References (list) - juergenlerner/eventnet GitHub Wiki

This page provides a commented literature list of own work on relational event models (REM) or relational hyperevent models (RHEM), using the eventnet software or (published or unpublished) predecessor software in its empirical analysis. The listed papers are expected to complement the eventnet tutorials in providing more formal details, a better embedding into related work, and a more thorough discussion of objectives and contributions. They may also point to further illustrating use cases of REM and RHEM.

We emphasize that this is not intended to be a general overview of literature on REM - it exclusively lists own work. To get a better overview of the general REM literature, you may have a look at the references of the listed papers, search for indirect citations and for citations to these papers or references, etc.

The list below is divided into two parts, work on dyadic REM and work on RHEM, in each of which papers are roughly listed chronologically from older to more recent and/or clustered into further thematic sub-divisions.

Dyadic REM

Conditional event-type models

Brandes et al., 2009 propose models for networks of typed, weighted, or signed events by decomposing their probability into a REM for events of any type and a conditional event-type model, specifying a conditional distribution for the type of interaction among two actors, given that these two actors interact at all. The paper also proposes to let the influence of past events decay exponentially over time. Models are illustrated by an analysis of positive and negative interaction in international relations, among others testing the predictions of balance theory.

The above work is extended in Lerner et al., 2013 by controlling for the effects of politically relevant node and dyad covariates and providing a better embedding in international relations research.

Wikipedia edit networks (undo and redo probabilities)

A series of papers applies REM to interaction networks among contributing users of Wikipedia articles. Users can add new text, can delete the contributions of others ("undo"), or can revert previous deletions ("redo"). Togeher these "edit events" determine the evolution of the article's text. In this work we are interested in who typically makes contributions of whom undone (a "negative interaction" interpreted as disagreement) and who typically defends whose contributions against deletion (a "positive interaction" interpreted as agreement). Explanations range from simple network patterns (e.g., signed degrees of users), global reputation of users, signed repetition or reciprocation, over to the predictions of balance theory. Lerner and Lomi (2020a) apply case-control sampling to sample from the large space of "non-events", that is, user dyads on which undo events could have happened but did not.

Wikipedia edit networks (explaining article quality)

Building on the above articles on the Wikipedia edit network, the following two articles explain the quality of a Wikipedia article by the interaction structure of its team of contributing users, quantified through REM effects in the edit network. A notable finding is that teams adhering more to the predictions of balance theory typically produce articles of lower quality, which may be interpreted as pointing to a "price of polarization" in team work. Seen from a higher level, this work seeks to explain the "outcome" of team work by the structure of team interaction networks.

Wikipedia: allocation of attention

Complementing the above papers on the Wikipedia edit network, which model user interaction separately by article, there are two papers seeking to explain "who contributes when to which article". These papers propose REM for the two-mode network of users and articles, in which relational events indicate that a user edits an article at a given point in time. Lerner and Lomi (2018) assess interdependencies between contributions to editing articles and contributions to the articles' talk pages. Both papers apply case-control sampling. The main contribution of Lerner and Lomi (2020b) is a series of experiments seeking to determine the limits of how small a sample could be used to get sufficiently reliable parameter estimates from a network of dozens of millions of nodes and hundreds of millions of events. (In most empirical applications, analysts would typically choose larger samples, determined by available data and/or computational ressources, since there is no need to reduce the sample size to such small numbers.)

Self-organizing attention networks

Being technically related to the two papers above, but much more focused on the theoretical and substantive contribution, Tonellato et al., 2023 theorize that the structure of self-organizing attention networks results from four mechanisms: focusing, reinforcing, mixing, and clustering. This theory is empirically tested by REM fitted to sequences of attention allocation events in an open-source software community.

Animal collective behavior

The article below applies REM to analyze data from a feeding experiment with wild jackdaws that have been randomly assigned to two groups. Individuals appearing at specially prepared feeders with a member of their own group get access to high quality food (whereas they get low quality food if they turn to the feeders alone and no food at all if they arrive with a member of the other group). REM have been used to analyze how interaction probabilities change as individuals learn which peers provide the best reward. As a novel methodological contribution, the article designed a variety of permutation procedures to generate controls (i.e., dyads that could have interacted, but did not).

RHEM

Relational hyperevents are time-stamped events connecting a varying and potentially unbounded number of nodes (actors, items, papers, ...) at a time. RHEM have first been published in the non-peer-reviewed preprint Lerner et al., 2019. This preprint may still serve as a concise document providing the core definition of (undirected and directed) RHEM. Different topics from Lerner et al., 2019 have been elaborated in subsequent publications: Lerner et al., 2021 for the analysis of meeting events from the Thatcher contact diaries, Lerner and Hancean, 2023 for the analysis of coauthoring networks, and Lerner and Lomi, 2023 for directed RHEM.

Undirected RHEM for meeting events

The two papers below specify RHEM for undirected hyperevents (that is, events that may have any number of participants but no distinction into "sources" and "targets") representing meetings or social gatherings. Lerner et al., 2021 illustrates RHEM with the meetings events from former British Prime Minister Margaret Thatcher's contact diaries, which was the initializing application motivating the development of RHEM. Lerner and Lomi, 2022, analyzes the canonical Davis, Gardner, and Gardner "Deep South" data, suggests RHEM as a dynamic model for Breiger's "duality" in the case of time-stamped events, proposes ways for visually linking model results back to the original data, and demonstrates that RHEM do not only scale up to big data - but may also scale down to small event networks.

RHEM and RHOM for coauthoring networks

Lerner and Hâncean, 2023 apply RHEM for coauthoring networks and propose relational hyperevent outcome models (RHOM) to explain the outcome (in this case, scientific impact measured by normalized citation counts as proxies) of publication events. Outcome of past publications may have an influence on future scientific collaboration and, vice versa, the structure and outcome of past publications may explain the outcome of future publications of the same or overlapping coauthor teams. In contrast to the work in Jürgen Lerner and Alessandro Lomi (2019a), listed above, where "outcome" were measured at the network level, in RHOM, outcome varies by event.

Covid-19 contact-nomination networks

The following two papers by Hâncean et al. propose directed RHEM for case-contact networks, applied to data where positively tested persons ("cases") report a list of other persons ("contacts") they have met during the last few days. Such network data can potentially be used to uncover or test hypothetical pathways of infections. RHEM can potentially solve the issue that contact nominations are inherently non-independent and also allow to test for "network effects", such as repeated co-nominations or reciprocated nominations.

RHEM for multicast communication networks

Lerner and Lomi, 2023 propose directed RHEM for multicast communication events, such as email messages that have one sender and any number of receivers. The paper elaborates the differences between more general "hyperedge covariates" possible in RHEM and "dyadic covariates" proposed in the work of Perry and Wolfe, 2013.

Diviák and Lerner, 2025 apply the model from Lerner and Lomi, 2023 to study multicast communication networks related with three corruption cases.

A much older form of multicast communication networks, namely letter exchanges and mentions of actors in letters, has been analyzed and modeled with RHEM in:

RHEM for co-offending networks

Bright et al., 2023 propose RHEM for co-offending networks, that is sequences of events in which varying number of actors (offenders) jointly commit crimes.

In Bright et al., 2024, RHEM are applied to explain the offense types (crime categories) of group or solo offenses with the goal to test whether group offenders are more versatile than solo offenders and whether group offenses facilitate social learning in which newcomers "learn" to commit new types of crimes from more experienced group members.

While the previous article explains the crime categories of offense events, the following one "looks at the other side of the coin" and seeks to explain the (group of) offenders of crime events with an emphasis on explaining re-offense probabilities, among others estimating transition probabilities from one crime category to another.

The following article introduces a variation of RHEM for two-mode, geo-located co-offending networks, where crime events have participants (first mode) and geographic locations (second mode). Several new effects are proposed that test and control actors' tendency to repeatedly commit crimes in the same locations, adjacent locations, or nearby locations. The model can also be applied to other data on events associated with geographic locations.

Coevolution of collaboration and references to prior work

Variants of RHEM for modeling the coevolution of coauthoring networks and citation networks, as discussed in this tutorial, are proposed in:

This model has been applied in the more substantively oriented paper

Moreover, the same model family has been applied in a study of cultural production to analyze collaboration and use of shared stylistic references of filmmakers by: