PROJECT TITLE :

Unsupervised Ensemble Classification with Sequential and Networked Data

ABSTRACT:

Ensemble Learning, a paradigm of Machine Learning in which multiple models are combined, has shown promising performance in a variety of tasks. This can be attributed to the fact that it combines more than one model. The unsupervised classification of ensemble data is the primary focus of this work. The ensemble combiner that does not know the ground-truth labels that each classifier has been trained on is referred to as unsupervised, and the term "unsupervised" describes this type of combiner. In contrast to the majority of earlier works on unsupervised ensemble classification, which were designed for data that is independent and identically distributed (i.i.d.), the current work presents an unsupervised method for learning from ensembles of classifiers even when there are dependencies between the data. Consideration is given to two distinct kinds of data dependencies: sequential data and networked data, the dependencies of which can be represented by a graph. Innovative algorithms for moment matching and expectation-maximization are developed for both of these problems. Knowledge of data dependencies within the meta-learner is beneficial for the unsupervised ensemble classification task, as shown by the evaluation of these algorithms' performance on both synthetic and real datasets.