The Role of Geography in the Complex Diffusion of Innovations

We illustrate patterns of spatial diffusion and churn with unprecedented detail by analyzing data from the full life-cycle of a social media website in Hungary. This unique dataset—covering an online social network (OSN) that preceded Facebook and became the country's most visited website by the mid-2000s—enables us to track when geo-located individual users registered profiles and when they logged in for the last time.

This complete life-cycle perspective provides empirical evidence of non-trivial problems in spatial diffusion modeling, including the role of local network structure, settlement size, and geographical distance. These findings point toward necessary corrections for existing diffusion models and lead to novel understanding of collective behavior in adopting new online products—and crucially, how churn becomes collective at the end of their life cycle.

Our analysis reveals fundamental patterns in how innovations spread and eventually decline. We fit differential equation models to cumulative distribution functions (CDFs) of adoption and churn, analyzing the local features of these dynamics using statistical methods. A striking finding emerges: churn happened early in those towns that adopted the OSN early, suggesting that pioneering communities also experience fatigue or saturation first.

The spatial dynamics of diffusion evolved dramatically over the platform's life cycle. In the early stages, diffusion took place across large distances—early adopters in different regions independently discovered and adopted the platform. However, as the network matured, diffusion became progressively more local, spreading through existing geographic communities and social ties.

Most remarkably, we find that users churn when approximately half of their nearby friends have already churned, while distant friends remain active. This finding reveals that cascades of churn are fundamentally local phenomena. Users are more likely to follow the behavior of spatially close friends than distant ones, even though both types of connections exist in their social network. Geographic proximity amplifies social influence in abandonment decisions.

Animation showing the spatial diffusion and churn patterns across Hungarian towns throughout the network's life cycle. (Special thanks to Balázs Lengyel)

In the second part of our research, we test the spatial aspects of an agent-based social contagion model on our network. While the model describes global diffusion dynamics better than differential equation models, we find systematic failures in predicting local adoption timing. Predictions of when adoption peaks in specific locations fail in the majority of towns, with a surprisingly wide time range between early and late predictions.

By examining various characteristics of local networks, we identify the sources of these prediction errors. Predictions tend to be late in large towns where early adopters are only loosely connected to late adopters—the model underestimates the barriers to diffusion within large, fragmented urban networks. Conversely, predictions are early in small towns where network density and transitivity are high—the model overestimates how quickly adoption spreads in tightly-knit communities.

We also demonstrate that urban scaling of early adopters differs between reality and the model, indicating that current models don't properly capture how city size affects innovation adoption patterns. Furthermore, contagion in the early stage occurs more easily between distant locations in reality than in the model, suggesting that existing models underestimate long-range transmission mechanisms such as media coverage, travel, or weak ties between communities.

These findings emphasize the non-trivial challenges we face in correcting spatial diffusion models. The spatial distribution of early adopters—the main unknown in any prediction exercise—profoundly shapes diffusion patterns in ways current models don't fully capture. Improving predictions will require better understanding of how geography, network structure, and settlement characteristics interact to govern innovation adoption and abandonment.