On the ?sthetics of Knowledge Graphs

For many of us, a knowledge graph is a thing of beauty – an esthetic object – that inspires both wonder and awe. This in addition to their many practical uses like data integration , data aggregation , and taming LLMs .

Why is that??

In one classic treatise, Dutton (2002) identifies the signature dimensions of esthetic objects:

• Esthetic objects display recognizable virtuosity:? the care and expertise of their creation is clear to anyone who stops to contemplate them thoughtfully. Knowledge graphs are not heaps of facts but a finely woven embroidered tapestry.? Experts will focus evaluation of knowledge graphs on their coherence and parsimony rather than their scope and coverage, in keeping with a philosophy of better data is better than more data. This focus enables much more reliable inference and reasoning that more data does not.
• Esthetic objects are created and appreciated apart from their usefulness. Many of us who build knowledge graphs, ontologies, and taxonomies are motivated more by excellence ("data quality") than by usefulness. This of course creates conflict with managers who are driven by speed and coverage rather than by quality and coherence – which cultivates technical debt. We can resolve this conflict by focusing the most careful curation on the most crucial, most impactful keystone concepts and scaling with automation along the long tail concepts.
• Esthetic objects leverage rules of composition and knowledge of the rules enhances our appreciation of these objects. Knowledge graphs systematically demonstrate these rules through the coherence and accuracy of the relations they document. Craftsmanship in our domain focuses on the design and interplay of these rules of composition that guide how to build and relate facts. Rules of composition – the knowledge architecture – lay bare the clean lines of structure hidden in the shifting sands of accumulated facts.? Careful curation of these rules – the knowledge graph design – is a key component of automation.
• Esthetic objects invite criticism in that they provoke judgements, positive or negative. Many people react viscerally to gaps and errors in knowledge graphs. As Cassie Kozyrkov phrases it, inherited datasets (and knowledge graphs) are like inherited toothbrushes : using them is an act of desperation. If you've ever presented a draft taxonomy to a mixed audience then you will be painfully aware of this. Each team in your audience will have a different perspective on, hence different priorities for, what should be included and how it should be organized.? We can resolve much of this conflict by using knowledge graphs instead of canonical taxonomies which encapsulate a single, frozen point of view.
• Esthetic objects simulate experiences of the world. Knowledge graphs simulate our experience of knowing about the world and of manipulating that knowledge in reasoning. In this sense they are very different from the early LLMs and transformer models that simulate only how strings co-occur with other strings. Many tasks focus narrowly on very small parts of the world, so even small knowledge graphs can add significant value .
• Esthetic objects are often set aside from ordinary life to create drama and emotional response by focusing on the individual experience of a tiny part of our reality – a scene in a painting, an auditory experience at a concert. Large-scale knowledge graphs, on the other hand, create wonder and awe by focusing on the scope and grandeur of what we know as a community, not as an individual.? This is reflected in the use of knowledge graphs as external resources (via RAG) set aside from LLMs rather than merged indiscriminately in training .

• Esthetic objects are created and consumed in an imaginative experience rather than direct perception. Knowledge graphs faithfully represent only some of our experiences of the objects in the world. Some features and facts are omitted according to the perspectives and priorities of how we imagine the world in a given context or for a given task. Other features we can infer leveraging rules and patterns over incomplete knowledge – another aspect of imaginative experience. These operations on knowledge graphs – reasoning and imagination – are clearly key topics for future research.?

The wonder and awe that knowledge graphs often inspire derive to a great extent from their nature as esthetic objects or works of art.

Ironically, treating knowledge graphs as esthetic objects rather than heaps of facts leads us to build higher quality resources that are much more useful and reliable in practical applications.