摘要
为解决当前的主动学习算法在采样时通常忽略样本特征表达信息的问题,提出一个基于样本特征表达和模型预测的主动学习模型。针对主动学习算法在模型训练早期阶段引起的冷启动问题,提出一个标注集初始化算法。利用聚类技术提取样本特征表达信息,通过分类器得到样本的模型预测信息,致力于使初始标注集的样本分布尽可能接近原始数据集的分布。实验结果表明,该模型分类准确率优于多个主动学习基线算法,该算法能够有效缓解模型的冷启动问题。
To address the problem of existing active learning algorithms that typically ignore sample feature representation information during sampling,an active learning model was proposed based on sample feature representation and model prediction.To alleviate the cold-start problem caused by active learning algorithms in the early stages of model training,a labeling set initialization algorithm was proposed.Clustering methods were utilized to extract sample feature representation information and the sample model prediction information was obtained through a classifier.The sample distribution of the initial labeling set was made as similar as possible to that of the original dataset.Experimental results demonstrate that the proposed active learning model outperforms multiple active learning baseline algorithms in classification accuracy,and the labeling set initialization algorithm effectively alleviates the cold-start problem.
作者
姜海涛
邱保志
李向丽
JIANG Hai-tao;QIU Bao-zhi;LI Xiang-li(School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,China)
出处
《计算机工程与设计》
北大核心
2024年第9期2757-2763,共7页
Computer Engineering and Design
基金
国家自然科学联合基金项目(U21B2037)
国家自然科学基金项目(62172371)。
关键词
主动学习
特征表达
模型预测
冷启动
聚类
图像分类
标注集初始化
active learning
feature representation
model prediction
cold start
clustering
image classification
labeling set initialization
