Beyond Correlation: Identifying Direct Influences in Networks

New method by Jan Nagler and colleagues identifies direct influences from time-series data
Data-driven decisions often require more than identifying associations. What matters is which factors cause change. How do different regions of the human brain influence one another? Which genes directly affect others? And how are changes in ecosystems linked to processes in the atmosphere? The challenge is identifying which components directly influence one another and which are connected only indirectly.
Shuai Li from Beihang University, Wei Chen from Beijing Normal University and Jan Nagler from the Centre for Human and Machine Intelligence at Frankfurt School of Finance & Management have developed a mathematical method to identify direct connections in complex networks more accurately from time-series data. The study, entitled “Predicting Physical Links in Networks”, was published in the scientific journal Physical Review X in 2026.
Complex networks consist of numerous components that influence one another. For example, if one component influences a second, which in turn affects a third, a relationship may also be observed between the first and third components. However, this relationship arises through an intermediate component rather than a direct connection. Conventional analytical methods can mistakenly interpret such relationships as direct links, leading to an inaccurate picture of the network's underlying structure.
To address this challenge, the researchers developed a method called InVaXMap (Incremental Variable Construction-based Cross Mapping). It takes into account the time delays with which influences spread between components of a system, helping to distinguish direct from indirect effects. The approach is based on the assumption that direct effects often occur more quickly than those transmitted through several intermediate steps. Rather than considering all available variables at once, the method gradually selects those most relevant to the analysis. This distinguishes direct connections from indirect effects and reduces the risk of false identifications.
The research team first assessed the method using a range of mathematical models in which the actual connections between components were already known. They then applied it to real-world datasets from different research fields. The results show that InVaXMap achieves greater accuracy than the comparison methods, particularly in networks with around five to twenty components. The method also showed accuracy gains in tests on larger model networks with up to 100 components.
One application is brain research. The researchers analysed brain activity data from people with alcohol dependence and a control group. The estimated influences within and between brain regions were substantially weaker in the group with alcohol dependence. These findings are consistent with previous research on the effects of alcohol on brain function.
Another application involves gene regulatory networks. Here, the challenge is to identify which genes directly influence others and which interactions occur through additional genes. When analysing the DREAM4 test datasets, InVaXMap identified known connections more accurately than the comparison methods.
InVaXMap was also applied in environmental research. Using measurement data from four sites in China, the team examined interactions between ecosystems and the atmosphere. The analysis revealed differences between locations and across seasons. At all four sites, the number of identified connections was lowest during summer.
The method’s accuracy depends, among other factors, on the quality and length of the available time-series data. It also assumes that all relevant variables are observed and that interactions and their time delays remain constant over the period studied. In certain systems, indirect effects may occur more quickly than direct ones. In such cases, the method may fail to identify genuine connections.
InVaXMap provides a more accurate picture of direct interactions between individual components from time-series observations, including where experimental interventions are difficult or impractical. For Frankfurt School, the underlying question is also relevant to risk management and AI-supported decisions: Which factors drive change, and where could interventions be effective? Applications to financial or business data were not investigated in this study.
Jan Nagler
