Missing data imputation and classification of small sample missing time series data based on gradient penalized adversarial multi-task learning.

In practice, time series data obtained is usually small and missing, which poses a great challenge to data analysis in different domains, such as increasing the bias of model predictions, reducing the accuracy of model classification, and affecting the analysis data. This paper aims to address the problem of missing data imputation and classification of small sample time series data. By exploring and implementing efficient data interpolation strategies to improve classification accuracy, the robustness and accuracy of classification models in the face of incomplete data. To achieve this, we propose a new model that can effectively classify time series data with missing values. Our model utilizes a bi-directional long short-term memory network combined with an extreme learning machine for the imputation task, which can recover the missing time series values. For the classification task, we employ a self-attentional Inception Time network, which is regularized by a classification loss to effectively mitigate network overfitting. To improve the performance of the model on small sample time series datasets, we use a gradient penalty adversarial training approach. Our model integrates the advantages of multiple network modules, the gradient penalty adversarial multi-task model achieves optimal imputation and robust classification of missing small sample time series data. To evaluate the overall performance of our model, we selected forty datasets from the UCR time series datasets, and selected the German emotional speech datasets and the EEG epilepsy datasets, with the plant electrical signal datasets obtained from real measurements. A series of experiments were conducted to evaluate the effectiveness of our method compared to other methods, the datasets were set up with multiple missing rates, with root mean square error and coefficient of determination to assess the accuracy of imputation, and with accuracy to assess the performance of the classification task. The results show that our proposed method outperforms existing methods in terms of imputation accuracy and classification performance. To better understand the deep learning model, we used the Grad-CAM + + method to enhance the reliability and credibility of the model by visualizing the important features of the temporal data when the plant electrical signal datasets was tested. In summary, this paper presents a model framework for the imputation and classification of missing small sample time series data, and the experimental results show that our model provides an effective solution for dealing with the analysis of missing small sample time series data. [ABSTRACT FROM AUTHOR]