Multi-label Classification Basedsemi-supervised And Localized Dimension Reduction

At present, a lot of valuable multi-label dataset is applied to all areas of life, however, it is very difficult to get the marked and significantive dataset, this will cost a lot of labour, material and financial resources. The proposal of semi-supervised learning or transductive learning is to utilize the marked and unmarked samples at the same time, Thus to study multi-label datasets. So the semi-supervised learning for multi-label data is the current hot issues in the field of machine learning and data mining.Existing related transductive learning for multi-label learning are one-way, that is learning is done from labeled to unlabeled samples or vice versa. Considering that one-way transductive learning does not take into account data’s full information, this paper proposes a new method of transductive learning-Two-way Markov Random Walk algorithm (TMRW), the algorithm uses information about labeled and unlabeled data to predict the labels of the unlabeled data by taking random walks between the labeled and unlabeled data. And then it use Adboost MH as the learning framework based on TMRW to improve the learning performance. Experiments show that it can achieve good performance than other classic algorithms.High-dimens ional multi-label data are easily availab le, but there is few good classification algorithm for high dimensional multi-label data. This paper proposes a local principal components analysis method (LPCA) for dimension reduction, dimension reduction was done in each cluster where the sample structures are consistent. Comparing the LPCA with PCA and MSDA, we use TMRW to classify the new dataset obtained by the dimension reduction. Experiments show that dimension reduction improves the high-dimensional multi-label data classification performance.This thesis provides an effective way to reduce the dimension of high-dimensional multi-label dataset and to construct a robust classifier with limited number of the labeled samples, and it also take Adboost.MH as the learning framework to improve the classification performance. It will have a certain reference value in multi-label classification.