Adaptive Neural Decision Tree For Eeg Based Emotion Recognition

An adaptive neural decision tree is investigated to recognize electroencephalogram (EEG) emotion signal with ability of intelligently selecting network structure. Firstly, to overcome lack of position information of the vectorized input EEG signal, the original one-dimensional EEG vector signal is cast into a two-dimensional matrix signal added with channel position information. Secondly, the depthwise convolution is adopted in transformers to deal with each channel by one convolution kernel, thus decreases the number of parameters. Then, to overcome model interpretability problem, an adaptive neural decision tree (ANT) based emotion recognition method is explored by embedding neural networks into the decision tree to perform feature extracting, path selecting, and label classifying. Further, ANT can automatically search optimized parameters by using the adaptive moment estimation algorithm, and explore tree architectures by using the exploration–exploitation trade-off reinforcement learning method to get the global optimal network structure without manually setting complex hyperparameters. Finally, by 5-fold cross-validation, binary, four-class and eight-class classification experiments are carried out on DEAP datasets, the average accuracy of the ANT algorithm are 99.14 ± 0.456%, 98.95 ± 0.84%, 97.58 ± 2.311%, respectively, which verifies the effectiveness of the proposed method, compared with the traditional decision tree method.