Pathways to Consumers’ Minds: Using Machine Learning and Multiple EEG Metrics to Increase Preference Prediction Above and Beyond Traditional Measurements

Abstract

A basic aim of marketing research is to predict consumers' preferences and the success of marketing campaigns in the general population. However, traditional behavioral measurements have various limitations, calling for novel measurements to improve predictive power. In this study, we use neural signals measured with electroencephalography (EEG) in order to overcome these limitations. We record the EEG signals of subjects, as they watched commercials of six food products. We introduce a novel approach in which instead of using one type of EEG measure, we combine several measures, and use state-of-the-art machine learning algorithms to predict subjects' individual future preferences over the products and the commercials' population success, as measured by their YouTube metrics. As a benchmark, we acquired measurements of the commercials' effectiveness using a standard questionnaire commonly used in marketing research.We reached 68.5% accuracy in predicting between the most and least preferred items and a lower than chance RMSE score for predicting the rank order preferences of all six products. We also predicted the commercials' population success better than chance. Most importantly, we demonstrate for the first time, that for all of our predictions, the EEG measurements increased the prediction power of the questionnaires. Our analyses methods and results show great promise for utilizing EEG measures by managers, marketing practitioners, and researchers, as a valuable tool for predicting subjects' preferences and marketing campaigns' success.

A gateway to consumers' minds: Achievements, caveats, and prospects of electroencephalography‐based prediction in neuromarketing

“…It has been previously shown that prediction accuracy is far better for products that are distant in their preferences, while prediction accuracy is at chance level for products which subjects are indifferent to (Hakim et al, ; Levy et al, ; Telpaz et al, ). Another study has shown that it is harder to discriminate between the neural activation of similar brands (Chen, Nelson, & Hsu, ).…”

Deep Learning for EEG-Based Preference Classification in Neuromarketing

“…We classified these predictive features based on the EEG signal types: (1) rhythms; and (2) transient activities. [17] Viewing brands and products ERP P300 BCI [11] Watching TV ads EEG Alpha (for emotional state) EMG, BCI and GSR [36] Viewing brands and images ERP N200 and P300 BCI [37] Watching TV ads EEG Alpha band frontal asymmetry BCI [35] Watching TV ads EEG Theta and gamma Heart rate, BCI and GSR [28] Watching TV ads EEG Asymmetrical increase in theta and alpha in PSD BCI [38] Viewing brand names ERP N400 BCI and EOG for eye movement [18] Viewing products and prices ERP FN400, LPC, and P200 BCI [39] Viewing products EEG Alpha, beta, theta, gamma, and delta BCI, Eye tracking [40] Viewing products ERP P300 BCI [41] Watching TV ads EEG Theta and alpha Heart rate, BCI and GSR [3] Viewing products ERSP and ERP Theta, N200, and FRN BCI [24] Watching ads (movie trailers) EEG (64 electrodes) Beta and gamma oscillations BCI and EOG for eye movement (2 electrode) [20] Watching TV ads dense-array EEG Three epochs: 200-350, 350-500, and 500-800 BCI [42] Viewing brand names ERP LPP BCI [43] Viewing product images ERP N200, LPP, and PSW BCI [30] Watching ad videos EEG Theta and alpha Heart rate, BCI and GSR [22] Viewing product images EEG Delta, theta, alpha, beta, and gamma BCI [29] Viewing product images EEG Alpha BCI [44] Viewing ads of food products EEG Delta, theta, alpha, beta, and gamma BCI [32] Viewing products and prices EEG Theta BCI [31] Tasting drinks EEG Alpha BCI [33] Viewing and touching products EEG Alpha and theta BCI [27] Viewing product images ERSP, ERP Theta, beta and N200 BCI [45] Viewing tourism images, videos and words EEG Delta, theta, alpha, beta and gamma BCI and GSR…”

“…Low-pass filter 3 [22,24,38] Band-pass filters 10 [3,11,15,20,28,[39][40][41]44,53] Down sampling 3 [11,15,44] Average 7 [3,15,27,37,38,40,54]…”

“…Statistical features: median, standard deviation, t-test, z-score transformation 9 [3,11,24,28,35,37,38,40,41] Time-frequency analysis 2 [15,54] Wavelet domain: entropy, energy » using wavelet transform 2 [22,39] Frequency domain: power spectral density, band power, spectrum power » using Fourier transform 8 [11,27,37,39,44,[53][54][55] Correlation analysis: mutual information, correlation coefficient 3 [18,27,39] Regression/Classification [22] Linear regression (Lasso, Ridge) 1 [27] Some preference studies involved the use of more than two classification algorithms to discover well-matched classifiers for a definite feature set [12]. Chew et al [54] measured the user preferences for the aesthetics presented as virtual 3D shapes by using EEG.…”

“…In their extended study [55,56], the authors increased the number of subjects to 16 but better results were not obtained. Hakim et al [44] achieved an accuracy of 68.5% by using the SVM to predict the most and least favored products by combining EEG measures with questionnaire measures.…”