# A Semantics-Guided Class Imbalance Learning Model for Zero-Shot Classification

**Authors:** Zhong Ji, Xuejie Yu, Yunlong Yu, Yanwei Pang, Zhongfei Zhang  
**arXiv:** [1908.09745](https://arxiv.org/abs/1908.09745)  
**Date:** 2019-08-26  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.09745v1/

## Key points

- The paper addresses the issue of class imbalance in Zero-Shot Classification (ZSC)
- ZSC is prone to class-imbalance issues as it prioritizes class-level knowledge transfer
- Class samples follow a long-tailed distribution in real-world scenarios, making it difficult to transfer discriminative information from sample-scarce seen classes to related unseen classes using traditional batch-based training methods
- The authors propose a sample-balanced training process to mitigate the class imbalance problem in ZSC
- They achieve this by randomly selecting an equal number of images from each class across all training classes to form a training batch
- An efficient semantics-guided feature fusion model is developed, which assigns different weights to selected samples based on their representativeness
- The proposed approach is evaluated through extensive experiments on three imbalanced ZSC benchmark datasets for both Traditional ZSC (TZSC) and Generalized ZSC (GZSC) tasks
- The results demonstrate promising outcomes, particularly for unseen categories closely related to sample-scarce seen categories
- Overall, the paper presents a refined methodology for addressing the class imbalance issue in Zero-Shot Classification by introducing a sample-balanced training process and leveraging semantics-guided feature fusion.

## Summary

The paper "A Semantics-Guided Class Imbalance Learning Model for Zero-Shot Classification" addresses the issue of class imbalance in Zero-Shot Classification (ZSC). ZSC focuses on recognizing visual instances from novel classes by establishing interactions between visual and semantic modalities. However, ZSC is prone to class-imbalance issues as it prioritizes class-level knowledge transfer. In real-world scenarios, class samples follow a long-tailed distribution, making it difficult to transfer discriminative information from sample-scarce seen classes to related unseen classes using traditional batch-based training methods. This limitation significantly hampers the overall generalization ability of ZSC. To mitigate the class imbalance problem in ZSC, the authors propose a sample-balanced training process that ensures all training classes contribute equally to the learned model. They achieve this by randomly selecting an equal number of images from each class across all training classes to form a training batch. This approach guarantees that sample-scarce classes contribute as much as those with sufficient samples during each iteration. Recognizing that instances within the same class may differ in their representativeness, the authors further develop an efficient semantics-guided feature fusion model. This model obtains discriminative class visual prototypes for subsequent visual-semantic interaction processes by assigning different weights to selected samples based on their representativeness. The proposed approach is evaluated through extensive experiments on three imbalanced ZSC benchmark datasets for both Traditional ZSC (TZSC) and Generalized ZSC (GZSC) tasks. The results demonstrate promising outcomes, particularly for unseen categories closely related to sample-scarce seen categories. Overall, this paper presents a refined methodology for addressing the class imbalance issue in Zero-Shot Classification by introducing a sample-balanced training process and leveraging semantics-guided feature fusion. The experimental results highlight its effectiveness in improving generalization performance, especially for unseen categories linked to sample-scarce seen categories.