Collaborative filtering, market basket analysis, and network motifs all try to find significant substructures in real-world market basket data, i.e., the information which customers bought which products. An important substructure is the number of customers which bought a given subset of products: if this number is significant with respect to an appropriate expectation, then the finding can be used to give recommendations to those customers which only bought a part of these products so far. I will show that the classic way of computing the expected value is based on the wrong expectation model and introduce a new expectation model which remedies the problem. The results of the new algorithm are quantitatively evaluated on a new benchmark set from the Netflix-prize data set.