Discovery of classifications from data of multiple sources

We study a learning paradigm that bridges between supervised learning and unsupervised learning. In this paradigm, the learner is given unlabeled examples described by several sets of attributes. The task of learning is to (re)construct class labels consistent with the multiple sets of attributes. We design a novel learning algorithm, called AutoLabel, for this type of learning tasks, and we identify the source of power in the algorithm. We test AutoLabel on artificial and real-world datasets, and show that it constructs classification labels accurately. Our learning algorithm removes the fundamental assumption of providing class labels in supervised learning, and gives a new perspective to unsupervised learning.