Elevate Your Preference Mapping Efforts with Machine Learning

Preference mapping has traditionally been highly regarded for its effectiveness in uncovering and visually representing consumer preferences derived from sensory data. Its impact on tasks such as product formulation, refinement of marketing strategies, and enhancement of product quality is without a doubt significant.

Nevertheless, machine learning, endowed with the ability to comprehend intricate, non-linear relationships and predict individual consumer preferences with remarkable accuracy, emerges as a compelling force in modern consumer research.

In an era marked by intense competition and the prevalence of data-driven decision-making, it is imperative for consumer research to embrace these technologies. This enables the unearthing of hidden insights and facilitates the making of well-informed, scientifically-backed decisions in product development.

Preference mapping (PM)

Preference mapping (PM) stands as a pivotal tool in consumer research, enabling examination and visualization of consumer preferences derived from sensory data. Its contributions are far-reaching, offering invaluable insights into product formulation, marketing tactics, and the fine-tuning of product quality.

The merits of preference mapping can be summarized as follows:

  • Visualizing preference relationships: This method excels in presenting the intricate web of relationships that underlie consumer preferences.
  • Discerning key influential attributes: PM identifies the pivotal attributes that influence preferences, a crucial aspect in refining product offerings.
  • Suitability for small-scale studies with linear relationships: In scenarios where studies are more compact in scale and the relationships between attributes and preferences are relatively straightforward, PM offers valuable insights.

Machine learning (ML)

While PM has certain advantages as outlined above, machine learning (ML) presents a broader toolkit capable of comprehending intricate (non-linear) relationships and making accurate predictions of consumer preference on an individual level. 

The advantages of ML are multifold and can be summarized as follows:

  • Complex non-linear relationships: ML can capture complex non-linear relationships between sensory attributes and consumer preferences. This is crucial for understanding how combinations of attributes affect liking, which may not be easily captured by traditional methods.
  • Enhanced prediction accuracy: ML models have the capacity to learn complex patterns in (sensory/consumer) data, leading to more accurate predictions of consumer preferences. This can result in products that better align with consumer tastes.
  • Optimized product formulation: ML can help identify the optimal combination of sensory attributes to create products that are highly preferred by consumers. This can streamline the product development process.
  • Feature importance identification: ML models can automatically identify the most influential (sensory) features. This information is highly valuable for prioritizing efforts in product optimization and formulation.
  • Personalization and segmentation: ML facilitates the provision of personalized recommendations, enabling targeted approaches towards specific consumer groups with tailored product offerings.

ML in preference mapping

To elucidate the distinction between PM and ML, a dummy example that serves as a practical illustration of their contrasting approaches is presented. In this example, PM and ML techniques were employed to describe and predict consumer preferences for a range of products (50 products) based on user preference data (5 users) and associated product sensory features (4 features).

Two distinct models were utilized to capture the complexities associated with the data: 

  • a conventional linear model and 
  • a non-linear machine learning model.

Preference Mapping Graph

Above figure illustrates the difference in predictive performance between a (left) traditional linear and (right) more advanced non-linear ML model. As the latter is able to capture non-linear relationships more effectively than the former, it can provide more accurate predictions, particularly in situations where the underlying relationships are intricate, which is clearly the case in this example.

Traditional PCA

Example of a preference map, biplot of sensory features (red) and users (blue). It is apparent that the user group likes products which score high in the Sensory_2 feature, while the opposite is true for Sensory_3 and _4. Preference mapping also reveals that there is 1 cluster of consumers with similar preferences.

Impact of sensory features on model output

While the PM visualization offers a glimpse into sensory-related user preferences, the ML approach goes a step further by providing crucial insights into key feature identification, as well as opportunities for product optimization and targeted user engagement.The figure above illustrates the impact of the sensory features on the model output for the ML model. Clearly, an order in influence is present, where user preference will be largely impacted by Sensory_2. Followed by Sensory_3 and Sensory_1. Sensory_4 shows far less influence, almost negligible. A positive relationship can be observed between Sensory_2 and model impact: high values for Sensory_2 will result in high values for user preference (i.e., a positive impact on the model output, here right from the vertical line). This is not the case for Sensory_3, where high values will mostly result in lower preferences (i.e., negative impact on model output, left from the vertical line in the figure above). The relationship between Sensory_1 and model output is less straightforward and more intricate.

Sensory 2 pd sensory 4







Partial dependence plots reflect the expected output of the model if we were to intervene and change exactly one of the model parameters. The influence of Sensory_2 is obvious, larger values will result in higher user preference in general, conform figure above. The influence of Sensory_4 is negligible as also explained in the previous figure.

Employing ML yields distinct outcomes compared to the traditional methodology during product optimization (see table). In this example, the utilization of ML not only yields a product with a more intricate combination of features but also garners a significantly higher user preference score (+30%) compared to the product obtained through traditional methods when it is presented to the same user group.

Table Pref. Map Traditional vs ML 1

Preference mapping at Foodpairing AI

The choice between employing ML and PM seems contingent on various factors. In instances where extensive, diverse and complex datasets are available, ML presents a potent toolset for product optimization. Industries that rely heavily on sensory evaluation, on the other hand, may find PM indispensable for optimizing product attributes. 

However, at Foodpairing AI, we have established an integration of both methodologies. By leveraging our extensive sensory database, we concurrently harness the predictive power of advanced non-linear ML techniques while also extracting invaluable sensorial insights for the purpose of product optimization. This approach embodies a convergence of the strengths of both ML and PM methodologies.

It should be stressed that ML can uncover hidden insights and patterns in (sensory and consumer) data that may not be immediately apparent through traditional PM techniques. In a competitive landscape, businesses that leverage advanced analytics and ML gain a strategic advantage. Adopting these technologies can lead to better-informed product development decisions.