Optimizing AI Models for Identifying Buying Intention Using Time-Frequency Domain EEG Features

Abstract

This study investigates the use of Electroencephalography (EEG) to classify consumer buying intention elicited by advertising stimuli. The main objective is to identify the most relevant EEG features and evaluate the effectiveness of various machine learning algorithms in predicting consumer intent. EEG signals were collected from 28 participants using six electrodes (Fp1, Fp2, F7, F8, O1, and O2) while they viewed video advertisements. The signals underwent preprocessing involving filtering, Independent Component Analysis (ICA), and amplitude clipping. Subsequently, wavelet-based segmentation was employed to extract alpha, beta, and gamma frequency components, from which 15 statistical features per channel were computed, grouped into (1) Time-Domain Metrics, (2) Entropy and Energy Features, and (3) Fractal Measures, yielding a total of 270 features. Dimensionality reduction was performed using three feature selection techniques: Pearson Correlation (PC), Linear Discriminant Analysis (LDA), and Mutual Information (MI). Each feature selection method was evaluated in combination with five classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), Naïve Bayes (NB), and Decision Tree (DT). The RF classifier consistently achieved the highest accuracy across all selection methods, peaking at 93%. Among the feature selectors, MI proved most efficient, achieving 90% accuracy with only 20 features. The most discriminative features were primarily derived from the gamma and beta bands in frontal and occipital regions, reflecting their role in attention and decision-making processes. Overall, the findings demonstrate the potential of EEG-based approaches for compact and accurate prediction of consumer intent, offering valuable insights for applications in neuromarketing and advertising analytics.