Long Tail Multi-label Learning
Mengqi Yuan, Jinke Xu, Zhongnian Li · 2019
Multi-label learning is an activity research area that many methods arise recently to solve this problem. However, according to the results of current researches, the class imbalance which appears in the most of labels makes the network unable to be trained. In this paper, we propose a Long Tail Multi-label Classification Processing Algorithm (LTMCP) method to deal with this problem which leads training successfully and greatly reduces the parameters of the model. Firstly, we train model for class balance label that only used one network in view of that the performance of deep learning is better than linear networks. Secondly, we introduce a LTMCP method that trains linear network for each long tail label with unbalanced distribution. The final experimental results clearly show that this LTMCP method improves the accuracy while saving a lot of training time.