Building an Enterprise-Wide Knowledge Graph

At Foodpairing, we have been long-lasting advocates of the guidelines provided by McKinsey & Company and especially by McKinsey Digital, the division helping businesses transform their operations, processes, and strategies by leveraging the latest digital technologies and innovation. In their “playbook”, McKinsey Digital defined six enablers for successful AI implementations at scale (see figure below). 

The six enablers of AI implementations, according to McKinsey & Company

These enablers constituted the basis for our enterprise-wide project on developing and deploying a semantic Knowledge Graph (KG) that integrates the various diverse data sources we have been working with. Below we briefly showcase how the six enablers facilitated our overall process and how we adhered to the encompassed guidelines.

Strategic Road Map

Amongst our top underlying priorities was the strategic scaling of our AI efforts through the deployment of a semantic KG. The initial phase of this process involved identifying priority use cases and promoting horizontal collaboration among the business, data-science, and engineering teams. The KG design and development efforts were linked to Foodpairing’s core business objective of rapidly adapting our AI-powered analysis approaches to intelligently fill in the gaps of product analysis and increase the speed and quality of the outputs to our clients. This effort constituted a critical building block for our strategic roadmap.

Talent & Organization

The immediate next step was to acquire in-house expertise to lead this initiative, retaining thus full control over all IP, instead of investigating commercially available solutions that are challenging to customize or partnering with third parties that may compromise IP. The newly hired experts are now members of the Core Data and Product Enablement Teams, and, due to the rich and diverse professional backgrounds existing within Foodpairing, they can closely collaborate with our in-house culinary experts and food scientists for designing and validating the KG, as well as with data scientists, engineers and infrastructure architects from the Platform Team for scaling up the pilot use cases.

Technology

Implementation-wise, our strategic decision was to have the KG act as a non-disruptive semantic data integration layer on top of the existing data landscape, unifying vast volumes of diverse data (see also next subsection) under a common framework. All underlying technologies are W3C recommendations belonging to the Semantic Web tech stack, which means that we rely on globally established open standards and that we are vendor agnostic when it comes to the implementation and persistence of the KG.

Data

The Foodpairing KG ingests vastly heterogeneous data generated in-house through various chemical and Machine Learning analysis initiatives as well as data being continuously collected online (see figure below).

Data building blocks comprising the Foodpairing KG

(Sub)sets of the data are being used interchangeably in multiple use cases, ensuring thus a substantially bigger impact than a single initiative. We are now in the process of establishing a set of data governance policies, according to McKinsey Digital’s principles, that will ensure that both existing as well as newly generated data are ready for use, are of high quality, and trustworthy.

Adoption & Scaling

Having a massive endeavor in mind, such as an enterprise-wide Knowledge Graph, adoption and scaling were always within our critical considerations. Regarding adoption, we made sure to involve multiple stakeholders and experts from within Foodpairing even from the start of the design phase, to actively participate in all the major strategic design decisions underlying the KG. Moreover, ensuring that the deployed use cases were integrated seamlessly into the rest of the workflows within Foodpairing, guarantees the wider adoption of the KG solution by the rest of the company. 

On the other hand, our constant aim was to prioritize the scalability of the most important use cases from the design phase and to prioritize use cases for deployment based on their value after scale-up. Since the lack of proper data integration is typically a significant challenge in scaling AI-based use cases, we made sure that proper mechanisms are in place for ensuring the proper semantic data integration via the KG.

Agile Delivery

No matter how ambitious our end-goal is, we faithfully follow an agile approach during the development of the KG, which is driven by biweekly sprints that pave the way for gradually building the foundational components of the Knowledge Graph from the bottom up. As already mentioned, we initially focused on the smaller (and more urgent) use cases first, aiming at developing each time the minimum viable graph (MVG) per use case, as recommended also by Gartner. Several months later, the Foodpairing KG has reached sufficient maturity to drive the extraction of deeper customer-driven insights.

Recap

Designing and deploying an enterprise-wide Knowledge Graph, especially one with the size and diversity of the Foodpairing Knowledge Graph, is a non-trivial endeavor, which needs to adhere to well-established principles in order to succeed. Combining the rich and diverse expertise within the organization and the wealth of data and knowledge, along with the well-defined guidelines by McKinsey Digital, we are confident that we are on the right track! Stay tuned for more exciting news to come soon!