This paper proposes to integrate two very different kinds of methods for data mining, namely the construction of Bayesian networks from data and the detection of occurrences of Simpson’s paradox. The former aims at discovering potentially causal knowledge in the data, whilst the latter aims at detecting surprising patterns in the data. By integrating these two kinds of methods we can hope fully discover patterns which are more likely to be useful to the user, a challenging data mining goal which is under-explored in the literature. The proposed integration method involves two approaches. The first approach uses the detection of occurrences of Simpson’s paradox
as a preprocessing for a more effective construction of Bayesian networks; whilst the second approach uses the construction of a Bayesian network from data as a preprocessing for the detection of occurrences of Simpson’s paradox