Dynamic Model Selection With its Applications to Novelty Detection

Abstract

We are concerned with the issue of dynamically selecting optimal statistical models from time series. The goal is not to select a single optimal model over the data as in conventional model selection, but to select a time series of optimal models under the assumption that the data source may be nonstationary. We call this issue dynamic model selection (DMS). From the standpoint of minimum description length principle, we first propose coding-theoretic criteria for DMS. Next, we propose efficient DMS algorithms on the basis of the criteria and analyze their performance in terms of their total code lengths and computation time. Finally, we apply DMS to novelty detection and demonstrate its effectiveness through empirical results on masquerade detection using UNIX command sequences.