Constrained AdaBoost for Totally-Ordered Global Features

This paper proposes a constrained AdaBoost algorithm for utilizing global features in a dynamic time warping (DTW) framework. Global features are defined as a spatial relationship between temporally-distant points of a temporal pattern and are useful to represent global structure of the pattern. An example is the spatial relationship between the first and the last points of a handwritten pattern of the digit "0". Those temporally-distant points should be spatially close enough to form a closed circle, whereas those points of "6" should be distant enough. For a temporal pattern of an N-point sequence, it is possible to have N(N -- 1)/2 global features. One problem of using the global features is that they are not ordered as a one dimensional sequence any more. Consequently, it is impossible to use them in a left-to-right Markovian model, such as DTW and HMM. The proposed constrained AdaBoost algorithm can select a totally-ordered subset from the set of N(N -- 1)/2 global features. Since the totally-ordered features can be arranged as a one-dimensional sequence, they can be incorporated into a DTW framework for compensating nonlinear temporal fluctuation. Since the selection is governed by the AdaBoost framework, the selected features can retain discriminative power.