基于直推式学习的网络故障诊断算法

网络故障的及时诊断能够保证日常工作、学习和生活能够正常进行。传统的基于监督式学习的诊断方法依赖于大量具有鉴别意义的样本,这在实际情况中通常难以得到满足。针对上述问题,本文提出了一种基于直推式学习的诊断算法。针对大规模的网络管理的特征数据,本算法利用主成分分析对特征进行降维,并利用新的度量下的特征数据来构建拉普拉斯矩阵；该矩阵能够很好的描述带检测样本和训练样本之间的关系。在此基础上,本文设计了基于直推式学习的目标函数,并利用拉格朗日乘子法完成了优化。实验部分证明了本算法能够在有限数目的带标签的样本的前提下获得精确的分类结果,能够显著提高网络故障诊断的检测率。

Transductive learning-based network fault diagnosis

Daily work, learning and life can be guaranteed by in-time diagnosis of network fault. Traditional methods, based on supervised learning, rely on a large number of discriminative data, which is difficult to be met in practice. To deal with those problems, our algorithm utilizes the principle component analysis to reduce the dimension of the features, and then constructs the laplacian matrix based on the new features. The corresponding matrix can describe the relationship between data well. On the basis of that, we design the objective function of transductive learning and apply the lagrange multiplier to optimize that. Experimental results verify the high accuracy of our algorithm while dealing with low number of labeled data. Our algorithm can significantly improve the diagnosis of network fault.