Queueing to publish in AI and CS

Summary

Written by David Martínez-Rubio in collaboration with Sebastian Pokutta. Does the common CS conference publication model with a fixed low acceptance rate over submissions make sense? What are some consequences of it? Here, I analyze some interesting properties that model the reviewing and acceptance system of machine learning conferences, but applies to CS more generally. $ \def\N{\link{#def_number_new_papers}{\color{black}N}}} $ $ \def\T{\link{#def_max_retries}{\color{black}T}}} $ $ \def\p{\link{#def_rate_of_acceptance}{\color{black}p}}} $ $ \def\xast{\link{#def_fixed_point_ideal_case}{\color{black}x^{\ast}}}} $ Disclaimer: This is a toy model and more knowledgeable people have devoted greater effort to other models and ideas [1, 2, 3, 4, 5], among many others. Below there is simple and to-the-point food for thought. I don’t know yet if I consider these conclusions based on simplified models to be valid for the real case. The ideal case: no giving up First, let’s assume that authors keep resubmitting their unaccepted papers indefinitely, without restrictions on how papers are accepted. Assume a sequence of non-overlapping conference calls (e.g. 3 per year), and each time $\N$ new papers are added to the pool of papers to be published, and we let $\p \in (0, 1]$ be a fixed rate of acceptance. The pool of unaccepted papers evolves like the following dynamical system for $x_1 \gets \N$: \[x_{t+1} \gets x_t (1-\p) + \N.\] This converges fast to a fixed point $\xast$, which is the solution to $\xast = \xast(1-\p) + \N$, yielding \[\xast = \frac{\N}{\p},\] \[\textit{#accepted_papers}= \xast \cdot \p = \frac{\N}{\p} \cdot \p = \N.\] Amazing!! #accepted_papers does not change with $\p$. If we reduce $\p$, the only effect is that the pool size grows until it’s so big that a fraction $\p$ of it ends up being the same number of papers $\N$, and we review more for nothing $(\propto \N/\p)$. We’d accept the same amount of papers in each conference! More easily: at the fixed poi...