The viability of any Citizen Science (CS) research program is absolutely conditioned to the engagement of the citizen. In a CS framework in which participants are expected to perform actions that can be later on validated, the incorporation of a reputation system can be a successful strategy to increase the overall data quality and the likelihood of engagement, and also to evaluate how close citizens fulfill the goals of the CS research program. Under the assumption that participant actions are validated using a simple discrete rating system, current reputation models, thoroughly applied in e-platform services, can be easily adapted to be used in CS frameworks. However, current reputation models implicitly assume that rated items and scored agents are the same entity, and this does not necessarily hold in a CS framework, where one may want to rate actions but score the participants generating it. We present a simple approach based on a Bayesian network representing the flow described above (user, action, validation), where participants are aggregated in a discrete set of user classes and we use the global evidence in the data base to estimate both the prior and the posterior distribution of the user classes. Afterwards, we evaluate the expertise of each participant by computing the user-class likelihood of the sequence of actions/validations observed for that user. As a proof of concept we implement our model in a real CS case, namely the Mosquito Alert project.