Research On Semi-supervised Classification Algorithm Based On Temporal Relationship Learning

Time series data processing is widely used in various fields in real life such as medical-aided diagnosis and machine part life span prediction.And time-series data classification is one of the important methods for time-series data processing.Traditional time-series data classification methods based on supervised learning have made progress with the help of large amount of labeled data.Compared with image data,the nonintuitive nature of time-series data leads to the fact that the labeling of time-series data usually requires a large amount of expert knowledge,which is sometimes not even accessible.Therefore,Semi-Supervised Learning(SSL)methods that combine large amounts of easily accessible unlabeled data with small amounts of labeled data have gained the attention of researchers.However,for time series data,the existing semi-supervised time series classification algorithms are either based on similarity measurement or model-based methods.The method based on similarity measurement is mainly classified according to the similarity between unlabeled data and labeled data.The classification accuracy has a strong correlation with the proportion of labeled data,and it is not easy to design a reasonable similarity measure.Model-based methods also have the problem of dependence on the proportion of labeled data.Different from other data types,time series data has an underlying temporal structure.However,the current semi-supervised learning models seldom pay attention to the potential timing structure within time series,which leads to a large number of unlabeled time series can not be fully utilized in model training.In this paper,by studying the internal timing relationships of unlabeled time series,we design the irregular time sampling algorithm and propose a semi-supervised time series classification framework,which can learn unlabeled time series in a self-supervised way through the potential timing structure generated by the timing data,so as to improve the classification accuracy of semi-supervised time series.Specifically,we propose four different irregular time sampling functions to sample the original time series into different sampling time series.Then,the model uses supervised learning module to classify the labeled time series directly,and predicts the irregular sampling function types of the sampling time series in the self-supervised learning module.Finally,the self-supervised learning module captures the underlying timing structure within the unlabeled time series.By jointly training supervised and self-supervised learning modules,the feature space can be made consistent between labeled and unlabeled data,thus improving the ability of model learning and the quality of feature representation.Extensive comparison experiments with seven benchmark models on multiple real-world datasets demonstrate the effectiveness of the proposed approach in this paper.