EM-Clustering when double isn’t enough…

One of the more interesting developments in clustering (in my opinion) is clustering of data the is on a unit hypersphere. It sounds like some rare special case at first, but appears quite frequently in real life applications such as Spectral Embeddings in Spectral Clustering, some subdomains of Bio-Informatics data or text-data in TFIDF representation. The data can be analyzed as unit vectors on a d-dimensional hypersphere, or equivalently are directional in nature. Spectral clustering techniques generate embeddings in the normalization step that constitute an example of directional data and can result in different shapes on a hypersphere.

The first paper published that suggested a good clustering algorithm presented an Expectation Maximization (EM) algorithm for the von Mises-Fisher distribution (Banerjee et.al, JMLR (6) 2005). Avleen and myself started to work on extension for this that utilizes the Watson distribution, a distribution for directional data that has more modeling capability than the simple von Mises-Fisher distribution. We just published our results for the Watson EM clustering algorithm in the AISTATS 2007 conference to be held in March (Matlab code will be available soon).

One problem with both algorithms is that they require a high precission number representation in order to work well for high-dimensional problems such as bio-informatics data and text. Most prior work with directional data was limited to maybe 3 dimensional cases, and most Kummer-function approximations (another problem we had to address) work well only for the lower dimensional cases. In our AISTATS paper we only presented results for lower dimensional embeddings as we had some problems getting it to work for higher dimensional data (also, the root-solver that was involved is just incapable of handling larger problems). We have been working on a speedup with some success, but I have to say that it was mostly the numerical problems that gave us a hard time.

More and more Machine Learning techniques require a more careful consideration of numerical problems (Support Vector Machines, my manifold clustering algorithm etc.) and I run into numerical problems every other day. While trying to improve our Watson-EM algorithm I found out that Continued Fractions have many desirable properties such as the unique, finite representation of rational numbers. Numbers can be represented exact with no numerical error. In the EM algorithm we use them to approximate the Kummer function. Maybe a more exact number representation for fractions can be made out of this?

So I started looking and found in an tech-report that the number representation of continued fractions can be nicely implemented in Prolog. It also explains how to add up numbers in a Continued Fraction Representation and so on with an arbitrary precision.

I haven’t found any papers yet that suggest a suitable hardware implementation of Continued Fractions to replace the IEEE floating point numbers we use nowadays, but it can probably be done.

Leave a Reply

You must be logged in to post a comment.