Graph's Optimal Research In Graph-based Semi-supervised Learning

With the rapid development of computer technology, it is more convenient for people to collect data and information. Comparing with the growing amount of information, the capacity of information processing is far from people's satisfaction in their daily works and studies. As for the learning on large amounts of data, people almost remain in the method of traditional machine learning. To ensure the learning quality, traditional supervised learning requires a lot of manual intervention; thus the complicated feature and learning speed are naturally the soft spots. The unsupervised learning can improve the learning speed without manual intervention, but it can not ensure the learning quality because of being short of reliable monitoring information. As a new method of learning, semi-supervised learning is treated as a new hot spots in machine learning for it can not only use a small number of marked data but also enable the majority of unlabeled data involved in the learning process.With gradually deepening the study on this topic, Graph-based semi-supervised learning algorithm has made tremendous progress in the complexity, computational speed and accuracy. At the present time in which the two-class classification being a simple application and multi-label classification task becoming complicated, the graph-based semi-supervised learning algorithm is sure to receive more attention because of its superior performance. In view of multi-label classification, this paper mainly focuses on the graph-based semi-supervised learning, with emphasis on the key technologies of optimizing the classification effect.By deep analysis and researches with the existing methods, the following achievements are gained:1. On ML-GRF algorithm which is suitable for multi-label classification, a improved algorithm is proposed by using Spearman correlation coefficient correlation matrix to construct the label module. This algorithm can reduce the uncertainty of the provisional classification marking. The experimental result shows that the algorithm has an good stability of temporary classification marking and it can improve the accuracy of classification.2, On the importance of graph made in graph-based semi-supervised learning, the structure has been optimized. The weight matrix in the graph is analyzed and adjusted by block so as to get a better result and reduce the negative influences made by the unmarked data in semi-supervised learning process as well.