The interconnectedness of individuals is a key factor shaping social, economic, and demographic processes. Although much is known about how interactions within particular domains - such as households, schools, workplaces, and neighborhoods - influence individual outcomes, much less is understood about how these domains are interlinked, thereby constraining our understanding of the broader architecture of social interconnectedness.

In this presentation, we introduce the Complete Network of Sweden project (netreg.se), an initiative that leverages population-wide Swedish register data to construct large-scale networks of social exposure that integrate multiple domains of affiliation and kinship over three decades. After outlining the scope and anatomy of the underlying data—covering millions of individuals and billions of dyadic relations—we turn to one of the core questions of the project: how far apart people and groups are in the social fabric.

Using a Milgram-style, sample-based approach to estimate average shortest paths, we begin by examining how average distances between individuals have changed over time. Building on this, we explore an alternative approach using random walkers, a stochastic and arguably more sociologically realistic method, to capture how "far away" different categories of individuals are from one another. Finally, we present ongoing work on Threadle, a new C#-based command-line tool offering a memory-efficient and functionally enhanced alternative to igraph for storing and handling the large, multiplex, and hybrid-mode networks that register data entail.

The effects of social networks among the economically powerful individuals in a society are hard to substantiate, mainly because network data are hard to come by. For some research questions, it makes sense to approximate them by data on interlocking directorates, i.e., bipartite networks connecting individual directors to the companies on whose boards they serve. These data can be derived from annual reports that stock exchange noted companies are required to publish. 

In my presentation I will sketch an actor-oriented model for the co-evolution of interlocking directorates and company policies. Based on historical data from the Amsterdam stock exchange, the model will be used for testing whether the adoption of takeover defense measures diffuses between companies via social contagion, i.e., by sharing board members.

Many real world networks come with an incredible richness in attributes, particularly on their nodes. Yet, this richness is often under-exploited, as node attributes are normally limited to provide ground-truth for community discovery, or they are used for simple analysis testing homophily. In this talk I will show how a few simple measures can open the door to analyze node attributes and their relationships with the edge structure in many more interesting ways.

In this short presentation I provide two examples of situations where ambiguity should be valued as substantive knowledge rather than noise to get rid of. Both examples have been implemented in Gephi Lite, an online tool to visualize and analyze networks. My first example is "connected closeness", a visualization and metric combination that quantifies the range at which a network layout is most successful. I contend that although the visual mediation of a force-driven algorithm is inherently ambiguous, measuring the scale at which it is the least ambiguous provides a remarkably stable characterization of the network structure. My second example measures and visualizes the ambiguity of community detection, at the level of nodes and edges. It helps contextualize some outlier nodes whose belonging to one particular cluster should not be taken for granted. Ultimately, I aim to illustrate that a blurry picture of a sharp thing should not be confused with a sharp picture of a blurry thing: network ambiguity often lies in the phenomena we study rather than in our imperfect apparatus.

Image networks are methodological constructs that map relations between images, their attributes, and their web contexts using publicly available online data. Nodes represent images and the researcher’s decisions about how to view them, from image metadata associations to AI interpretation. The different network types—such as image-hashtag, image-emoji, and image-AI outputs—can be as open-ended as the researcher’s creativity. As a Method, the visual inquiry of image networks is responsive to what precedes and conditions a network visualisation, beyond the topic itself.

This talk situates image networks within digital methods, outlining what informs the practice of visual inquiry—image origins, curation, what computational media do with image data, and layout choices. It provides insights from digital media research on Vision-AI gender bias, Generative-AI racial stereotypes of Black women and a longitudinal prompt method for interrogating LLMs outputs. The contribution foregrounds a technicity thinking-praxis for image networks as reproducible digital methods.

How can we visualise complexity without reducing it? This question drives the design of the Grounding AI Map, a public-facing network visualisation of over 2 million scientific papers that examines what algorithms are doing across diverse research contexts. Rather than confining our corpus into simplified categories or statistical summaries, we propose that the visual form itself becomes an analytical tool for reading epistemic structures in scientific knowledge.

Drawing on Latour's principle of irreduction, we designed the map to resist collapsing heterogeneous research into pre-defined subjects or clear-cut narratives. Instead, these methodological choices open the possibility for exploratory visual analysis, where ambiguity is preserved as a deliberate feature. By refusing reduction, the map transforms visual features into potential epistemic markers, rather than mere aesthetic elements.

With a focus on cluster analysis, we ask: What does cluster morphology tell us about scientific practice? Can we read visual differences as meaningful epistemic markers? What new stories about the organisation of knowledge become visible when we focus on network topology as a research instrument? With this approach, the Grounding AI Map invites further inquiry into how visual analysis might generate new questions for engaging with complexity in scientific networks.

How can we design annotation strategies that go beyond explanation, that open rather than close interpretation, in order to “narrate” large-scale networks? How can annotation become an epistemic practice that preserves complexity and invites exploration? These questions motivate our inquiry into how annotation can serve as a narrative device to make large networks legible while acknowledging

that the act of annotation entails reduction and trade-offs that have shaped our approach.

We explored these questions through the Grounding AI project, a public-facing network visualising over two million scientific papers to examine what algorithms are doing in scientific research. The network’s scale and complexity created barriers for non-experts to engage with and interpret the visualisation. To address this, we developed layered annotation strategies for both physical and digital

formats. These textual, visual, interactive, and AI-generated annotations create new modes of interfacing with the network, enabling viewers to contextualise clusters, navigate the structure, explore topics, and engage in reflection.

This presentation reflects on the design, challenges, and lessons learned from translating Grounding AI into a public-facing network. We highlight how layered annotation can act as a narrative and epistemic device, supporting exploration, interpretation, and dialogue between viewers, data, and design.