Data Balanced Bagging Ensemble of Convolutional- LSTM Neural Networks for Time Series Data Classification with an Imbalanced Dataset

Abstract

A system was developed using a bagging (bootstrap-aggregating) ensemble of neural networks to classify time-series data with class imbalanced datasets. The proposed system uses a Data Balanced Bagging Ensemble (DBBE) of Convolutional-LSTM (CLSTM) Neural Networks (DBBE- CLSTM) to classify accelerometer and EEG datasets. The base neural network (CLSTM) that is used in the DBBE-CLSTM contains three convolutional layers with a batch normalization layer following each convolutional layer, one max-pooling layer, one LSTM layer, and one sigmoid classification layer. A bagging ensemble was created which used the CLSTM paired with the proposed data balancing technique as the ensemble's base learner. The proposed bagging ensemble achieves a validation set average accuracy of 90.42% and a validation set recall of 93.23% on an accelerometer dataset. Ultimately, the proposed DBBE- CLSTM achieves the best overall performance of the models developed in this paper when evaluating both accuracy and recall, with the DBBE-CLSTM achieving greater than 90% for both metrics. As a secondary verification of the proposed DBBE- CLSTM, we evaluated the model on the UCI Epileptic Seizure Detection dataset and show that the model achieved high performance across all evaluated metrics. The DBBE-CLSTM achieved an average validation set accuracy of 99.23%, and average validation set f1-score of 0.9809 on UCI Seizure dataset.