Title: Spectral methods for unsupervised ensemble learning and latent variable models

 

Abstract

With the availability of huge amounts of unlabeled data, unsupervised learning methods 

are gaining increasing popularity and importance. We focus on

"unsupervised ensemble learning", where one obtains the predictions of multiple 

classifiers over a set of unlabeled instances. The classifiers may be human

experts as in crowdsourcing, or prediction algorithms developed by research

groups worldwide. The challenge is to estimate the accuracies of the different classifiers 

and combine them to an accurate meta-learner. To tackle this

problems we show how it relates to latent variable models, and derive simple

estimates for the classifiers' accuracies based on a spectral analysis of the ob-

served data. On the experimental side, we apply our methods to a problem in

Computational Biology, where for various classification tasks one combines the

results of multiple algorithms for improved accuracy.

In the second part of the talk, I will focus on extending the techniques developed 

for unsupervised ensemble learning to a specific family of linear latent

variable models. For cases where the latent layer is binary, we derive an interesting 

relation between the model parameters and the relatively recent notion

of tensor eigenvectors of the data higher order moments. We apply our methods

to the problem of inferring global ancestry in population genetics.