Introduction

A relational event dataset records a sequence of interactions. A row might say that actor A sent a message to actor B at 10:03, followed by B replying at 10:05. A relational event model (REM) uses that order directly. For each observed event, it compares what happened with the events that were possible at the same point in the sequence.

This differs from fitting a model to an aggregated network. Counts and final edges alone cannot show whether a reply followed immediately, whether an actor became more active over time, or which alternatives were available before an event occurred. In a REM, those details enter through the event history, risk set, and time-varying statistics.

REMFlow exposes four functions for fitting and assessing these models:

  1. remify validates an event table and constructs its risk sets;

  2. remstats calculates statistics for every observed event and alternative;

  3. remstimate fits the selected ordinal, exact-time, actor-oriented, or duration model;

  4. diagnostics reports fitted probabilities, ranks, residuals, and related checks.

RelationalEventModel is the general high-level facade for untyped and typed tie-oriented event models and next-event probabilities. MisinformationModel inherits from it and adds action and stance conventions, actor-role scores, group-similarity measures, source ranking, and actor-blocking analyses. These domain-specific quantities do not identify causal sources or estimate the causal effect of an intervention.

Implementation choices

The default NumPy backend runs on the CPU. For model families with a JAX implementation, jax:cpu and jax:gpu place likelihood and derivative calculations on the requested device. Requesting jax:gpu without an available GPU raises BackendUnavailable.

Histories, statistics, estimates, and diagnostics are regular typed Python objects. They retain event and coefficient order, expose documented metadata, and support serialization where noted in the API reference. Formula strings are parsed by a restricted parser; arbitrary Python code is never evaluated.

See the usage guide for model-specific examples. The project’s README contains the same summary of combinations that are not yet available.