The K2 metric is a well-known evaluation measure (or scoring function) for learning Bayesian networks from data [7]. It is derived by assuming uniform prior distributions on the values of an attribute for each possible instantiation of its parent attributes. This assumption introduces a tendency to select simpler network structures. In this paper we modify the K2 metric in three different ways, introducing a parameter by which the strength of this tendency can be controlled. Our experiments with the ALARM network [2] and the BOBLO network [17] suggest that - somewhat contrary to our expectations - a slightly stronger tendency towards simpler structures may lead to even better results.
展开▼
