半监督学习研究的述评

监督学习需要利用大量的标记样本训练模型，但实际应用中，标记样本的采集费时费力。无监督学习不使用先验信息，但模型准确性难以保证。半监督学习突破了传统方法只考虑一种样本类型的局限，能够挖掘大量无标签数据隐藏的信息，辅助少量的标记样本进行训练，成为机器学习的研究热点。通过对半监督学习研究的总趋势以及具体研究内容进行详细的梳理与总结，分别从半监督聚类、分类、回归与降维以及非平衡数据分类和减少噪声数据共六个方面进行综述，发现半监督方法众多，但存在以下不足：（1）部分新提出的方法虽然有效，但仅通过特定数据集进行了实证，缺少一定的理论证明；（2）复杂数据下构建的半监督模型参数较多，结果不稳定且缺乏参数选取的指导经验；（3）监督信息多采用样本标签或成对约束形式，对混合约束的半监督学习需要进一步研究；（4）对半监督回归的研究匮乏，对如何利用连续变量的监督信息研究甚少。

Review of Semi-Supervised Learning Research

Traditional supervised learning methods require a lot of labeled samples to accomplish training tasks,but in practical application,the collection of labeled samples is difficult.Although unsupervised learning methods do not require prior information,it is difficult to guarantee the accuracy.Semi-supervised learning breaks through the limitation of traditional methods that only consider labeled samples or unlabeled samples,and can mine a large amount of information hidden in unlabeled data and assist a small number of labeled samples for training,becoming a research hotspot of machine learning.This paper summarizes the general trend and detailed research contents of domestic semi-supervised learning,and summarizes six aspects including semi-supervised clustering,classification,regression and dimension reduction,unbalanced data classification and noise reduction.There are many semi-supervision methods,but there are some shortcomings:(1)Although some of the new and effective methods are proposed,they are only verified through specific data sets and lack theoretical basis and proof.(2)When the data is complex,the semi-supervised model needs many parameters,but lacks the experience of parameter selection.(3)The supervision information is mostly in the form of sample labels or pair constraints,and the semi-supervised learning of mixed constraints needs further study.(4)There is a lack of research on semi-supervised regression and the prior information of continuous variables.