Abstract
This paper presents a Python-based workflow for bridging the gap between subjective emotional ratings and objective physiological data in food science. Utilizing Subgroup Discovery (SD) via the pysubgroup library, the authors identify human-readable rules linking skin conductance responses to the perceived pleasantness of food odors. Key innovations include individualized discretization of subjective ratings and the use of descriptive data mining to uncover specific physiological signatures. Results demonstrate that specific skin conductance metrics, such as rise-time and latency, can reliably characterize hedonic responses, providing an interpretable alternative to "black-box" predictive models.
Introduction
Understanding human emotional responses to food stimuli is a complex challenge in sensory science. Researchers typically balance subjective self-reported ratings with objective Peripheral Nervous System (PNS) activity, such as skin conductance. However, subjective ratings are highly individual, and physiological signals are inherently noisy and multidimensional. Traditional statistical methods often fail to capture the nuanced "if-then" patterns that explain the relationship between a physiological state and a specific emotion. This study proposes a data science framework to provide descriptive and explanatory insights into these biological mechanisms, moving beyond simple prediction toward interpretable discovery.
Methods
The proposed protocol follows a structured workflow designed to handle the variability of human psychological data and the complexity of physiological signals.
The core of the methodology involves three critical stages:
- Individualized Discretization: To account for individual rating styles (where one person's "7" might be another's "5"), the authors apply K-Means clustering to each participant's scores independently. This transforms continuous pleasantness ratings into categorical classes relative to the individual's own scale.
- Subgroup Discovery (SD): Using the
pysubgrouplibrary, the system searches for subpopulations where a target property (e.g., "unpleasant") is significantly more prevalent. Unlike standard machine learning that aims for global prediction, SD generates explicit, human-readable rules like: “If Rise-Time is ≥ 5.35 seconds, then the odor is likely unpleasant.” - Quality Metrics and Validation: Rules are evaluated using Weighted Relative Accuracy (WRAcc) and Lift. To ensure statistical significance, results are validated against 10,000 bootstrap iterations to prove patterns are not random noise.
The framework was tested on a public dataset (Licon et al., 2018) containing 2,398 observations from 22 individuals, focusing on four skin conductance features: Amplitude, Latency, Rise-Time, and Number of Events.
Results
The application of individualized discretization significantly cleaned the data for mining, as shown in the processed dataset structure where raw scores are mapped to hedonic classes.
The algorithm successfully identified interpretable rules for hedonic classes:
- Unpleasant Odors: A Rise-Time ≥ 5.35 seconds was the strongest indicator, showing a Lift of 1.25, meaning this state is 25% more frequent in unpleasant trials than in the general dataset.
- Pleasant Odors: A specific Latency interval [1.92 to 2.26 seconds] was identified as a key marker with a Lift of 1.23.
Validation confirmed that these top-ranked rules were statistically significant, falling well outside the 90% confidence interval of random distributions, confirming the patterns were meaningful biological markers.
Discussion and Conclusion
By prioritizing descriptive discovery over simple prediction, this approach offers "interpretable AI" for food science. The use of per-subject clustering is a critical innovation for handling the inherent subjectivity of emotional data, treating pleasantness as a relative state rather than an absolute number. While the current study is limited by a small participant pool and a focus on odors, it establishes a robust foundation for future research into multimodal stimuli, such as the combined sensory experience of taste, texture, and visual appearance. Future work will explore alternative search strategies within the Subgroup Discovery framework to further refine these physiological signatures.
Summary updated