Research On Manifold Alignment Algorithm Based On Label Information

Data classification is a hot topic in the field of pattern recognition and data mining.With the popularization of the Internet,the classification data in practical application is not only large but also has a higher dimension,which caused a lot of inconvenience for the data analysis.How to extract effective features from these data is the most important issue we face.Manifold learning and metric learning are typical nonlinear feature extraction methods,which have been widely used in classification problems.However,traditional manifold learning and metric learning can only be used in a single field.Due to the rapid development of the web application field,the data updating speed is fast,and the new fields are endless.The training data in the new field is often limited and insufficient to meet the requirements of traditional classification methods for training data.Marking new data or retraining the model in the new field will bring huge costs.How to use the marked data in the related auxiliary domain to help the new field to learn an effective classification model is the problem that scholars have been trying to solve.Manifold alignment algorithm is a potential method to solve the above problems.The key idea of manifold alignment algorithm is to project different high-dimensional dataset to a common low-dimensional space,while matching the corresponding instances and preserving local geometry of each input dataset.It can achieve transfer learning.The traditional manifold alignment algorithm needs sufficient corresponding instances to mine the connections between different manifold sample points.However,for the data classification problem,the label information of the training sample is only given,and there is no corresponding point information between the data sets.How to use label information to mine the correlation between data in different fields is a key issue in applying manifold alignment to cross-domain data classification.Based on the classification problem,this paper uses the label information to mine the connections among different manifold sample points,and proposed a manifold alignment algorithm based on label information.In addition,manifold learning and metric learning are combined to improve the accuracy of data classification.Further,the metric learning is combined with the manifold alignment,and the metric learning is applied to the cross-domain data classification.The main research contents are as follows:1.The label space embedding of manifold alignment algorithm is proposed.The algorithm reconstructs the features of each sample point by the label information and local geometric structure of samples,and calculates the similarity of the sample points in the manifold with cosine similarity.Maintaining the local geometry of the target domain and similarity of manifolds,and project the source domain and the target domain simultaneously into a common label space.Sample classes are obtained without the use of classifiers while calculating the low-dimensional embedding of the unknown class samples of the target domain.Experiments on several datasets validate the effectiveness of label space embedding of manifold alignment algorithm.2.A semi-supervised local linear embedding algorithm based on neighborhood component analysis is proposed.The algorithm combines metric learning with manifold learning,makes full use of label information based on metric learning to learn new distance measurement,so can discover accurate local geometric structure of samples.Then maintaining sample local geometric structure,and project the samples into low-dimensional space.Experiments show that the algorithm can get a good dimension reduction effect,and the low-dimensional embedding result can be input into the classifier to get a better classification result.3.A semi-supervised manifold alignment algorithm based on neighborhood component analysis is proposed.The algorithm draws on the idea of neighborhood component analysis and manifold alignment.It extends the original neighborhood component analysis model on a single data set to two manifolds.Combining the labeled data of source domain and target domain,learning appropriate measure matrix for each domain respectively.The two manifolds are projected into the low-dimensional space which is more obviously differentiated by category.Experiments on some datasets demonstrate the effectiveness of this method in classification problems.