COST-tree: Identification of Cosine Interesting Patterns Based on FP-tree

The cosine similarity, also known as un-centered Pearson Correlation, has been widely used for mining association patterns, which contain objects strongly related to each other. However, it is often used as a post-evaluation measure and is computationally prohibitive for large data. To this end, we develop an FP-tree like algorithm, named COST-tree, for finding association patterns based on the cosine measure. A key idea is to combine the strength of the FP-tree structure and the Conditional Anti-Monotone Property of the cosine measure. Experimental results on real-world data demonstrate the effectiveness of COST-tree, in particular for finding rare but interesting patterns at extremely low support levels.