New Bayesian method infers friend or foe from contact patterns

A team of researchers from the University of Melbourne and Université Laval has introduced a Bayesian framework to infer signed social networks—networks that label relationships as positive (friendship) or negative (antagonism)—from indirect observations like proximity data. The challenge is that zero interactions can mean either that two people never had a chance to meet or that they actively avoided each other. The new method, detailed in a preprint on arXiv, uses MCMC (Markov chain Monte Carlo) inference to model interaction groups and separate chance from choice. On synthetic benchmarks, it significantly outperforms natural baselines, especially in detecting negative edges.

When tested on real-world contact data from a French high school, the model recovered a social structure that closely matched results from traditional friendship surveys. The work has implications for understanding group dynamics, workplace interactions, and even online behavior, where passive data streams (e.g., co-location logs) can now reveal who likes whom and who steers clear. The code and data are publicly available, allowing other researchers to apply the method to their own contact datasets.