Towards An Unified Replication Repository for EEG-based Emotion Classification

EEG-based emotion classification is characterized by a sequence of steps, where prepossessing and feature engineering have a preponderant role, given the richness and complexity of the data. This has led to a high number proposed methodologies in recent years. Although positive, this makes it difficult to visualize and compare the state of the art, as each method usually uses its own dataset and evaluation metric and does not provide enough detail (or source code) for replication. In an attempt to formalize the analysis and evaluation methodologies, this paper presents a initial step towards a unified analysis framework. We performed a replication study on two of the most representative approaches for classification and evaluate them using a unique metric on the DEAP dataset. Additionally, as feature extraction represents one of the most heterogeneous and also laborious steps, we propose an alternative based the use of a Convolutional Neural Network (CNN), with the aim to study if a setting where the model automatically learn to represent the data can be competitive in comparison with manually crafted features. Results of our empirical study show that the replication and comparison is feasible and that CNN based approaches are competitive to traditional signal-based classification.