# Decoding Non-Invasive Electroencephalography Signal via a Two-Discriminator Adversarial Network

## TL;DR

This paper introduces a new neural network method to improve emotion recognition from EEG signals by addressing individual differences and channel relationships.

## Contribution

A two-discriminator adversarial network (TD-DANN) is proposed to enhance emotion decoding from non-invasive EEG signals.

## Key findings

- TD-DANN achieved 98.45% accuracy for subject-dependent emotion recognition on the SEED dataset.
- The method showed 89.45% accuracy for subject-independent experiments on the same dataset.
- TD-DANN also attained 84.40% and 77.13% accuracy on the SEED-IV dataset for subject-dependent and subject-independent tasks, respectively.

## Abstract

Electroencephalography (EEG), as a typical non-invasive biosensing signal, reflects individual emotional changes by recording the brain’s neural activity in response to various external stimuli. However, the significant differences in brain activity among individuals and the complex interrelationships between EEG channels notably hinder the accuracy of emotion decoding in non-invasive biosensing scenarios. To address this challenge, this paper proposes a two-discriminator domain adversarial neural network method (TD-DANN). The proposed method aims to obtain more generalized and individualized emotion feature representations through adversarial learning. Specifically, graph convolution is utilized to extract features from EEG signals. By modeling the EEG channels as graph nodes, the adjacency matrix can be dynamically learned to capture the complex relationships between different channels during emotion generation. Moreover, we design a domain discriminator and an individual discriminator. The domain discriminator is used to minimize the difference in feature distribution between the source and target domains. It is able to obtain discriminative features with universality. The individual discriminator is used to learn discriminative features consistent with the individual’s brain activity. It can enhance the adaptability to the individual’s emotion. The experimental results show that the TD-DANN achieves promising recognition accuracies of (98.45 ± 2.38)% and (89.45 ± 5.87)% for subject-dependent and subject-independent experiments on the SEED dataset, respectively. The proposed method attains recognition accuracies of (84.40 ± 8.70)% and (77.13 ± 7.97)% for subject-dependent and subject-independent experiments on the SEED-IV dataset, respectively. These results validate the effectiveness of the TD-DANN in the emotion decoding problem.