A method was preposed for mining conditional preferences and learning CP-nets on the data stream based on the inverted matrix structure. The transaction layout using the inverted matrix reduces the number of times the database is scanned, and through random access, frequent preference items can be found in less than one full scan. In addition, through the establishment of frequent preference tree FP-Tree, the generation of candidates is reduced. Experimental results show that, compared with other methods of learning CP-nets structure, this method can obtain accurate CP-nets faster, show better performance in large transaction databases, and reduce memory requirements.