Research Of Semi-supervised Learning Method Based On Attribute Partial Order Structure Theory

With the advent of data explosion, how to extract useful information from the mass of data and make better use of unlabeled data is becoming more and more important. However, the existing semi-supervised learning methods require that data distribution must be consistent with the hypothesis model and adding unlabeled data can lead to reduce the generalization ability. In this paper, attribute partial order structure theory is applied to semi-supervised learning to improve the generalization ability. Attribute partial order structure theory is not only able to learn a variety of models without help of prior knowledge, but be able to analyze attribute partial order structure with the analysis models of logical concept to mine the connection among data sets.Semi-supervised learning method is based on attribute partial order structure theory, based on formal concept analysis and mathematical partial theory as the theoretical basis, data discretization as the drive, formal context of data as the project, attribute coverage maximum project generating attribute partial order structure diagram as the research subject. This method expresses attribute structure relation of discrete data, finds the distribution rules of data attribute and achieves a similar data attribute match.Firstly, this paper elaborates various of definitions and theories for formal concept analysis and attribute partial order structure theory. Comparing with concept lattice, the advantages of attribute partial order structure diagram which can be express more clearly and has simpler and clearer hierarchy and clustering relationships is found. Then this process includes: data discretization by splitting the data range, formal context generation by transforming data attribute into several regions, attribute rules discovery of formal context for label data and matching rules discovery of unlabeled data for semi-supervised learning, unlabeled data tag, the matching model generation of data attribute, feasibility and effectiveness verify of this method and so on. It is important to the development of machine learning theory based on semi-supervised learning.Finally, in the chapter of software analysis and evaluation, the overall architecture of software design has been given out firstly, then the function of each part and the workflow of this software has been described in detail. At the same time, this software has been combined with the attribute partial order structure diagram to show the match of unlabeled data, and gives the process of Iris data set and Wine data set in detail. In the end, this classification results are compared with other semi-supervised learning classifiers.