Optimize a Reasoning Ontology Map for Your AI Expert Graph

Using InfraNodus, you can build AI experts with the knowledge graphs that represent a reasoning ontology — a system of rules that tell your LLM not only what to think, but also — how to think.

Once we represent the reasoning logic as a graph, we can apply various metrics from network theory to optimize the graph structure. This helps us build reasoning maps that avoid tunnel vision and bias, on the one side, while also avoiding dispersal and low coherence, on the other. 

The most optimal graph structure has a medium level of modularity (distinct communities of concepts will be present and interlinked globally) while the distribution of influence (measured with betweenness centrality) across the different concepts will be more or less equal, without too much bias towards the central concepts. 

InfraNodus can be used to help optimize such structures:

• adding more connections between the prevalent clusters when the reasoning ontology graph is too dispersed
• developing less represented clusters (and relations between them) when the reasoning ontology graph is too biased

For instance, consider this example of a reasoning ontology for a popular negotiation framework "Getting to Yes":

While this framework is relatively coherent (all the main ideas are there), we can see that the level of Topical Diversity (and Semantic Variability, by extension) is Low. 

The problem is that there's too much focus on the concept of "principled negotiation", so the reasoning framework is too focused on that central idea — all the rules converge towards that concept, so the framework may lose more nuanced, smaller cluster of ideas (such as "Fair Standards" or "Negotiation Power" and "tricks"). 

In order to optimize this structure, we can use the Content Gap feature and the built-in AI: Question to Diversify function to generate research questions that develop and bridge less represented clusters — this helps us shift the weight to the other important topics in the discourse that we mentioned above. 

After several iterations, the graph's structure is Optimal — we now have a more even distribution of influence across the different clusters and more connections between other important secondary ideas that play an important part in this negotiation framework.