Making Sense of Scents: Integrating Consumer Sensory Perceptions and Expert Assessments of Perfumes with Bayesian Networks

Please join us for a free live webinar on Thursday, May 22, 2025, at 11:00 a.m. (EDT, UTC-4).

Background

In this webinar, we return to one of our best-known tutorials, the so-called "Perfume Study." We have widely used this example of a key driver analysis in BayesiaLab courses and seminars, and it also features prominently in Chapter 8 of our book: https://www.bayesia.com/bayesialab/e-book/chapter-8-probabilistic-structural-equation-models-for-key-driver-analysis

The original study was based on a monadic consumer survey on perfumes conducted by a market research agency in France. Each respondent evaluated a single fragrance, allowing for independent assessments without direct comparison. For this example, we focus on survey responses from 1,320 women who evaluated 11 perfumes across a wide range of sensory and perceptual attributes.

The Perfume Study Revisited

We now examine this familiar dataset again, but with the benefit of additional expert assessments of the studied fragrances and related metadata.

Building a Probabilistic Structural Equation Model

The new study mirrors the original one in developing a Probabilistic Structural Equation Model, i.e., creating latent factors to summarize the sensory perceptions of respondents manifested in the survey. 

Multi-Quadrant Analysis

A key feature of the key driver analysis workflow in BayesiaLab is that an overall model representing all products in a market can be split into submodels representing each product separately. As a result, we can examine the market dynamics as a whole, but perform optimization for each product individually.

Sensory Analysis by Experts

Consumer perceptions are critically important, but ratings are typically expressed in layman's terms, which may not translate directly into the technical attributes used by product developers. In the context of perfumes, we require a structured sensory evaluation by trained experts.

Linking Consumer Perceptions and Expert Evaluations

Combining the consumer view and expert assessment to create a unified understanding of products has long been a challenge. The workflow proposed in this webinar provides a robust framework for amalgamating the opinions of untrained consumers and the professional judgment of experts.

Consumer Comments Provide Structure for Learning

Large Language Models (LLMs) have recently made great advances in providing narratives as answers to questions. Hellixia, BayesiaLab's GenAI assistant, utilizes LLMs not to produce new narratives but to obtain structure from existing narratives. More specifically, Hellixia taps into LLMs to translate free-form consumer feedback, such as that obtained in open-ended survey questions, to structured causal knowledge. 

Semantic Analysis of Comments

A central task for this purpose is the semantic analysis of consumer narratives. Hellixia performs semantic and conceptual mapping to bring meaning to often informal and imprecise expressions contained in consumer feedback. 

From Comments to Causal Hierarchy

Hellixia's semantic analysis of consumer feedback produces a hierarchy of causal drivers, which BayesiaLab can use to perform a formal causal driver analysis and, subsequently, a causal driver optimization.