Determining Flavor-Attribute Sensitivity of Panelists

Controlling for Individuals' Sensitivities to Flavor-Attributes


Rishab Gupta

Data Science Intern

Jason Cohen

Founder & Chief Executive Officer


Every individual experiences flavor differently. While I might think that a particular beer is mildly bitter, others may be overwhelmed by the same flavor. While it would be convenient for sensory scientists to alter the taste buds of every human being on Earth (or at least in their panel) to be the exact same, that is (as far as we know) impossible and thus it is necessary to account for differences in each panelists sensitivity to each flavor-attribute measured. Historically, measuring individuals' flavor-attribute sensitivities at scale was challenging, if not impossible. But recent developments in sensory science and machine learning have led us to a new kind of predictive intelligence - one that is able to accurately measure individuals' sensitivities to flavor-attributes. It is now possible for any food and beverage manufacturers to better utilize sensory testing by controlling for individual panelists' biases, helping save time and money.


At Gastrograph AI, accurately modeling human sensory perception is our bread and butter, which makes the problem of measuring individuals' sensitivities a necessary prerequisite for our other quality control, flaw detection, and flavor profiling models.

For the purpose of this discussion, I am presenting a brief introduction to the Gastrograph application. If you have used the application before and are aware of how it works, feel free to skip the rest of this section.

Gastrograph is used to review beer, coffee, chocolate, and other food and beverage products. In each review, a user can input data for 24 flavor-attributes. Each flavor-attribute is rated on its intensity on a discrete scale from zero to five (zero: not present, to five: most intense; initializing at 0) which constitutes the flavor profile of a sample. Once all of the flavor-attributes’ intensities have been entered, the user inputs a value for perceived quality, which is a discrete value from one to seven (lowest to highest “quality”) that reflect the user's perception of the quality of the product. Shown below is a typical beer review:


The Data

For the sake of brevity I will only talk about beer, although this approach works identically for all types of homogenous food and beverages. The data consists of approximately 10,000 beer reviews (sampled from roughly 2,000 unique products). For these reviews, we have used the 24 flavor-attributes and the review weights in our analysis, which is a value between zero and ten. The review weights represent the adjustment we assign to each review, and they are calculated as the product of predicted review number and review trust. A discussion on the calculation of review trust is beyond the scope of this paper, but in simple terms it can be thought of as being a value between zero and three that expresses how much we trust that review based on a variety of factors.

The Data Science

It’s important to understand that user sensitivity analysis must be conducted by testing a user’s response to tasting individual foods or beverages. Let’s look at a common situation faced when studying craft beer drinkers. If a person has only reviewed sour beers, a generic analysis might say that the individual is more sensitive to sour flavors. Instead, the analysis should be focused on how their review varies from every other review of the same product. To do so, we generate an Objective Flavor Profile (OFP) for the product and find the deviation of each review of that product from the OFP. The OFP is generated as a weighted and normalized average of flavor-attributes using the review weights and certain other variables as categorical groupings.

With the deviations of a review from the product’s OFP, we have succeeded in translating original flavor-attributes into units of absolute deviation. The next step is to perform a summation of every user’s reviews to determine their sensitivities to the 24 flavor-attributes. However, this translation by itself is not enough to intuit the sensitivities because the absolute values of these deviations don’t convey much information. An alternative metric that better represents internal deviations between individuals is needed. The summation of absolute deviations from OFPs per user were translated into units of standard deviations to represent how susceptible a user is to a particular flavor when compared to the other users.

To summarize, for each review, the original flavor-attributes are converted into units of deviations from the Objective Flavor Profile. Then the standard deviation for each flavor-attribute per individual is calculated in comparison to other users to predict their sensitivities. The final prediction is then composed of user sensitivities to each of the flavor-attributes in units of standard deviations.

The Results

User sensitivities can now be plotted individually or with respect to other individuals to provide a metric for comparison. Shown below is a graph of my sensitivity to the 24 flavor-attributes versus other employees at Gastrograph AI. The X-axis represents each of the 24 flavor-attributes, sensitivities to which are measured in units of standard deviation on the Y-axis. For each flavor-attribute, my sensitivity (denoted by a blue triangle) is plotted along with a box- plot that illustrates the spread of sensitivities of other Gastrograph AI employees.

The generated plot provides a rough idea as to how my reviews differ from that of the majority of individuals. The most notable deviations from other users are my low sensitivities to flavors of “sugar”, “roasted”, and “spices” and high sensitivities to flavors of “marine”, “herbaceous”, and “astringent”. This could mean that I have particularly deviant sensitivities to these flavors or that I use the application incorrectly, especially when it comes to these flavors. In either case, it provides reason for my reviews to be adjusted during model building and my employers can analyze the graph to decide if I need to be subjected to further training.

Rishab Sensitivity

Another analysis that could be done using such a plot would be a market-targeted product development. For a moment, consider Gastrograph AI to represent a demographic. Using the box-plots, we can draw an inference regarding the sensitivities of that demographic (note: Gastrograph AI is a small team of diverse individuals, so the range of deviations is more pronounced than it would be for an actual demographic). This insight can now be used to create products for the targeted audience.

The Applications

As you can imagine, an analysis such as this has a plethora of applications (which we have helpfully built into our service as an automated analysis). This information is not only relevant to Gastrograph AI, but also to producers in the food and beverage industry and people who want to know more about their own flavor preferences. Let’s jump into a discussion of some of the most interesting and important applications of this analysis.

  1. Increased market intelligence for food and beverage producers (breweries, coffee roasters, etc):

    a. Formulating products for a targeted audience: This analysis helps explain the preferences of consumers in a given market. In order to tailor a product for a specific market, producers can generate a flavor profile that more closely matches the preferences of the consumers who share a clustering of sensitivities.

    b. Targeted sales: The Objective Flavor Profile of each product can be compared with the sensitivities of different demographics within the general population, and Gastrograph AI can help producers determine the population where the product would be most well-received. These predictions can also be bolstered by existing sales data to achieve a higher precision.

    c. Employee training: Most producers do in-house reviews of their products at various stages of production. It is crucial that members of the tasting panel are well-trained so that the company can trust their reviews and make decisions based on them. Gastrograph AI can help identify specific patterns and sensitivities in the employee's’ reviews and identify those whose sensitivities are highly deviant, in which case it might be a wise option to have them retrained.

  2. Users:

    This analysis can help users better understand their own flavor preferences and sensitivities and even use this information to train themselves. For instance, I love IPAs and am usually hesitant to move out of my comfort zone and try other beers. Although this analysis reinforced the fact I am more sensitive to bitter flavors, which are predominant in IPAs, it also revealed that I have an affinity to sour flavors. As an experiment, I decided to test this prediction and bought myself a few sour beers and I quite liked them! So not only does this analysis help you pick the right beer, but it can also be used to explore the limitless boundaries of flavors and taste products you never would have otherwise experienced.

  3. Gastorgraph AI:

    This analysis was shared with all employees at Gastrograph AI, all of whom take pride in being professional tasters for two or more products. This triggered a journey of self- exploration as people tried to learn more about their flavor preferences. Moreover it served as a means to perform employee training as discussed above. As for the data team, we couldn’t wait to play with our shiny new toy. Some of the internal applications of user sensitivities include:

      a. Retraining models that suffered due to generic analysis of data.
      b. Clustering of users based on sensitivities to predict appropriate targeted audiences for particular products.
      c. Assigning trust scores to reviews and reviewers.

Future Work

As mentioned in the sections above, this analysis has a wide range of applications; some have been explored and others need further analysis before being realizable. Some options for future work include but are not limited to:

  1. Building more Market Intelligence tools for our clients.

  2. Predicting reviewer preference changes as sensitivity changes over time with training and experience.

  3. Predicting a range of flavors that a user is likely to like. This can be used to make a beer recommendation system in the future!

  4. Using user sensitivities as the starting point, drawing correlations amongst flavors.

  5. Predicting the subset of users that are more likely to well- receive a product based on the products OFP and the user’s sensitivity. This could be extended to predict user demographics, age, sex, etc.

Notice anything backwards, missing, misleading or incorrect? ! 

Create a better-tasting world

Get in touch to speak with our team so we can learn about the challenges you're facing.