@c0rrad0.bsky.social presenting our work on the Russia-Ukraine war on social media #ic2s2

Causality center stage at #ic2s2 with opening keynote by @eckles.bsky.social

ALICE is surprisingly fast and can be used in a number of application scenarios, not only for graphs but also for transactional and sequence datasets. For example, the number of frequent itemsets in this dataset is perfectly preserved under ALICE, but not under the standard model

Compared to the standard "configuration" model that preserves the degree sequence, ALICE preserves a number of interesting additional properties: the number of caterpillars (paths of length 3) in the graph, and the degree assortativity.

We also get estimates of the latent variables (which change over time) for free! This is something that simulation-based calibration cannot do. Our estimates are quite accurate even in high-dimensional spaces (here for the initial vector of opinions of 100 different agents).

Our estimates are maximum likelihood estimates: we write the likelihood function of the ABM and then maximize it via gradient ascent. To do so, we transform the ABM into a probabilistic graphical model, with observed and latent variables. Details on how to do it are in the paper.

We look at an opinion dynamics model: the well-known bounded-confidence model (BCM). Estimating the confidence bound parameter produces 4 times more accurate estimates than the ones by a traditional method of simulated moments (MSM). The method is also up to 200 times faster!

Have you ever calibrated an Agent-Based Model? Typical simulation-based solutions can be slow and inaccurate. With
Jacopo Lenti and Corrado Monti we show that an alternative likelihood-based approach is faster and more accurate. #WSDM2024
arxiv.org/abs/2310.02766