The Art, Science, and Data-Science behind Food Pairing
Simply put, food pairing is the process of identifying the combination of flavors, textures, and aromas that can excite and capture the imagination of human taste-buds.
This process involves unravelling the flavours, textures, and aromas of the individual ingredients that can serve as the core foundation for the new product development and innovation for the Food & Beverage, FMCG, CPG, and Flavour businesses.
Food pairing is both art and science!
It is an art because it demands:
- Culinary excellence
- In-depth understanding of the taste-buds of target-audience
- Learning from past successes and failures
- Your quotient of self-belief, intuition, and risk appetite
Food pairing is also a Science because it involves:
- Identification of the aroma profile of ingredients through High-performance liquid chromatography (HPLC), Gas Chromatography (GC), and other laboratory processes
- Evaluation of which of these aromas will be discernible by the human sense of smell and taste
- Quantitative analysis of ingredients pairings, and more
Traditionally, these aspects related to both art and science have been at play behind some of the most loved food pairings like honey + lemon, tomato + basil, and even caviar + white Valrhona chocolate (courtesy Michelin Star Chef and Culinary Scientist, Heston Blumenthal)
However, with the emergence of big data along with Natural Language Processing (NLP) and Machine Learning (ML) capabilities of Artificial Intelligence (AI) platforms, data science is increasingly becoming a crucial element in the process of food and flavor pairing.
In this blog, Kasun Perera (Head of Data Science at Ai Palette) and I (Co-founder of Ai Palette) will try to decode the technology behind AI-driven food pairing.
We will also discuss how AI can help you to identify ingredient pairings that are otherwise beyond the realm of Human Intelligence and traditional methods.
Tracing the origins of the chemistry and science behind food pairing
Heston Blumenthal began his tryst with gastronomy at age 16, when he visited L’Oustau de Baumaniere in a small village in Provence, France.
His curiosity, creativity, and culinary genius led him to experiment with salty ingredients and chocolate. He happened to put the salty caviar and a piece of white Valrhona chocolate simultaneously into his mouth.
The resulting combination turned out to be exquisite, and the eventual recipe ended up creating one of the best-known dishes on the menu of his first restaurant – The Fat Duck.
Heston partnered with François Benzi, from the flavor house – Firmenich, for flavour analysis of the food ingredients, laying the stepping stone for the lab-based scientific approach towards food pairing. And the rest is history!
The benefits of the lab-based scientific method of food pairing:
- The scientific method evaluates the aroma profile and the chemical composition of the ingredients. It provides quantitative evidence to support human decision-making.
- It is a time-tested approach. Over the decades, many organizations and individual chefs have achieved the desired maturity in tools, systems, databases, and processes required to ace this method.
The limitations of the lab-based scientific approach towards food pairing
- The aroma profile evaluation, at times, favours unknown food combinations that may not find acceptance among the target audience leading to product failure, post-launch. It has widespread consequences, specifically for F&B, CPG, and FMCG companies.
- Though this scientific method focuses on the sensations of sense, touch, and smell of humans; but it may lack the depth of understanding required for cuisine specific cooking styles. E.g. – Asian cooking has more spikes in taste (spicy, salty, sweet, and sour) as compared to the European palette that has a more blended taste.
This method deploys conventional tools, databases, and processes that are not capable of deriving real-time consumer insights.
The ‘Who’,’Why’, and then ‘How’ of the real-time consumer insights in food pairing
It is the F&B and FMCG/CPG stakeholders, who can benefit the most from real-time consumer insights.
This is because Big Data analytics can enable organizations to make decisions that are based on the prevailing popularity of the ingredients, amongst the larger set of target audience.
Extraction of such real-time consumer insights from the large volume, variety, and velocity of the unstructured Big Data brings Natural Language Processing (NLP) capabilities of an AI and ML at play.
We will find out from Kasun, how AI and ML technology works behind the scenes to derive insights in the context of your specific business use-cases or pain-points and how it can ensure that the existing Human Intelligence works in collaboration with the Artificial Intelligence.
Over to you, Kasun 😊
How NLP-NER capabilities of an AI platform and graph database can help you identify your next flavor or food-pairing
As Som mentioned, the lab-based scientific approach relies on insights from the underlying chemical composition of the ingredients.
On the other hand, the AI-driven (data-labs based) food pairing can evaluate the strength and potential of the food pairs and provide recommendations by exploring additional dimensions and deriving correlations, like:
- the existing affinity of your target audience towards ingredients/recipes
- existing or emerging food-habits and trends in a target geography
- identifying new ingredients (which were not specifically listed as ingredients) in recipe documents, social media posts/comments, and other sources of data
The list of these probable correlations can be practically never-ending and depends entirely on the availability and quality of your Big Data, the diversity of your data sources, and also on the F&B product knowledge and problem-solving capabilities of your data-science and product development teams.
To understand the AI-driven food pairing process, let’s get introduced to the NLP (Natural Language Processing) and ML (Machine Learning) technology-stack of an AI (Artificial Intelligence) platform.
The following are the key components of this tech-stack:
- Source of Big Data:
- Target-audience generated unstructured data from social media
- Recipes and menus from the foodservice industry (from cafes, bars, restaurants to fine-dining and even street food stalls)
- Recipe databases built, maintained, and driven by various communities within the F&B ecosystem
- Customer reviews and other data from E-commerce marketplaces and platforms
- SKUs’ related data from retail outlets
- Named-Entity Recognition engine, which is one of the features of Natural Language Processing (NLP).
- Frequency Analysis to determine the strength of the pairing of the ingredients.
- Graph Database(s) to store and process the correlation data regarding the identified ingredients a.k.a nodes (as per the nomenclature of the graph databases). You can also retrieve the food pairing recommendations from the graph database by executing database queries that represent your business use-case or problem statements.
Now, with the help of a fictitious problem statement, let us find out how all these various components play their part in determining the most probable food pairing recommendations:
- Assume that we are on a mission to find a recommendation for the following problem-statement:
- A popular F&B company wants to strengthen its product portfolio by launching new variants of their Parmesan based Snacks in the USA. They are looking for recommendations of new ingredients that can pair well with Parmesan
- Step#1 would be to identify available sources and procure Big Data that can provide consumer insights regarding ‘parmesan’, as an ingredient and ‘snacks’, as a product category for our target geography.
- In Step#2, the raw, unfiltered, unstructured, and voluminous Big Data is put under the scanner of the Named Entity Recognition (NER) engine or model. This NER model is trained and configured to identify ingredients from this target-audience generated data.
- Once relevant ingredients are identified by the NER engine – the numbers can go up to a couple of thousands, or more depending on the input data – in Step#3, frequency analysis is conducted to determine the strength of the correlations between the identified ingredients.
- All the identified ingredients, their various correlations with each other, the quantitative value of these correlations, and other information is then stored in a Graph Database. Such a database is best suited and cost-effective for storing, processing, and retrieving information that involves large-scale interconnected relationships across the data points.
A sneak-peek into how Graph Database stores various identified Ingredients and their relationships
- Now, with the help of the data science teams, the product development teams of this F&B company can query the Graph Database to fetch recommendations for their specific problem-statement(s):
- As seen in the above snapshot of the Graph Database, Chicken is one of the ingredients that has been paired with Parmesan. We can query the Graph Database to share with us the ingredients that can be paired as a third ingredient with Chicken and Parmesan for the product category ‘Snacks’, for USA geography.
- The results can be as fascinating as below:
A typically recommended output from an AI-driven Food Pairing process
After an immense data-crunching and a lot of learning/unlearning, the AI/ML platform powered by the NER engine and aided by Graph Database has been able to divulge the necessary recommendations.
But if you ask Som, this is half the battle won! So, over to Som who will help us see the larger picture of Human and Machine collaboration.
Welcome to the era of Augmentive Intelligence
Kasun’s love for data is very evident from his well-articulated explanation of how AI can empower us with food pairing recommendations based on the insights from Big Data.
And he is right in saying that this is half the battle won. This is because, once we have the recommendations, it is not the time to declare victory but to leverage ‘Augmentive Intelligence’.
Augmentive Intelligence is all about collaboration between Human Intelligence and Artificial Intelligence to find solutions that matter.
Bounded by the data-bubble that Big Data creates for the AI/ML Platform, we have a list of the best possible ingredient pairing combinations as per the criteria defined by our problem-statement.
It is important to highlight the fact that the AI platform didn’t deliver a specific solution for our problem statement, but it provided us with a list of possible solutions.
For bounded human rationality, it wouldn’t have been possible to swift through Big Data and find a way out to arrive at these probable solutions.
But it is very much possible for Human Intelligence to leverage the knowledge about the target audience, their geography, past product successes and failures, conduct laboratory tests, and identify a likely winning ingredient pair from the list provided by Artificial Intelligence!
All this can result in your organization launching new products faster and have a better shot at success.
Dr. Peter Angeline, a Strategic Futurist at The Hershey Digital Innovation Lab, explained Augmentive Intelligence during his keynote speech at the Transforming Food Insights Summit in November 2020 (you can watch the recording here – AI’s Inevitable Disruption of CPG).
Kasun and I believe in the potential of Augmentive Intelligence. It can enable organizations to see beyond the obvious and leverage this opportunity to understand customers better and act faster.
And when you and your organization are ready to test the waters, you will have to embark on the journey of enterprise-wide Change Management.
I will conclude this blog with my two cents on Change Management.
The art of Expectation Management for successful Change Management:
During my several conversations with CPG, F&B, and FMCG organizations, who are at various stages of adoption of Artificial Intelligence (AI), I have realized that Expectation Management is a prerequisite for Change Management.
In this context, we are talking about expectations of an organization from an AI platform, what it can do and what it can’t!
I have developed this belief that the onus of Expectation Management is with both the AI Technology Partner (like Ai Palette) and the product and consumer insights teams of your organization.
Hence, I felt a need to share my thoughts on this topic through this blog – My Encounters with Limitations of AI and ML in the Food and Beverage Industry: Lessons in Expectation Management.
In this blog, I have thrown some light on three critical expectations from AI and have shared a framework that may help you with expectation management.