基于伪标签的二阶段时序半监督学习框架OA

In view of the problem of scarce labeled data in some scenarios of time series classification,this study proposes a two-stage time series semi-supervised learning framework based on pseudo-labels.In the first stage,contrastive learning is used for training to construct a base classification model and label the unlabeled data.In the second stage,appropriate pseudo-labeling techniques are employed to retrain the model,so as to make full use of the close association between labeled data and unlabeled data to improve the model performance.Experiments are conducted on multiple public time series classification datasets to verify the effectiveness of the proposed framework,and an in-depth discussion is carried out on the applicable conditions of different pseudo-label training methods in the second stage.The experimental results show that when the proportion of labeled data is only 1%and 5%,the proposed learning framework can increase the average accuracy by approximately 5.1%and 3.5%respectively on two base models and multiple datasets.This fully demonstrates that the proposed method can effectively solve the problem of semi-supervised time series classification.