| 摘 要: 针对高维数据降维存在的计算复杂度高及方向不可控问题,提出一种基于正交约束与最小重构误差的对比学习降维方法(RoCRL)。在编码器中加入正交约束层,控制降维方向,降低特征之间相关性。定义最小重构误差对比损失,使低维嵌入保持与原始数据相似的分布结构。实验表明:RoCRL降维后数据的平均识别准确率最高可达99.67%,与7种基线模型相比提升了2.21%,实验时间最大缩短283.65s。RoCRL增强了上游降维任务提取信息的能力,从而提高了下游分类任务的效率和精度。 |
| 关键词: 特征降维 正交约束 重构误差 对比学习 自编码器 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(12171378);冶金工业过程系统科学湖北省重点实验室项目(Y202105) |
|
| Contrastive Learning-Based Dimensionality Reduction with Orthogonal Constraints and Minimal Reconstruction Error for Image Classification |
|
GONG Yicheng1,2, SUN Yingying1, WANG Xiaoyu1, LI Sunbowen1,2
|
(1.College of Science, Wuhan University of Science and Technology, Wuhan 430065, China;
2.Hubei Province Key Laboratory of System Science in Metallurgical Process, Wuhan 430081, China)
gongyicheng@wust.edu.cn; sunyingying082@163.com; xiaoyuwang@wust.edu.cn; bw1863@wust.edu.cn
|
| Abstract: To address the challenges of high computational complexity and uncontrollable projection directions in high-dimensional data dimensionality reduction, this paper proposes RoCRL—a contrastive learning-based
dimensionality reduction method incorporating orthogonal constraints and minimal reconstruction error. By integrating an orthogonal constraint layer into the encoder, RoCRL controls projection directions and reduces feature correlations.A minimal reconstruction error contrastive loss is defined to preserve distribution structures consistent with the original data in low-dimensional embeddings. Experiments demonstrate that RoCRL achieves a peak average recognition accuracy of 99.67% on reduced-dimensional data, surpassing 7 baseline models by 2.21% while reducing runtime by up to 283.65 seconds. RoCRL enhances information extraction capability in upstream dimensionality reduction tasks,thereby improving the efficiency and accuracy of downstream classification tasks. |
| Keywords: feature dimensionality reduction orthogonal constraint reconstruction error contrastive learning autoencoder |
