proposed algorithms, and their results are shown in the thesis. In brief, the results are a bit below satisfactory. This may be due to the insufficient use of music features and deficiency of the algorithms proposed. Further enhancement are possible on the construction of classifiers exploiting music knowledge and the sequence classifiers using emerging substrings. (481 words) DOI: 10.5353/th_b2974297 Subjects: Pattern recognition systemsMachine learning