Most existing studies on association rules discovery focused on finding the association rules between all items in a large database that satisfy user-specified minimum confidence and support. In practice, users are often interested in finding association rules involving only some specified items. Meanwhile, based on the search results in former queries, users tend to change the minimal confidence and support requirements to obtain suitable number of rules. Under these constraints, the existing mining algorithms can not perform efficiently due to high and repeated disk access overhead. In this research, we present a novel mining algorithm that can efficiently discover the association rules between the user-specified items or categories via feature extraction approach. At most one scan of the database is needed for each query; hence, the disk access overhead can be reduced substantially and the query be responded quickly.