Using Network Science to Understand Human Thinking
Imagine going to a dinner party at a friend’s house. You might imagine being welcomed into their home and sitting around the dinner table. Perhaps before coming you pick out a nice bottle of wine, or a desert to share. You likely will not eat anything before you arrive because you’re anticipating being served dinner. From the minute you receive the invitation to the dinner party you understand the activity that you are being invited to participate in.
Cognitive Scientists have been able to learn a lot by having people describe simple activities (like going to a dinner party). Though this seems like a simple task for participants, looking at these descriptions give Cognitive Scientists rich data sets that can be analyzed with tools from Network science. Network science is an important area at the intersection of Cognitive Science and Neuroscience. One major push in Cognitive Science has been to develop Large Language Models (LLMs) such as ChatGPT that can be used to perform numerous functions. Although this push has been tremendously successful, there are other exciting ways to utilize Network Science. This article flips the script on traditional network science models; instead of using LLMs to try to make networks think like humans, this article uses network science models of human data to try to learn more about how HUMANS think.
What is Network Science?
Network science is a broad term used to describe a way of thinking about data. For instance, you could understand communication in a group with a network science approach, or make inferences about how to boost traffic on your website. Network science describes data as collections of nodes and edges. In these models, nodes are the units, and edges are the connections between these units. Edges describe how units relate to each other, for example, how you might get from one unit to another, how often two units might co-occur, or the order in which units might be organized. Any data set that has nodes and connections between them can be described with network science. For instance, in describing communication in a group, nodes could represent people in the group and the edges could be the number of text messages exchanged between them. Or, in the case of website traffic, a node could be a webpage and the edge could be visits from one webpage to the next.
Broadly, network science allows us to understand the patterns in large data sets by describing what is happening in the connections between nodes. In the case of the dinner party example, each activity that you generated when you imagined the dinner party would be one node in the network. The order in which the activities were generated would be used to connect the events. Thus, the entire activity of going to a dinner party can be described as a set of activities (nodes) that take place in a specific order (edges).
This is exactly the approach that, Kevin Brown and the rest of his team at Western University pioneered in their article. The team used network science to describe how we, as humans, tend to think about the things that have happened to us and how we imagine the events of our lives.
How do Networks Model Event Data?
The data for this project came from human participants who were given a simple task: Order this set of events in the order that you would do this activity. For example, a participant might be assigned to the event ‘serve nice wine at dinner’, and presented with activities like ‘research wine’, ‘open wine’, ‘get out glasses’, ‘make toast’. Participants created a list of ordered events indicating how they would perform that activity, from beginning to end, using the activities provided. From this data, the authors constructed network models for each event. Within the network models, each activity is a node, and the connections between the nodes were based on the order that the participant reported doing the activity. For example, if I decide the first steps to ‘serving nice wine’ are ‘research wine’ then ‘make toast’, then there would be a directed connection leading from ‘research wine’ to ‘make toast’.
After collecting ratings from 145 participants, the authors were able to build network that show how people in general tend to think about 80 different events. In the aggregate models (like the one shown in the image below), the nodes of the network still represent activities, but the connections between the nodes describe the consistency with which participants rate events as following when completing a specific activity.
What does this tell us about human thinking?
When you look at the shape of the resultant graphs one thing probably jumps out to you: these networks do NOT look linear! If every person who completed this task agreed on the order in which activities should be completed, then the graphs would be straight lines. Instead, in the dinner party event, we see activities with connections to an assortment of other activities. Turns out, even for a simple event like a dinner party, people prioritize different activities, leading to differences in the ordering. These differences create a web of connections, rather than a step-by-step plan that everyone would follow. This is an important finding and indicates that when we replay the events of our lives, we are unlikely to be remembering a perfect chain of events. Historically, some perspectives in cognitive psychology have suggested that event knowledge involves a replay of the event in the order in which it was experienced. These models of event knowledge suggest that when we think about events we replay them from beginning to end (like watching a home video). However, the network models from the current paper tell us that the story is a little more complex. At least some of the time, this is not the case! We do not always think of events as linear plans. Instead we might think of the most important element first, or our favourite event. Future research is needed to determine what other factors shape the way in which we imagine events.
Another thing you might notice looking at the network graph is that some events are grouped closer together than other events. For example, in the blue square, ‘make toast, ‘drink wine’, ‘smell wine’, and ‘taste wine’ are grouped closer together. One idea that the authors explored was whether these groups of events could represent “scenes”. A scene is a subgroup within an event that describes a part of the activity. For example, the events shown in the blue box could represent something like ‘enjoying the wine’, an activity within the broader activity of ‘having a dinner party’. There have been important insights in cognitive psychology suggesting that scenes are an important property of events. In this study, authors were able to identify many instances that they think reflect the use of scenes when thinking about activities. Authors identified these instances mathematically, by quantifying the ‘modularity’ of the event networks. Modularity refers to the extent to which the network has clusters in which events are closer to each other than the rest of the network; in other words, it is a mathematical way of looking for the clusters like the ones visually identified in the blue box. Locating clusters in some event networks represents an exciting observation which paves the way for new research that Kevin Brown, and his team in the McRae Cognitive Science lab, are excited to take on.
Take home message:
Overall, network science is a powerful tool to learn about, and to better understand the world around us. There is a lot of current work on the development and training of LLMs; scientists work to get these networks to perform everyday tasks like picture recognition, creating project plans, and writing papers for undergraduate classes. In conjunction with the development of sophisticated LLMs for the performance of everyday tasks, there are exciting opportunities within cognitive science to flip-the-script and use network science principals to better understand human data.
The current paper exemplifies the power of such approaches, demonstrating that by applying a network science framework we can gain new insights into human behaviour, and specifically into human thinking. In this paper, authors demonstrated that when we think about events we do not necessarily adhere to a strict linear replay. Further, authors found instances where events are organized into scenes. Both of these demonstrations speak to existing theories about how events are represented in the mind.
Original Article: Brown, K.S., Hannah, K.E., Christidis, N., Hall-Bruce, M., Stevenson, R.A., Elman, J.L., McRae, K. (2024). Using network science to provide insights into the structure of event knowledge, Cognition, (251), https://doi.org/10.1016/j.cognition.2024.105845