HPC-enabled GA-SVM feature selection model for large-scale data

• 主要作者: Tay, Darwin Jia Xian.
• 其他作者: Stephen John Turner
• 格式: Final Year Project
• 語言: English
• 出版: 2009
• 主題: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• 在線閱讀: http://hdl.handle.net/10356/18903

With the explosive growth of data to be processed in multiple areas such as bioinformatics, scientific simulation and e-commence, data mining techniques are essential in making proactive, prudent and knowledge-driven decision. Support vector machine (SVM), pioneered by Vapnik has been chosen in this work as the data mining tool due to its excellent generalization performance. In particular, LibSVM has been selected as the software package to perform classification because of its sound performance and popularity. In this paper, an hybrid model for solving the problem of model selection associated with SVM is proposed. This model, HPC-enabled GA-SVM, takes advantage of genetic algorithm (GA) and high performance computing (HPC) technique like parallelism via OpenMP and MPI to conduct the process of model selection. GA was selected due to its capability of performing effective feature selection while HPC techniques have the capability of enhancing the computational performance. Exploration technique like ‘Uniform Design’ (UD) has also been employed to enhance the performance of the proposed model. A speedup of 29.02 times was achievable when compared to the traditional ‘grid’ search algorithm which is an exhaustive search approach without compromising much accuracy. Moreover, a caching policy known as “relaxed” caching policy has been proposed to avoid re-evaluations of previously evaluated combination that are in vicinity. This allows a speedup of 72.83 times when compared to the ‘grid’ search algorithm.