image

Have you cake and eat it: Flexibility vs. interpretability in Data Science

Más información

FECHA :
Octubre 04, 2018
De 10:00 a 11:00 hrs

University of Santa Cruz
Estados Unidos

Over the last 10 years, Deep Learning (DL) methods (which are based on exible models build by the repeated composition of non-linear kernels) have shown success in tasks as diverse as image recognition and automatic machine translation. Although this success has made DL extremely popular, the methods present important practical drawbacks. Of these, the inability of DL to provide “human-friendly” explanations is one of the most serious. Consider, for example, so-called autoencoders, which aim to learn a sparse, low-dimensional representation for a set of observations. Unlike standard statistical methods such as principal component or factor analysis, the low-dimensional features constructed by an autoencoder typically have no straightforward interpretation in terms of the original application. While this is usually not a problem if the features will be used to construct a predictive model, the lack of interpretability is a serious issue in many social science applications where the main interest lies precisely on the latent features. This talk will provide an example of how carefully-constructed statistical methods can play a key role in the derivation of data science methods that are both exible and interpretable. In particular, we will introduce a new class of embedding models for matrix-valued binary data, and discuss a particular instantiation of the class that is motivated by practical issues associated the analysis of legislative voting data in the United States and the United Kingdom. The resulting model is a generalization of the class of spatial voting model from political science that allows the preference space to be any arbitrary Riemannian manifolds. The features of the model are illustrated using a number of datasets of voting records from the US House of Representatives.

¿Te pareció interesante?
Dirígete a la plataforma de inscripción y regístrate en línea

Quiero Inscribirme