Neuromarketing is one of the most novel topics in neuroscience. This new topic is intended to detect, analyze and predict customer’s emotions in decision making towards a product. Traditional marketing methods such as designing questionnaires and polls are not sufficient and contain multiple errors due to the environmental, moral and temporal limitations. Neuromarketing has eliminated these obstacles and made it possible to access more precise data by enabling the analysis of human brain signals. One of the first steps towards the progression of this field, is the design and eva‎luation of detection and classification algorithms based on artificial intelligence. In the past decade, many studies in customer’s emotion classification have been employed based on artificial intelligence methods. These studies have been conducted in two major groups of machine learning and deep learning algorithms using extracted data from physiological signals. One of the most applicable approaches in emotion classification is the investigation of electroencephalogram signals. For instance, in one of the recent neuromarketing studies, the classification accuracy of machine learning algorithms has been investigated by analyzing EEG signals acquired from participants while watching pictures of different sport shoes. In yet another research, the linear and non-linear EEG features in decision making have been considered and the level of brain activity in different regions has been investigated for men and women groups. Also, some of these studies have focused on the performance of deep learning algorithms compared to machine learning algorithms. The results have encouraged researchers to more utilize deep learning models. In this study, the performance of AI-based algorithms in like/dislike classification will be assessed. To this end, SVM, KNN and RF classifiers along with the two proposed architectures, namely fully connected and convolutional networks, will be investigated. The input data for these algorithms are extracted from EEG signals in the public DEAP dataset in the form of features based on power spectral density. The overall results report that KNN classifier has the best performance among machine learning classifiers with the accuracy of 85%. Also, the classification accuracies of the two proposed neural networks are reported about 84% for each. Comparing the results with previous studies indicate that investigating the frontal and pre-frontal cortices lead to favorable results in customer’s emotion classification. Moreover, classifiers based on deep learning are reportedly competing well with conventional machine learning classifiers. Further through this study, the generalizability of the classifiers is eva‎luated using the local IUTED dataset. This dataset has been gathered at the biomedical laboratory of Isfahan University of Technology and is considered a simpler, compact version of public DEAP dataset. The classification accuracies of previously employed classifiers are then calculated after the required features have been extracted from the dataset. The final results indicate that despite the lower quality of the local dataset compared to DEAP dataset, the fully connected and convolutional neural networks have presented favorable classification accuracies. This is an indication of the generalizability of the proposed models.

رسول امير فتاحي ورنوسفادراني

فرزانه شايق بروجني , جلال ذهبي