MgRCL: a multi-granular representation contrastive learning approach for hierarchical multi-label classification

Xiaoxia Zhang, Xin Xu, Xianjun Zhou, Hong Yu, Feng Hu

Abstract: Contrastive learning has been widely used in image tasks in recent years. With the increasing amount of large-scale image data, most of these data often have few or no labels, which is not conducive to the exploration and discovery of their information. Traditional supervised machine learning methods tend to improve the effectiveness of downstream tasks by increasing the correlation between features and label information, the process of these approaches often lack effective interpretation and with high label dependency. In recent years, unsupervised learning has improved its effectiveness and attracted much attention as free of label dependency. In this paper, we propose a mesoscopic contrastive learning method based on multi-granularity representation for hierarchical multi-label classification, short for MgRCL, wherein we obtain the pseudo-labels at each level by extracting features from the dataset and applying hierarchical clustering techniques. In the training phase of the model, data enhancement approach is employed to process the images through an encoder network, a projection layer and a softmax layer to fully integrate the hierarchical information that embedded in the pseudo-labels. Our model designs an improved loss function that combines the mesoscopic information, it lies between macro (or coarse-grained) information and micro (fine-grained) information, which provides an interpretable perspective on the learning process. The extensive experiments on CIFAR10, CIFAR100, STL10 and FMNIST show MgRCL achieves excellent results compared to other existing methods and can explain the classification process from micro, mesoscopic and macro perspectives. Our code in github:https://github.com/ZXXSTUDENTGITHUB/-/tree/zhouxianjun.

External IDs:dblp:journals/apin/ZhangXZYH25