Predictive modeling problems involving time-series as inputs are quite common in the medical and financial domains (and certainly not restricted to these areas). For example, consider a medical application where various monitored measurement (heart rate, glucose levels, immune activity) over time are predictive of survival of an animal. In such cases, it is natural to expect the entire profile or particular continuous subsequences of the profile to be responsible for the response. In this work we present a feature selection scheme that attempts to address this problem by discovering and modeling highly predictive contiguous subsequences.

Our model is called LAPS (the Lasso with Attribute Partition Search) and is inspired by the Fused Lasso (Tibshirani et al., 2005) and the Group Lasso (Yuan and Lin, 2006; Meier van de Geer and Buhlmann, 2006) models.