Recurrent Predictive Models for Sequence Segmentation

Many sequential data sets have a segmental structure, and similar types of segments occur repeatedly. We consider sequences where the underlying phenomenon of interest is governed by a small set of models that change over time. Potential examples of such data are environmental, genomic, and economic sequences. Given a target sequence and a (possibly multivariate) sequence of observation values, we consider the problem of finding a small collection of models that can be used to explain the target phenomenon in a piecewise fashion using the observation values. We assume the same model will be used for multiple segments. We give an algorithm for this task based on first segmenting the sequence using dynamic programming, and then using k-median or facility location techniques to find the optimal set of models. We report on some experimental results.