A Peephole Bidirectional the Long Short-Term Memory Network for Multi-Scale Imbalanced Time Series Classification

For unbalanced multivariate time series with missing values, a multi-scale prediction model based on long short-term memory network variant (multi-PB-LSTM) is constructed. The multi-PB-LSTM utilizes multi-task Gaussian processes to transform asynchronous time series into unified and complete input data by learning the correlation between timestamps and features. At the same time, several classifiers based on LSTM variants are constructed aiming to capture dependencies over a long span in sequence data. Finally, through the decision fusion component of the multi-PB-LSTM model, the prediction results of each classifier are combined to obtain the final classification result. The proposed model can achieve satisfactory performance on some public datasets (MIMIC-III, PhysioNet).