Provenance-based Belief

Provenance has been touted as a basis to establish trust in data. Intuitively, belief in a hypothesis should depend on how much one trusts the relevant data. However, current proposals to assess trust based solely on provenance are insufficient for rigourous decision making. We describe a model of provenance and belief that is necessary and sufficient to incorporate —trust in the data— in a way that supports normative inference. The model is based on the observation that provenance can be viewed as a causal structure which can be used to compute belief from assessments of the accuracy of sources and transformations that produced relevant data. In our model, data sources are like sensors with associated conditional probability tables. Provenance identifies dependencies among sensors. Together, this information allows construction of causal networks that can be used to compute the belief in a state of the world based on observation of data. This model formalizes the role of source accuracy, and provides a method for formally assessing belief that uses only information in the provenance store, not the contents of the data.