基于类别感知与重加权的多源域自适应算法

Category-aware and reweighting-based multi-source domain adaptation algorithm

多源域自适应是迁移学习中的一个重要分支,类别偏移是多源域自适应领域的热点难题之一,其本质是源域和目标域类别分布不匹配的问题。针对此问题,提出了一种基于类别感知与重加权的多源域自适应算法,该算法通过类别感知策略增强相似类别间的正向迁移;同时,引入重加权矩匹配策略,减少不同层面的分布差异;此外,利用伪标签构建自适应权重,有效降低类别偏移的影响。在Digits-five和Office-Caltech10两个数据集上的任务分类准确率分别达到了94.11%和97.18%,实验结果表明,所提算法相比于当前典型的多源域自适应算法在类别偏移场景下的准确性方面有显著提升。

Multi-source domain adaptation is an important branch of transfer learning.Category shift,a prominent challenge in this field,stems from the mismatch between category distributions in the source and target domains.To address this problem,a category-aware and reweighting-based multi-source domain adaptation algorithm is proposed.The algorithm enhances positive transfer between similar categories through a category-aware strategy and introduces a reweighting moment matching strategy to reduce distribution differences at various levels.Additionally,adaptive weights are constructed using pseudo-labels to effectively mitigate the impact of category shift.Experimental results on the Digits-Five and Office-Caltech10 datasets show that the proposed algorithm achieves classification accuracies of 94.11%and 97.18%,respectively.These results indicate that the proposed algorithm significantly improves accuracy in scenarios with category shift compared to current typical multi-source domain adaptation algorithms.