Cognitive Diversity: How to Measure and Reduce Bias in Your Thinking
InfraNodus has a built-in recommender system for cognitive diversity modulation. It identifies the network structure of your thought or discourse — biased, focused, diversified, or dispersed — and helps optimise its diversity through gap analysis.
The approach is based on the assumption that an optimal network structure should be heterogenous but not dispersed. Therefore, if your discourse network is too biased, InfraNodus will highlight and steer your (and your LLM's) reasoning towards the less represented clusters of ideas. If it's too dispersed, InfraNodus will direct the focus towards the topics that recur most frequently but have gaps between them, in order to help make it more coherent.
If a discourse is focused, InfraNodus will steer attention towards the mid-sized clusters to introduce more diversity while maintaining focus. If the discourse is diverse, InfraNodus will maintain that diversity by focusing on the biggest gaps in order to keep optimal coherence and to bridge the structural holes that exist in the network.
In the end, the objective here is to bring the network of ideas to its optimal state through a redistribution of influence via attention.
Benefits of Cognitive Diversification
This approach can be very useful for modulating collective discussions, organising one's thoughts, or better understanding how to develop an existing discourse.
For instance, many neurodivergent behaviours can be characterised by obsessive loops or, alternatively, more scattered forms of thinking. While these patterns may be beneficial in some contexts (obsession for problem-solving, scattered thinking for creativity), there are also multiple instances where such behaviours may be not very desirable.
That's when cognitive diversity modulation can play an important role in equalising the structure of thought and ensuring that — on the one side — no single cluster of ideas takes over (too much bias), and — on the other side — the ideas do not become incoherent (too much dispersal).
This is achieved through a redistribution of influence and attention towards the less represented clusters and gaps, to ensure that the structure of thought stays coherent but diverse.
How Network Science Is Used for Cognitive Optimization
We use a network representation of thought and knowledge graphs to create a structural representation of the relations between ideas and to identify their community structure. Then we apply advanced graph theory algorithms to detect the main nodes and their relative influence — based on betweenness centrality: how well they connect distinct clusters of nodes in the network [2].
We then run community detection algorithms (Blondel et al.) [1] to identify the clusters these nodes form, measure these clusters' modularity, as well as their relative influence based on the betweenness centrality of the nodes that belong to each community.
This allows us to estimate how equally influence is distributed among the different topical clusters and to redistribute it by directing the user's focus or the LLM's attention to the clusters and gaps that could be further developed for maximum optimisation.
There are four different states that can be detected with this approach:
- biased
- focused
- diversified (optimal)
- dispersed
In our article on the network diversity index we describe in detail how this algorithm works.
Biased Network
The objective here is to reduce bias, and this can be done by activating the less influential clusters and bridging the gaps between them.
If the top clusters or nodes have too much influence, the network is tagged as "biased". InfraNodus then identifies the gaps between the existing clusters and chooses the gap whose clusters have the least total influence first (measured via betweenness centrality — which means they are disconnected from the rest of the network). It then reiterates through the remaining gaps that contain either of those two least-influential clusters and takes the cluster from the gap with the least total influence. The result is a triad of clusters that are least represented and also have gaps between them.
This helps shift attention towards the less prominent parts of the discourse. When used via the API / MCP or InfraNodus' AI chat module, InfraNodus' internal AI engine will add these least represented clusters into its context before generating a response that attempts to bridge them.
Focused Network
The objective here is to maintain focus and to not allow it to slip into a biased state, by promoting diversity.
In order to achieve this structural goal, InfraNodus reiterates through the existing structural gaps and chooses the gap between the least influential clusters first. It then reiterates through the remaining gaps that contain either of these two clusters and picks the one that has medium total influence (betweenness centrality). This ensures that the resulting triad, once activated, will be connected to the network through at least one cluster — which helps maintain focus and, at the same time, introduces diversity.
Diversified Network
In the diversified state, the objective is to maintain that state by not letting any of the dominant clusters gain too much influence, while also maintaining coherence by bridging the less represented clusters to the main structure.
In the diversified state, this is achieved by generating a triad of clusters that are represented in the biggest structural gap in the network, and then reiterating through the remaining gaps that contain at least one of the clusters in the biggest gap and where the other cluster has medium total influence.
Dispersed Network
In a dispersed network, the objective is to increase coherence. We need to bridge the clusters that already have high total influence, while also linking them to the less represented smaller clusters of ideas.
In order to do that, we reiterate through the biggest gaps to find the combination of clusters with the highest total influence (to support coherence). We then scan the remaining gaps that contain at least one of the clusters we have chosen, in order to pick an adjacent cluster with the lowest influence and generate the triad of insight clusters. This avoids the formation of orphaned clusters and ensures global connectivity.
References
- Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. (2008). "Fast unfolding of communities in large networks." Journal of Statistical Mechanics: Theory and Experiment, 2008(10), P10008.
- Freeman, L. C. (1977). "A Set of Measures of Centrality Based on Betweenness." Sociometry, 40(1), 35–41.
- Paranyushkin, D. (2019). "InfraNodus: Generating Insight Using Text Network Analysis." The World Wide Web Conference (WWW '19), 3584–3589. See also: infranodus.com/about/research-framework
Try It Yourself
Measure and modulate the cognitive diversity of your own thinking, discourse, or knowledge base using InfraNodus:
Sign Up Log In
Consulting Services
Contact us if you are interested in applying cognitive diversity modulation to your research, your team's collective discussions, or your organisation's knowledge management.
Contact Us