Association rule mining is one of the useful techniques in data mining and knowledge discovery that extracts interesting relationships between items in datasets. Generally, the number of association rules in a particular dataset mainly depends on the measures of 'support' and 'confidence'. To choose the number of useful rules, normally, the measures of 'support' and 'confidence' need to be tried many times. In some cases, the measures of 'support' and 'confidence' are chosen by experience. Thus, it is a time consuming to find the optimal measure of 'support' and 'confidence' for the process of association rule mining in large datasets. This paper proposes a regression based approach to improve the association rule mining process through predicting the number of rules on datasets. The approach includes a regression model in a generic level for general domains and an instantiation scheme to create concrete models in particular domains for predicting the potential number of association rules on a dataset before mining. The proposed approach can be used in broad domains with different types of datasets to improve the association rule mining process. A case study to build a concrete regression model based on a real dataset is demostrated and the result shows the good performance of the proposed approach. © 2012 Springer-Verlag.