Abstract:

In response to the problem that for dimensionality reduction, traditional manifold learning algorithm did not consider the raw data category information, and the degree of clustering was generally at a low level, a manifold learning dimensionality reduction algorithm with supervised discriminant projection (SDP) was proposed to improve the dimensionality reduction effects of network security data.On the basis of the nearest neighbor matrix, the label information of the raw data category was exploited to construct a supervised discriminant matrix in order to translate unsupervised popular learning into supervised learning.The target was to find a low dimensional projective space with both maximum global divergence matrix and minimum local divergence matrix, ensuring that the same kind of data was concentrated and heterogeneous data was scattered after dimensionality reduction projection.The experimental results show that the SDP algorithm, compared with the traditional dimensionality reduction algorithms, can effectively remove redundant data with low time complexity.Meanwhile the data after dimensionality reduction is more concentrated, and the heterogeneous samples are more dispersed, suitable for the actual network security data analysis model.

Key words: data dimension reduction, manifold learning, supervised learning, discriminant projection