多标记数据驱动的可变换算子值核OA

An operator-valued kernel is a binary function that takes the value of an operator on Hilbert space,which in the field of machine learning aims to better describe the correlation between different tasks in multi-task learning.Multi-label learning is a special kind of multi-task learning,in this paper,we learn operator-valued kernels from multi-label datasets based on the kernel alignment method and construct a prediction model for multi-label learning.Firstly,we use kernel alignment method to learn the instance-level feature importance distribution;secondly,we construct operator-val-ued kernel based on the instance-level feature importance distribution,and prove that it is not only partial trace kernel but also transformable operator-valued kernel,and that each block of its corresponding kernel matrix depicts the interac-tion information of label correlation among the samples;lastly,we design the multi-label learning algorithms based on transformable operator-valued kernel,and conduct comparative experiments with four high-performance algorithms on nine multi-label datasets,the results verify effectiveness of our proposed algorithm.