when attending the ICRC 2025, my first large in-person conference, i was struck by how hard it is to choose from among 5–6 parallel sessions. of course, sometimes you do have a clear preference, e.g. when a session is directly relevant to your research. but oftentimes it’s not so cut-and-dry, and you find yourself lost, torn between several poorly known but vaguely interesting topics.
technically, the issue is one of ranking the sessions by their interestingness $I$. presumably, $I$ is additive, so $I_s = \sum_{t \in s} I_t$. presumably then, one could carefully read through all the abstracts, evaluate them, assign $I_t$ to every talk, add them up, and pick the session with the largest $I_s$. however, this method is
it is, however, much easier to choose between two given talks. the only information such a choice conveys is $I_{t_1} > I_{t_2}$. this statement is also not strict: for example, you might misread/misunderstand one of the talks at first, then come around to actually preferring it. so it’s more of a probabilistic comparison: “right now i feel like $I_{t_1} > I_{t_2}$.”
there is a host of algorithms for converting such pairwise probabilistic comparisons into a ranking, the most famous being the Elo rating system. i opted to use a more recent algorithm from the gaming world, TrueSkill. it models $I_t$ as a normal random variable and uses Bayes’ theorem to update it after “conducting a match” (asking the user to choose one of the two talks).
the resulting workflow is as follows:
in practice, it’s a fun way to get a quick glance at the talks and force yourself to consider all the possibilities. are you sure you don’t want to learn about axion star mergers in the pi-axiverse? maybe the latest advances in random forests for proton-$\gamma$ separation methods in IACTs will pique your interest? in some cases, i have genuinely surprised myself, and the script did a good job of pushing me out of my default cosmic rays session.