The Research Of Data Classification Method Based On Metric Learning

Distance metric learning is an important task in visual classification.Learning an appropriate distance metric can greatly improve the performance of image recognition.In the field of biological information and text classification,multi-label learning is an important research topic.In contrast to single-label learning,multi-label learning needs to consider the internal relationship between multiple tags.Based on the distance metric,this paper constructs the regularized convex hull distance measurement learning model and the multi-label distance measurement learning model for the classification of image data and multi-label data.1.Based on the distance measurement learning model of convex hull,the regularized point-to-convex hull distance measurement(RPCHD)and the regularized convex hull to convex hull distance measurement(RCHCHD)are introduced in this chapter.The approximate optimization strategy is used to solve the coefficients in the distance measurement.Two distance measurement learning models similar to SVM are constructed in RPCHD and RCHCHD,which are transformed into standard SVM to learn distance measurement matrix M.The distance metric matrix is represented by positive and negative sample pairs.Experiments on three databases show that the proposed RPCHD and RCHCHD can effectively improve the performance of image recognition.2.A multi-label distance measurement learning model is proposed,which combines multi-label learning with measurement learning.The distance between multi-label in the sample is calculated to obtain a distance vector to represent the relationship between multi-label.The samples are sorted according to the value of the elements in the distance vector.Samples in the same order are considered to belong to similar classes,while samples in different orders are considered to belong to different classes.Mahalanobis distance was introduced to keep the order of label distance unchanged,and a multi-label distance metric learning model was built.Experiments on several datasets show that this method has considerable advantages over other advanced methods.