Classification by Pattern-Based Hierarchical Clustering

In this paper, we propose CPHC, a semi-supervised classification algorithm that uses a pattern-based cluster hierarchy as a direct means for classification. All training and test instances are first clustered together using an instance-driven pattern-based hierarchical clustering algorithm that allows each instance to "vote" for its representative size-2 patterns in a way that balances local pattern significance and global pattern interestingness. These patterns form initial clusters and the rest of the cluster hierarchy is obtained by following a unique iterative cluster refinement process that exploits local information. The resulting cluster hierarchy is then used directly to classify test instances, eliminating the need to train a classifier on an enhanced training set. For each test instance, we first use the hierarchical structure to identify nodes that contain the test instance, and then use the labels of co-existing training instances, weighing them proportionately to their pattern lengths, to obtain the most likely class label(s) for the test instance. In addition, CPHC increases the chances of classifying isolated test instances by inducing a type of feature transitivity. Results of experiments performed on 19 standard text and machine learning datasets show that CPHC outperforms a number of existing classification algorithms even with sparse (as low as 1%) training data.