基于增量切空间校准的自适应流式大数据学习算法

发布时间：2018-08-28 14:34

【摘要】：流形学习是为了寻找高维空间中观测数据的低维嵌入.作为一种有效的非线性维数约减方法,流形学习被广泛应用于数据挖掘、模式识别等机器学习领域.然而,对于样本外点学习、增量学习和在线学习等流形学习方法,面对流式大数据的学习算法时间效率较低.为此提出了一种新的基于增量切空间的自适应流式大数据学习算法(self-adaptive streaming big data learning algorithm based on incremental tangent space alignment,SLITSA),该算法采用增量PCA的思想,增量地构造子空间,能在线或增量地检测数据流中的内在低维流形结构,在迭代过程中构建新的切空间进行调准,保证了算法的收敛性并降低了重构误差.通过人工数据集以及真实数据集上的实验表明:该算法分类精度和时间效率优于其他学习算法,可推广到在线或流式大数据的应用当中.
[Abstract]:Manifold learning is to find low dimensional embedding of observation data in high dimensional space. As an effective nonlinear dimension reduction method, manifold learning is widely used in data mining, pattern recognition and other machine learning fields. However, for manifold learning methods such as sample outer point learning, incremental learning and online learning, the time efficiency of flow big data learning algorithm is low. In this paper, a new adaptive flow big data learning algorithm (self-adaptive streaming big data learning algorithm based on incremental tangent space alignment,SLITSA) based on incremental tangent space is proposed. The algorithm uses the idea of incremental PCA to construct subspace incrementally. The intrinsic low-dimensional manifold structure in the data stream can be detected on-line or incrementally, and a new tangent space is constructed during the iteration process, which ensures the convergence of the algorithm and reduces the reconstruction error. Experiments on artificial data sets and real data sets show that the classification accuracy and time efficiency of the proposed algorithm are superior to those of other learning algorithms and can be extended to the online or streaming big data applications.
【作者单位】：南京师范大学计算机科学与技术学院;
【基金】：国家自然科学基金项目(41471371,61702270) 江苏省高校自然科学基金项目(15KJB520022)~~
【分类号】：TP181

【相似文献】