Predicting how consumers, from a variety of backgrounds and demographics, will respond to a new product is hard. Most food and beverage producers today use a mix of qualitative and quantitative consumer surveys to determine if a product is “good enough” to launch, only to find out later that good enough doesn’t make for a great product development strategy.
The problem is that data generated from qualitative and quantitative consumer surveys isn’t predictive – it’s only explanatory. Statistical analysis by an innovation team or product development team can show if the consumers included in the survey liked a product or even how much they liked a product, but that data is only useful if those consumers are a stratified and representative sample of every cohort in your target consumer group. Which is unlikely, so the data collected most likely will not generalize well.
So after identifying which flavors are going to trend into long-term brand-making preferences with Deep Market Insights, how do you cut down on the iterative cycles and expensive consumer tests and use predictive data to create winning new products?
Background: The Problem
Gastrograph AI is the first AI platform for the food and beverage industry. The AI models the sensory perception of every consumer demographic across age, sex, race, socioeconomic status, and smoking habits, to predict consumer preference today and the evolution of preference into the future. Gastrograph AI uses hundreds of self-learning algorithms to various attributes of incoming reviews. Unfortunately, many of the highest performing algorithms are uninterpretable – we can prove that they work, but can’t see into the model to know why it works.
Thus, with models such as neural nets, autoencoders, and support vector machines, it is increasingly difficult to look inside and see which group of flavor attributes or somatosensations contributes to achieving a high perceived quality for a given style of product. Yet this is precisely what food and beverage producers care about: what flavor profile should I aim for in order to achieve a high perceived quality score among the general population?
How can the team at Analytical Flavor System help our clients solve this problem, while still using the highest accuracy models?
We could try to input every possible flavor profile, for every possible class and style of product, into an algorithm to predict perceived quality and consumer preference. If that worked, we could just choose the highest performing flavor profiles in each category. But there are over 4.7 x 1018 possible flavor profiles per class of product on the Gastrograph System, and that does not include specific reference flavors or account for demographic and geolocation biases.
Using brute force and trying out every flavor profile is not only inefficient, but also improbable: this would have to be rerun for every flavor-profile with every update of the model across every tasting demographic and preference archetype. Now, the problem becomes: what is a way of efficiently searching for the best flavor profile of a given style of product?
The Solution: Genetic Algorithms
Genetic algorithms are a way of efficiently searching through any range of values. They are the result of data scientists and mathematicians trying to model how evolution results in a population better fit to survive in a given environment. In biological evolution, when DNA from the parental generation makes the genetic information for their progeny, DNA crossover, mutation, and recombination to form the progeny. The best fit offspring in turn survive and reproduce, to produce offspring even better fit for the environment. In this way, the DNA of the population changes and converges over time to produce the best fit individuals for any given environment.
Genetic algorithms mimic this process. The algorithm first generates a random population of values for some given parameters. In subsequent 'generations', the algorithm chooses the high performing units of the population, chooses the parameter values of the subsequent child unit from either of the parental units, mutates the values with some low percentage, and evaluates the generated population with a fitness function to determine how fit it is to survive in this environment. The highest performing units are chosen to survive to the next generation, and the rest are either terminated or bred again, depending once again on how well they score according to the fitness function. After multiple iterations of this process, the population converges to a set of observations that result in a high value of fitness.
Genetic algorithms allow us to generate theoretical flavor profiles and test them against a prediction of consumer preference learned through reviews of existing products.
If we start with the flavor profile of the first iteration of a product under development, we can ask the AI to optimize it for a target demographic – say, millennial female new-premium consumers. The AI sets off mutating the product around inferred axis of variance (keeping clear of flavor combinations which can’t be produced) and tests each of its new hypothetical flavor profiles against the predicted preferences of the target consumer cohort. The AI can spawn hundreds of mutations at every generation, and learns which combination of changes has the greatest effect on preference. It then culls the worst performing hypothetical flavor profiles (like a flavor eugenics program!) and mutates a new generation of flavor profiles from the best of the last generation. After a few generations, the population converges to a set of flavor profiles that lead to the highest perceived quality.
Figure 1. Starting Flavor Profile
Figure 2. Generation 2
Figure 3. Generation 3
This results in a few possible target flavor profiles for the product under development – with highly accurate predictions for its success and distribution of consumer assessment across demographics.
Figure 4. Predicted Market Preference
Using Gastrograph AI’s genetic algorithms, your innovation and product development teams can create better, more successful, products in fewer iterative cycles using the power of data.
Since all of these parameters can be controlled, we can help producers that want to create specific categories or styles (for instance, an IPA) for their target consumers (i.e. female millennials).
You and your team can also limit the possible flavor profile values however they want. For example, if a brewery wants to make an IPA with a low level of herbaceous flavor (or aroma), they could specify that, and the AI will find the optimal flavor profiles with low herbaceous perception. Our clients can put as many restrictions on the output flavor profile as they desire. Linking the output flavor profile with the production process needed to achieve it provides an end-to-end new product development application, from conception to actualization.