Author(s)

Shengxuan Xu, Yuanyuan Tan and Kaiyu Zhang

Corresponding Author

Shengxuan Xu

ABSTRACT

Due to the excellent performance of multi label learning in the process of solving the problem that a single object may have multiple associated categories, the researchers have attracted extensive attention. In this paper, according to the information and needs of different users, we use data mining to analyze the correlation between many label features, and use the sample clustering information to adjust the similarity matrix of weak label samples. We propose an evaluation standard classifier reliability to measure the classifier, and construct the similarity matrix based on k-means. In addition, we introduce different levels of multi instance learning algorithm as our classifier model, and propose two kinds of classification distance: the minimum classification distance and the average classification distance. Finally, we apply the model to the natural scene image classification and text classification, and compare the method proposed in this paper with the common samples. It is found that under the same training samples, the method in this paper can achieve better classification performance in various evaluation indexes.

KEYWORDS

Label probability matrix; multi example multi label learning; weak label; multi label learning