Friday, 16 December 2016

Mixed-type data analysis IV: Representing multivariate ordinal data

Multivariate ordinal data is popular when human judgement is involved. For example, in collaborative filtering, we rate multiple items, each of which with a number of stars. In a typical survey, we provide ordinal assessment of many things, ranging from feeling of the day (happy, OK, sad) to the current situation of worldwide security (safe, OK, dangerous). Since these come from the same person, they are correlated, and thus we need to model multiple ordinal variables simultaneously. This blog will present an overview of the area.

Much of existing work, however, is focusing on single ordinal variable, typically under the umbrella of "ordinal regression". How about multiple ordinal variables?

There are several approaches. One way is to assume that ordinal data are just quantized version of an underlying continuous variable. Thus, each ordinal value corresponds to an interval of the underlying variable. This is intuitive, for example, when we says salary levels A, B and C, and they refer to ranges like A = $[50K,60K], B = $[60K,70K] and C = $70K+.

This thinking is convenient, especially when the underlying variable is assumed to be Gaussian. We can build a multivariate Gaussian distribution. The problem is that we will never observe these Gaussian variables directly but indirectly through the intervals dictated by the ordinal levels. Things get more interesting when the intervals are unknown. The only requirement is that the intervals have to be consecutive (i.e., no gaps). With this, we need to estimate the interval boundaries from data.

This is basically the main idea behind this paper published in ACML'12. However, we go further because the multivariate Gaussian distributions are hard to sample from under interval constraints. We leverage Gaussian-Bernoulli Restricted Boltzmann Machines instead. This makes MCMC sampling quite efficiently. The RBM style can also make it easy to extend to model the matrix with row and column RBMs linked together.

The other way is to use log-linear model, treating the ordinal as categorical but with log-linear constraints among the ordered levels. This is the idea behind this work published in UAI'09.

Updated references:

  • Ordinal Boltzmann Machines for Collaborative Filtering. Truyen Tran, Dinh Q. Phung and Svetha Venkatesh. In Proc. of 25th Conference on Uncertainty in Artificial Intelligence, June, 2009, Montreal, Canada. 
  • A Sequential Decision Approach to Ordinal Preferences in Recommender Systems, Truyen Tran, Dinh Phung, Svetha Venkatesh, in Proc. of 25-th Conference on Artificial Intelligence (AAAI-12), Toronto, Canada, July 2012
  • Cumulative Restricted Boltzmann Machines for Ordinal Matrix Data Analysis, Truyen Tran, Dinh Phung and Svetha Venkatesh, in Proc. of. the 4th Asian Conference on Machine Learning (ACML2012), Singapore, Nov 2012.
  • Ordinal random fields for recommender systems, Shaowu Liu, Truyen Tran, Gang Li, Yuan Jiang, ACML'14, Nha Trang, Vietnam, Nov 2014.

No comments:

Post a Comment