Unsupervised Ensemble Classification With Correlated Decision Agents

Abstract

Decision-making procedures when a set of individual binary labels is processed to produce a unique joint decision can be approached modeling the individual labels as multivariate independent Bernoulli random variables. This probabilistic model allows an unsupervised solution using EM-based algorithms, which basically estimate the distribution model parameters and take a joint decision using a Maximum a Posteriori criterion. These methods usually assume that individual decision agents are conditionally independent, an assumption that might not hold in practical setups. Therefore, in this work we formulate and solve the decision-making problem using an EM-based approach but assuming correlated decision agents. Improved performance is obtained on synthetic and real datasets, compared to classical and state-of-the-art algorithms.

Correlated Binary Data for Machine Learning

“…The covariance matrix characterizing a database containing binarized pathway activation measurements is included to asess the methods' ability to resemble an actual, real-world case. Also, we present testing results using two state-of-the-art binary ensemble meta-learners-the Correlated Expectation-Maximization (CEM) and the Latent Spectral Meta-Learner (LSM), presented in [7] and [8], respectively. We include the relative errors featured by both meta-learners using both methods for Cases 1-12.…”

“…We present a new method to generate artificial correlated, binary data sets, denoted Method 1. This method is an upgraded version of the approach applied in [7], which has been entirely revisited to obtain easily tunable, realistic correlation patterns. The approach applied in [8], denoted Method 2, is also reviewed for comparative purposes.…”

“…Two recent works tackle the issue of inter-classifier dependencies, each developing an adapted meta-learner. In [7], the EM algorithm is used to estimate the underlying probability distribution of the instances' ground truth labels; in [8], an algorithm identifying strong inter-classifier correlations is developed and ultimately used to adapt an already existing meta-learner. These two works were developed on the basis of two different approaches as to the modeling of inter-classifier dependencies.…”

“…We focus extensively on the approaches presented in [7] and [8]. Our work exploits the one in [7] to obtain an artificial data generation method with realistic correlation patterns.…”

“…We focus extensively on the approaches presented in [7] and [8]. Our work exploits the one in [7] to obtain an artificial data generation method with realistic correlation patterns. Based on the potentially existing relationships between the different annotators, an intuitive designation of their correlation structure is introduced.…”