Domain self-adaptive dimensionality reduction method through keeping maximum dependency relationship between data conversion

Abstract

The invention relates to the domain self-adaptive correlation problem in the machine learning field and provides a domain self-adaptive dimensionality reduction method through keeping a maximum dependency relationshipship between data conversion. In order to reduce the distribution difference between a source domain and a target domain, the two-time regeneration nuclear hilbert space mapping is carried out on data, wherein two hilbert spaces are respectively marked as H1 and H2. The H2 is called as a common potential space between the source domain and the target domain. In order to conveniently measure the difference of edge distribution between the source region and the target region, the measurement is achieved by using a maximum mean value difference (MMD) method. The H2 is obtained through the mapping of the H1. In order to obtain the correlationship of the data mapping measurement between the space H1 and the space H2, the hilbert-schmidt independent criterion (HIC) method is used for measurement. The method mainly aims to minimize the distribution difference of the source domain and the target domain in the H2, namely the MMD value is minimized.The dependency relationshipship between data conversion in the space H1 and the space H is maximized. That is, the HIC value is maximized.

Description

technical field [0001] The invention relates to a domain adaptive technology oriented to the field of machine learning, in particular to a domain adaptive dimensionality reduction learning method. Background technique [0002] With the development of science and technology, the data information that people deal with is becoming more and more complex and huge. These data often have a high dimensionality, and these data generally have a lot of redundant information. Feature extraction is necessary. [0003] In the past machine learning research, there have been many studies on dimensionality reduction methods for feature extraction. The most typical linear dimensionality reduction methods include PCA and LDA (Document 1, Keinosuke Fukunaga. Introduction to statistical pattern recognition. Academic Press, 1972.). Typical popular learning nonlinear dimensionality reduction methods include ISOMAP (document 2, Tenenbaum JB, Silva VD, Langford JC (2000) A Global Geometric Framewo...

AI Technical Summary

Problems solved by technology

The method of self-labeling target domain instances is also not suitable for high divergence domains

Images