Integrating Cacao Physicochemical-Sensory Profiles via Gaussian Processes Crowd Learning and Localized Annotator...

Area of Science:

• Food Science and Technology
• Data Science and Machine Learning
• Sensory Analysis

Background:

• Integrating sensory perception and physicochemical properties of cacao products is vital for quality control and innovation.
• Challenges exist in combining heterogeneous data, especially subjective and inconsistent sensory evaluations from multiple experts.
• Existing methods struggle with noisy, subjective, and non-stationary sensory data.

Purpose of the Study:

• To develop a comprehensive framework for integrating sensory and physicochemical data in cacao-based products.
• To address the challenge of learning from noisy, crowd-sourced, and non-stationary sensory annotations.
• To enable accurate estimation of perceptual scores and infer annotator reliability.

Main Methods:

• Construction of an integrated database combining physicochemical parameters and sensory descriptors.
• Application of a correlated chained Gaussian processes model for learning from crowds (MAR-CCGP) to infer ground truth from noisy sensory data.
• Development of a localized expert trustworthiness approach using MAR-CCGP to dynamically adjust for annotator consistency variations.

Main Results:

• The MAR-CCGP framework demonstrated robust, interpretable, and scalable learning from heterogeneous and noisy sensory data.
• Experiments on semi-synthetic and real-world Casa Luker data showed superior predictive performance compared to state-of-the-art baselines.
• The method achieved more precise estimation of annotator reliability, adapting to input-dependent variations.

Conclusions:

• The proposed framework provides a principled foundation for data-driven sensory analysis and product optimization in food science.
• MAR-CCGP effectively handles multi-annotator regression settings with noisy and subjective data.
• The unique combination of a novel database, robust regression, and input-dependent trust scoring distinguishes MAR-CCGP.