Abstract
Neuromarketing is a continuously evolving field that utilises neuroimaging technologies to explore consumers' behavioural responses to specific marketing-related stimulation, and furthermore introduces novel marketing tools that could complement the traditional ones like questionnaires. In this context, the present paper introduces a multimodal Neuromarketing dataset that encompasses the data from 42 individuals who participated in an advertising brochure-browsing scenario. In more detail, participants were exposed to a series of supermarket brochures (containing various products) and instructed to select the products they intended to buy. The data collected for each individual executing this protocol included: (i) encephalographic (EEG) recordings, (ii) eye tracking (ET) recordings, (iii) questionnaire responses (demographic, profiling and product related questions), and (iv) computer mouse data. NeuMa dataset has both dynamic and multimodal nature and, due to the narrow availability of open relevant datasets, provides new and unique opportunities for researchers in the field to attempt a more holistic approach to neuromarketing.
Introduction
Neuromarketing—the study of how our brains respond to marketing stimuli—has long suffered from two major hurdles: secrecy and narrowness. Most high-quality datasets are proprietary "in-house" secrets of private firms, leaving academic researchers with very few open-source options. Furthermore, existing studies often look at brain activity (EEG) in isolation, ignoring the actual behavior (what the person clicked) or the person's underlying personality (e.g., impulsivity).
This lack of integration makes it hard to move from mere correlation to causation. The NeuMa dataset was introduced to provide a standardized, multi-layered view of the consumer decision-making process, bridging the gap between brain activity and actual behavior.
Methods and Experimental Design
The researchers designed a realistic "digital brochure-browsing" experiment involving 42 native Greek-speaking participants.
- The Scenario: Participants browsed six pages of a supermarket catalog containing 144 products. They were instructed to "buy" products as they normally would by left-clicking them.
- Multimodal Recording: While browsing, the system simultaneously recorded 19-channel EEG brain activity, eye-tracking (ET) gaze points and pupil size at 120Hz, and behavioral data (mouse click timestamps and cursor positions).
- Psychological Profiling: After the task, participants completed extensive questionnaires covering demographics, the "Big Five" personality traits, and specific consumer traits (e.g., "Bargain Hunter" vs. "Hedonic Motivation").
- Data Synchronization: All streams were synchronized using Lab Streaming Layer (LSL) to ensure temporal alignment between neural signals and visual fixations.
The NeuMa Dataset
The NeuMa dataset is provided in two forms: NeuMa Raw (untouched synchronized streams) and NeuMa Pre-processed (cleaned EEG and data segmented by product). For every product a user looked at, the dataset includes the specific "snippet" of brain and eye activity associated with that moment. This is significantly more robust than existing publicly available datasets, which typically feature fewer sensors and lack behavioral/personality integration.
Key Results and Insights
Technical validation using EEG Quality Indices (EQI) showed that 95% of the recordings were of "good" or "excellent" quality, despite the challenges of using dry EEG sensors in a dynamic task. On average, participants clicked 18 out of the 144 products, providing a healthy ratio of "Buy" vs. "NoBuy" data points for statistical modeling.
The dataset's structure allows for a "Buy" vs. "NoBuy" contrast, enabling researchers to compare physiological states during purchase decisions. By including consumer traits, researchers can investigate if different profiles exhibit distinct neural responses to stimuli like discount labels.
Limitations and Future Work
The primary limitation is the geographic and demographic specificity; all participants were Greek and relatively young (mean age 31.5). While 42 participants is solid for EEG studies, it is relatively small for training deep learning models from scratch. Future work should focus on exploratory research to identify new neural signatures of preference rather than just classification accuracy. The authors encourage the community to use the provided Matlab/Python scripts to develop alternative segmentation strategies to look at the data through different lenses.
Summary updated