The issue of mining association rules with composite items was proposed several years ago. Algorithms with composite items have the potential to discover rules which cannot be found out by other algorithms without composite items. However, much redundant rules which are of trivial significance or even incorrect will be also discovered by these algorithms in certain cases. In this paper, we design a Novel Frequent-Pattern tree for finding large composite items first. And then how to measure the reliability of these discovered rules with composite items in order to find out the most reliable association rules is discussed.