Understanding the factors influencing consumer purchase decisions is crucial in consumer behavior and advertising research. Traditional methods like questionnaires have limitations, failing to capture the nuances of consumer experience and potentially influenced by social desirability bias and the mere measurement effect. This study addresses these limitations by integrating electroencephalography (EEG), which provides direct insights into neural activity, with machine learning, which can handle complex datasets. Neuromarketing, the intersection of neuroscience and marketing, has shown promise in predicting consumer behavior with higher accuracy than traditional methods. However, many neuromarketing studies use laboratory settings and shopping simulations, lacking the ecological validity of real-world shopping experiences. This study uses a field experiment to collect real-time EEG data from consumers making actual online purchases, aiming to predict purchase decisions with high accuracy and ecological validity. The study enriches neuromarketing theory and offers practical implications for improving marketing strategies.

The study draws upon existing research in EEG-based emotion recognition and machine learning. EEG signals, reflecting brain electrical activity, are analyzed using machine learning algorithms (supervised and unsupervised) to identify emotional states. Supervised learning methods, like SVM, RF, KNN, and neural networks, use labeled data for training, while unsupervised methods, such as K-means clustering, are used for pattern detection in unlabeled data. Common EEG features include power spectral density (PSD), reflecting power distribution across brain frequencies (delta, theta, alpha, beta, gamma), and prefrontal asymmetry index (PAI), reflecting differential activation between left and right frontal areas. PSD is linked to various cognitive and emotional states, and PAI is associated with approach/withdrawal tendencies. Previous studies have demonstrated the effectiveness of using these features in emotion recognition and consumer behavior research, but often in controlled lab settings.

This study employed a field experiment involving 66 undergraduate students (after excluding participants based on financial stability), all right-handed with no neurological disorders and similar spending habits. Participants used two laptops; one for online shopping and the other for EEG data monitoring. They made purchase decisions for five items in their online shopping cart, purchasing at least one item and not purchasing at least one. The Emotiv EPOC+ EEG device with 14 electrodes recorded EEG data at 256 Hz during the decision-making process. Data preprocessing involved filtering, segmentation, artifact removal (using ICA), and marker reading. Two key EEG features were extracted: PSD (across delta to gamma frequency bands from 14 channels) and PAI (from three pairs of frontal electrodes, also across delta to gamma bands), resulting in 119 features. Traditional machine learning classifiers (RF, KNN, SVM with linear and Gaussian kernels) and a shallow neural network were used. Feature selection methods (correlation, T-value, F-score, PCA, RFE) were applied to optimize model performance. Leave-one-out cross-validation (LOOCV) and 1000-permutation tests were used to assess the statistical significance of the results.

The SVM classifier with a Gaussian kernel demonstrated the highest accuracy (87.1%), significantly outperforming other classifiers. The Gaussian kernel's ability to model non-linear relationships in EEG data likely contributed to its superior performance. KNN classifiers showed high accuracy in specific configurations, but only some passed the permutation tests. Random forest classifiers showed moderate performance and passed the permutation tests. The shallow neural network achieved a mean accuracy of 66.84%, showing moderate performance. Feature selection consistently highlighted the importance of frontal (PAI) and occipital (PSD) regions in predicting purchase decisions. Analysis of top EEG features across different methods revealed consistent involvement of the frontal and occipital regions, with PAI highlighting the prefrontal cortex's role in processing cognitive and emotional information, and occipital PSD reflecting visual processing and attention. The features identified spanned various frequency bands (delta, theta, alpha, beta, gamma).

The high accuracy achieved by the SVM with a Gaussian kernel, particularly when combined with RFE for feature selection, demonstrates the potential of using EEG and machine learning to predict consumer purchase decisions. The superior performance of the Gaussian kernel over the linear kernel underscores the non-linear nature of neural processes underlying decision-making. The consistent importance of frontal and occipital regions across different feature selection methods highlights the central roles of cognitive and emotional processing, as well as visual information processing in purchase decisions. The findings challenge the limitations of traditional self-report methods, suggesting EEG data provides a more accurate measure of implicit preferences. The moderate performance of the shallow neural network indicates the need for larger datasets and potentially more advanced neural network architectures in future research. The results support the idea that a comprehensive feature set, integrating both PAI and PSD, enhances the predictive accuracy of machine learning models.

This study demonstrates the effectiveness of using EEG data and machine learning algorithms, specifically SVM with a Gaussian kernel, to predict consumer purchase decisions in online shopping environments. The findings highlight the importance of frontal and occipital brain regions in the decision-making process. Future research could explore the use of higher-density EEG systems, deep learning models with larger datasets, and the inclusion of other relevant factors like brand appeal to further improve predictive accuracy and deepen our understanding of consumer neuroscience.

The study's limitations include the use of a relatively small dataset, a limited number of EEG electrodes (14), and the exclusion of advanced EEG features (brain networks, microstates, nonlinear dynamics) due to data processing limitations. The absence of deep learning models was also due to the limitations of the dataset size and computational resources. Future studies should address these limitations by using larger datasets, higher-density EEG systems, and more advanced analytical techniques.