Mining sophisticated patterns for classification and correlation analysis

Pattern mining has been a hot issue since it was first proposed for market basket analysis. Even though pattern mining is one of the oldest topic in data mining domain, there are still many ongoing challenges to overcome on this subject since the scale of the data size is getting bigger and the complexity of data structure is getting more complicated. This dissertation discusses several pattern mining tasks, challenges associated with them, and algorithm designs that overcome these challenges. Specifically, we design and implement techniques for (1) directly mining discriminative patterns from a numeric valued feature set of k-embedded edge subtrees given labeled training data, (2) mining top correlated patterns from transactional databases with low minimum support, and (3) mining flipping correlation patterns from transactional databases given item hierarchy. We evaluate our solutions by conducting comprehensive experiments on large-scale synthetic and real world datasets