Citizen Science games and gamified tasks have shown promise in engaging citizens in scientific work through often digitalized version of real-life scientific tasks while producing useful data.
This work has spurred an increasing interest in investigating how to design such interactions to entice players to produce as high quantities of scientifically useful data as possible.
Player modelling, the computational analysis of player interactions with and within games, has been applied to better understand which UI features or particular player interactions can best predict player behavior.
However, much of this work has relied on modeling approaches that exhibit low degrees of explainability, and that rely on large quantities of player data.
In this paper we present Saturable Dynamic Bayesian Network Models (SDBN), a novel player modelling approach that avoids these two issues using a \textit{feature activation network} that quantifies the user's gradually extending familiarity with the interface.
This makes our approach usable for more fine-grained design-decisions when building and improving citizen science games, and it makes our approach applicable to games with smaller numbers of players.
We apply our novel method to data collected from approximately 2,000 players of the citizen science game Quantum Moves 2.
Our study shows that the SDBN approach can reliably model player drop out down to 10 participants and allows for both bulk and individual level insights of how player interactions with and within our game predicts player engagement and the quality of consequent data.