Multi-label Green's Function Criterion inspired Transfer Annotation System

Abstract

Due to some datasets may acquire a large amount of information to train the annotation model, but some datasets may lack sufficient information, multi-label transfer learning has attracted a lot of attention recently. The correlation between inherent class label among multiple labels is integrated into data structure and can cover the relationship of samples. However, it is usually not accurate enough. In order to use their advantages and eliminate their weaknesses, this paper presents a new framework which integrates the techniques of Green's function inspired label correlation and label covariance analysis. Meanwhile, the framework adopts low-rank based method to learn the groups clearly. At the same time, due to the edge of annotation system be very difficult to determine, our framework is designed to carry out the appropriate hyperplane margin for multi-label datasets. In this framework, label covariance is first utilized by perceptron criterion analyzer, then linear discriminant function can dispose the unlabeled part of multi-label data easily. In addition, a low-rank based margin mechanism is presented to optimize the performance of our framework. We evaluate our framework in the performance of label correlation and covariance builders for transfer annotation. The derived results are encouraging and we achieve about 5% improvement on average accuracy.