Semi-Supervised and Unsupervised Machine Learning: Novel Strategies

Chapter 4

Semi-Supervised Classification Using Pattern Clustering

4.1. Introduction

In the previous chapter, a semi-supervised approach has been described, which gained advantage of unlabeled data by means of clustering. A minimum labeled seed was fed to a supervised classifier. Then, the lexicon features in these initial labeled seeds were automatically expanded through a set of synonym groups found by the clustering algorithm.

In this chapter, a new alternative to semi-supervised algorithm is introduced. In a similar way as the approach described in the previous chapter, clustering is also used to “augment” the small labeled seeds. However, in contrast to the previous approach, the cluster assumption is now applied to obtain groups of data instances instead of features. This cluster principal assumption — underlying class labels should naturally fall into clusters — has been frequently applied to other works in the semi-supervised machine learning (ML) literature [BLU 01].

Following the clustering step, the labeled seeds have been used to tag the clusters in such a way that the initial labels are augmented to the complete clustered data. In the previous chapter, an explicit labeling step was absent. However, by assuming no overlap of the terms from different categories inside each extracted group, an implicit labeling of semantic clusters was performed. The clusters that overlapped at least one term with one of the labeled prototypes remained labeled with the class label of the ...