The term “Knowledge Graph” (KG in short) was made popular by Google in their now famous “things, not strings” post from 2012, where they were introducing “an intelligent model that understands real-world entities and their relationships to one another”. This component operates in the background of the Google Search Engine and helps users to locate more effectively the right pieces of information during their search.
Since then, several other “big players” have joined the “KG wave”, like Facebook, LinkedIn, Microsoft, Amazon, Ikea, and many more. The respective global market size is estimated to reach $7.22Bn by 2028 from $1.9Bn in 2021, at a spectacular Compound Annual Growth Rate (CAGR) of 22.1%.
In the 2022 Gartner Hype Cycle™ for AI (see below), KGs are now exiting the Peak of Inflated Expectations and are gradually becoming more and more mature, playing a central part in Data-centric AI, which focuses on enhancing and enriching the use of data available by enterprises. In fact, Gartner predicts that, by 2025, KG technologies will be used in 80% of data & analytics innovations (up from 10% in 2021), facilitating rapid decision-making across the organization.
Under the Hood
The underlying idea revolves around representing data as graphs, with nodes representing the entities in a domain of discourse and edges representing the entity interrelationships.
For instance, the following diagram illustrates two sample recipes (gray nodes) surrounded by their respective ingredients (green nodes) as well as additional nodes representing the types of the entities in the graph (i.e., “Recipe”, “Ingredient”, etc.).
Enhancing Data in the CPG Industry
Compared to more traditional models (e.g., relational DBs), there are two additional dimensions that a KG brings into play, offering substantial value to modern AI-powered enterprises like Foodpairing that handle vast volumes of data:
- The graph dimension permits running graph and network algorithms on top of the data stored in the KG;
- The semantic dimension allows assigning meaning (i.e., “semantics”) to the entities and relationships in the KG, facilitating the execution of more sophisticated queries over the data, and providing access to deeper and more meaningful insights.
Additional benefits that KGs bring into play include the fact that they can represent information in an intuitive and human-friendly fashion closer to human cognition. Moreover, KGs are extremely well-suited for joining up heterogeneous data from diverse sources, offering a platform that can very well harness the inherent linkages and context existing within a graph.
In a nutshell, KGs can integrate data with knowledge at scale, making, thus, data smarter and turning it, as Michael Atkin aptly put it, “from a problem to be managed into a resource to exploit”.
The Foodpairing KG
Capitalizing on the benefits KGs bring to the table, the Foodpairing arsenal of AI tools naturally includes our own KG that integrates the diverse data we have been generating in-house through the various chemical and Machine Learning analyses, along with data being continuously collected online under a common framework.
A set of connectors analyzes the information from the input sources and ingests it into the Foodpairing KG, which currently encompasses:
- Over 15K unique food and non-food ingredients, categorized under 141 distinct ingredient categories that are organized within a 5-tier taxonomy. All the ingredients are accompanied by their chemical composition, functional benefits, and mood states associated with them, according to the literature.
- Tens of thousands of SKU products launched worldwide in the last decade, along with their ingredient lists, respective markets, companies, and brands.
- Around 10M recipes, categorized under a multi-level taxonomy of 828 dish types.
- Several billion time-stamped social media posts, linking ingredients to dish types and other entities of interest (e.g., locations, target audiences, etc.). This information leads to assessing the evolving trends of (combinations of) ingredients in certain dish types, like, e.g., the comparative evolution of the use of blueberry in cocktails, smoothies, and ice creams (see graph below).
- Over 8K links to Wikidata food-related entities, which leads to retrieving valuable additional information, e.g., country of origin of ingredients or dishes, alternative use and fabrication method per ingredient, or even the most popular colors found in ingredients (hint: red and yellow!).
As it currently stands, the multi-faceted Foodpairing KG generates deeper insights and holistic knowledge from the entirety of the stored information and allows us to:
- Intelligently fill in the gaps of “traditional” product analysis; e.g., discover the most popular vegetable or fruit ingredients in a market, retrieve highly popular or novel combinations of ingredients, etc.
- Rapidly adapt our AI-powered analysis approaches to other (non-food) domains – knowledge transfer – and map the product landscape, independently of domain;
- Perform a smart selection of and navigation over combinations of ingredients;
- Generate recipes that match the flavor profiles of specific target audiences;
- Recommend recipes complying with health parameters (e.g., obesity), dietary restrictions, or other constraints (e.g., allergens, sustainability, etc.).
As discussed in this post, the Foodpairing Knowledge Graph is a central tool in our quest for facilitating CPG organizations to launch better products, faster.
Containing unique in-house generated data, and linked with millions of data points from our CFI platform, the Foodpairing KG generates invaluable insights and offers a critical advantage to our partners wishing to create successful products and achieve business growth. And the best part? We have just only started to scratch the surface!
Stay tuned for more exciting KG news to come!