A Research On Emerging New Label And Incremental Learning In Mulit-label Data Stream

Multi-label learning usually studies the problem of one object with multiple la-bel simultaneously.In the previous studies,label space is generally considered to be fixed,but in practical applications,the data continuously grows in the open environ-ment over time,and the data will emerge new label.This is the problem of multi-label data stream.In order to maintain good predictive performance in the data stream envi-ronment,multi-label learning methods must be able to detect new label and correctly classify instances with new label.To solve this problem,this paper decomposes it into three sub-problems:new label detecting,classify for known labels datasets,and label-based incremental learning.The main work of this paper with regard to these three issues is as follows:(1)Based on the unsupervised characteristics and dynamic characteristics of the new label detect problem,the k-means clustering algorithm is applied to the data stream based on the anomaly detect idea.Besides,we propose a method named NLD-EC based on ensemble of k-means according to sampling of data and features.The experimental results show that the new label discovery problem is effective;(2)As for the classification problem of known labels datasets,based on the pair-wise label raning loss and misclassification loss,the penalty function in the form of harmonic series is introduced at the same time to construct the objective function,and the classification model can be obtained by working out the objective function based on a gradient descent method.The model can be updated well by new label data.(3)Concerning the label-based incremental learning problem,it is difficult for traditional algorithm to consider the relationship between the new label and the known labels.In addition,there are certain errors in the new label discovery model.There-fore,an iterative optimization strategy based on classification credibility adjustment is proposed with the help of the loss function in(2)mentioned above.Finally,a algo-rithm named L-ILPR is proposed to make the algorithm more stable in the data stream process.Experimental results show that the proposed algorithm has better performance and stability in multi-label data stream classification.