| Summary: | 碩士 === 國立臺北科技大學 === 電機工程系所 === 96 === High dimensions and huge volumes of datasets result in large loadings of feature selections and classififications. In order to overcome this drawback, we propose a class-based parallel mechanism to speedup the classifiations of high dimensional datasets.
In this paper, the proposed parallel method can be divided into three parts: parallel greedy modular eigenspace (PGME), feature scale uniformity transformation (FSUT) and parallel positive Boolean function (PPBF). First, the most significant features of each class are acquired by PGME/FUST from original features. A parallel classifier of optimal positive Boolean function is then trained from these extracted features. Finally, a classification process can be deployed to achieve a high performance computation for high dimemsional datasets in parallel.
Three parallel schemes are demonstrated, namely a share memory multi-processors (SMP), a cluster, and a hybrid parallel architecture (the combination of both cluster and SMP systems), to validate the flexibility of implementation of the proposed scheme. We estimate the speedup by different numbers of classes, dimensions and samples. The experimental results show that our class-based parallel technique can effectively enhance the performance for high volumes of datasets.
|
|---|
