Unsupervised ensemble learning with dependent classifiers
Llobet Turró, Martí · UPCommons institutional repository (Universitat Politècnica de Catalunya) · 2021
Data sets containing instances that are assigned values by an ensemble of annotators are becoming increasingly common. These annotators or classifiers are often of unknown reliability and no ground truth labels are provided for the several instances to assess it. There are many cases of interest, most notably crowdsourcing, though such data sets are also common in computational biology, [1] [2] computer vision, [3] natural language processing, [4] and 5G communications systems, [5] among others.Our work focuses on the classification of potentially correlated binary data through two different approaches present in the literature. To achieve this goal, two challenges emerge. First, finding out about the classifiers’ intrinsic correlation structure, and secondly, obtaining a meta-learner that classifies each instance making use of the classifiers’ correlation structure and assignments over many instances. We rely on different algorithms in the literature to address such challenges. The algorithms are extensively tested with artificially generated data and real data sets in the field of computational genomics to test the algorithms. Specifically, we focus on pathway-based classification of clinical occurrences in breast cancer samples.