Depth First Generation of Frequent Patterns Without Candidate Generation

Abstract

Mining frequent patterns has been studied popularly in data mining research. Most of the current studies adopt an FP_growth-like approach which does not bring the candidate generation. However, the cost of recursively constructing each frequent item’s conditional frequent pattern tree is high. In this paper, we propose a depth first algorithm for mining frequent patterns. Efficiency of mining is achieved with the following techniques: large database is compressed into a frequent pattern tree with a children table but not a header table, which avoids costly repeated database scans, on the other hand the mining algorithm adopts a depth first method which takes advantage of this tree structure and dynamically adjusts links instead of generating a lot of redundant sub trees, which can dramatically reduces the time and space needed for the mining process. The performance study shows that our algorithm is efficient and scalable for mining frequent patterns, and is an order of magnitude faster than Trie, FP_growth, H-mine and some recently reported new frequent patterns mining methods.