Penalty learning for changepoint detection

We consider the problem of signal segmentation in the setup of supervised learning. The supervision lies here in the existence of labelled change points in a historical database of similar signals. Typical segmentation techniques rely on a penalized least square procedure where the smoothing parameter is fixed arbitrarily. We introduce the alpin (Adaptive Linear Penalty INference) algorithm to tune automatically the smoothing parameter. ALPIN has linear complexity with respect to the sample size and turns out to be robust with respect to noise and diverse annotation strategies. Numerical experiments reveal the efficiency of ALPIN compared to state-of-the-art methods.