Please use this identifier to cite or link to this item:
Title: User Customization for Music Emotion Classification using Online Sequential Extreme Learning Machine
Authors: WONG, CHI MAN (黃志文)
Department: Department of Computer and Information Science
Faculty: Faculty of Science and Technology
Issue Date: 2015
Citation: WONG, C. M. (2015). User Customization for Music Emotion Classification using Online Sequential Extreme Learning Machine (Outstanding Academic Papers by Students (OAPS)). Retrieved from University of Macau, Outstanding Academic Papers by Students Repository.
Abstract: Machine learning techniques have been widely applied to handle complicated and advanced classification problem including classification of music emotion. In this work, traditional machine learning algorithms such as k-nearest neighbor, and state-of-the-art neural network methods such as support vector machine, and extreme learning machine are applied and compared with different combinations of feature sets extracted from a benchmark music set using MIRtoolbox for music emotion classification. The best classifier with the best feature sets combination is chosen for user customization. Because music emotion perception is subjective and can vary individual to individual, the model may not fit all users. To overcome this problem, we propose using online sequential extreme learning machine to update the model based on user’s preference because it is fast and accurate with good generalization ability. The result showed that with respect to user’s preference, the model can be updated immediately and remained a similar accuracy. Further more, it is also important to learn a good feature representation for music emotion classification task, so that we also investigate on the deep network such as Multi layer extreme learning machines (MLELM). MLELM inherits the fastness property of ELM while it can learn higher level of feature representation for classification. However, there are three problems for MLELM: 1) To determine the number of hidden neurons for each layer; 2) When the number of hidden neurons of ith hidden layer is different than the number of hidden neurons of (i+1)th hidden layer, the output from ith hidden needed to be scaled in order to fit in the (i+1)th hidden layer; 3) It assumes that the feature space for each hidden layer is the same. To solve these problems, we propose using a new kernel-based MLELM, to analytically determine the hidden neurons number for each layer. The result showed that the proposed kernel MLELM obtained a better performance than MLELM, SVM, and ELM.
Instructor: Prof. VONG, CHI MAN
Programme: Bachelor of Science in Computer Science
Appears in Collections:FST OAPS 2015

Files in This Item:
File Description SizeFormat 
OAPS_2015_FST_012.pdf2.59 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.